Merge branch 'main' of https://code.rizlum.com/ngpthanh15/IQA-Metric-Benchmark

update ressult
update humans label
2025-09-12 08:03:17 +00:00 · 2025-09-11 09:39:02 +00:00 · 2025-09-08 07:43:21 +00:00
21 changed files with 1556 additions and 1 deletions
--- a/data/.gitignore
+++ b/data/.gitignore
@@ -1,5 +1,4 @@
 *.png
 *.csv
 *.jpg
 *.jpeg
 *.gif
--- a/data/task/cni/Hallucination/anls_pipeline/correlation_results.csv
+++ b/data/task/cni/Hallucination/anls_pipeline/correlation_results.csv
@@ -0,0 +1,17 @@
 method,n_overlap,pearson
 brisque_per_image.csv,46,-0.104588
 clipiqa+_vitL14_512_per_image.csv,46,0.011055
 dbcnn_per_image.csv,46,-0.203747
 deqa_per_image.csv,46,0.094139
 hyperiqa_per_image.csv,46,-0.095212
 maniqa_per_image.csv,46,-0.007514
 musiq_per_image.csv,46,-0.116055
 nima_per_image.csv,46,-0.038136
 niqe_per_image.csv,46,0.001954
 nrqm_per_image.csv,46,-0.005415
 paq2piq_per_image.csv,46,0.048295
 piqe_per_image.csv,46,0.078599
 topiq_nr_per_image.csv,46,-0.136474
 unique_per_image.csv,46,-0.078535
 uranker_per_image.csv,46,0.122599
 human_per_image.csv,46,0.154902
--- a/data/task/cni/Hallucination/anls_pipeline/human-labels.csv
+++ b/data/task/cni/Hallucination/anls_pipeline/human-labels.csv
@@ -0,0 +1,82 @@
 filename,label
 ITE55336627_PASSEPORT_Abasse_GUEYE__Completed_0.jpg,Low
 ITE55340318_ID_kat__1__Completed_0.jpg,Low
 ITE55340822_passeport_pascal_2028_Completed_0.jpg,High
 ITE55341271_C_I_face_Completed_0.jpg,High
 ITE55341677_2023_-_CNI_Recto-Verso_Hassina_0.jpg,High
 ITE55343297_20250729_172626_Completed_0.jpg,High
 ITE55343716_CNI_Completed_0.jpg,High
 ITE55343716_CNI_Completed_1.jpg,High
 ITE55346966_17537774579547958575370370624241_Completed_0.jpg,Low
 ITE55347865_Snapchat-715567440_Completed_0.jpg,Low
 ITE55347866_Snapchat-1551171803_Completed_0.jpg,Low
 ITE55347926_DHONDT_CNI_0.jpg,High
 ITE55348016_image_Completed_0.jpg,High
 ITE55348464_IMG-20250721-WA0000_Completed_0.jpg,High
 ITE55348878_CNI_RECTO_0.jpg,Low
 ITE55348904_cni_verso_0.jpg,High
 ITE55349793_2_em_CNI_CHASTAING__0.jpg,High
 ITE55350967_PLI_INCONNU_-_ADH_625302_42296806_DMS_0.jpg,High
 ITE55351087_PLI_INCONNU_-_ADH_625120_42296075_DMS_0.jpg,Low
 ITE55351128_PLI_INCONNU_-_ADH_625123_42296103_DMS_0.jpg,Low
 ITE55351128_PLI_INCONNU_-_ADH_625123_42296103_DMS_1.jpg,Low
 ITE55351156_Passeport_Completed_0.jpg,High
 ITE55351372_RÉPUBLIQUE_FRANÇAISE_Completed_0.jpg,High
 ITE55351372_RÉPUBLIQUE_FRANÇAISE_Completed_1.jpg,High
 ITE55352532_PLI_INCONNU_-_ADH_625374_42297046_DMS_0.jpg,Low
 ITE55352984_CNI_Completed_0.jpg,Low
 ITE55352984_CNI_Completed_1.jpg,Low
 ITE55354550_20250728_113359_Completed_0.jpg,High
 ITE55354891_image_Completed_0.jpg,High
 image (1)_0.jpg,High
 image (1)_1.jpg,High
 image (10)_0.jpg,Low
 image (10)_1.jpg,Low
 image (11)_0.jpg,Low
 image (11)_1.jpg,Low
 image (12)_0.jpg,High
 image (12)_1.jpg,High
 image (13)_0.jpg,High
 image (13)_1.jpg,High
 image (14)_0.jpg,High
 image (15)_0.jpg,High
 image (16)_0.jpg,High
 image (17)_0.jpg,High
 image (17)_1.jpg,High
 image (18)_0.jpg,High
 image (18)_1.jpg,High
 image (18)_2.jpg,High
 image (18)_3.jpg,Low
 image (19)_0.jpg,High
 image (19)_1.jpg,High
 image (2)_0.jpg,High
 image (21)_0.jpg,High
 image (22)_0.jpg,High
 image (22)_1.jpg,High
 image (22)_2.jpg,High
 image (22)_3.jpg,High
 image (23)_0.jpg,High
 image (24)_0.jpg,High
 image (25)_0.jpg,High
 image (26)_0.jpg,High
 image (27)_0.jpg,High
 image (28)_0.jpg,High
 image (28)_1.jpg,High
 image (29)_0.jpg,High
 image (3)_0.jpg,High
 image (30)_0.jpg,High
 image (31)_0.jpg,High
 image (4)_0.jpg,High
 image (5)_0.jpg,Low
 image (6)_0.jpg,High
 image (6)_1.jpg,High
 image (6)_2.jpg,High
 image (6)_3.jpg,Low
 image (6)_4.jpg,High
 image (6)_5.jpg,High
 image (6)_6.jpg,High
 image (7)_0.jpg,High
 image (7)_1.jpg,High
 image (8)_0.jpg,High
 image (9)_0.jpg,High
 image_0.jpg,High
--- a/data/task/cni/Hallucination/anls_pipeline/per_image_anls.csv
+++ b/data/task/cni/Hallucination/anls_pipeline/per_image_anls.csv
@@ -0,0 +1,48 @@
 image,anls,hallucination_score,num_fields
 image (1),0.855357,0.144643,14
 image (11),0.440000,0.560000,15
 image (12),0.834921,0.165079,14
 image (13),0.837897,0.162103,14
 image (14),0.800000,0.200000,5
 image (15),0.944444,0.055556,18
 image (16),0.868056,0.131944,16
 image (17),0.849206,0.150794,14
 image (18),0.731783,0.268217,15
 image (19),0.849206,0.150794,14
 image (2),0.000000,1.000000,1
 image (21),0.549781,0.450219,18
 image (22),0.880174,0.119826,17
 image (23),0.565217,0.434783,14
 image (24),0.666667,0.333333,15
 image (26),0.833333,0.166667,18
 image (27),0.682692,0.317308,16
 image (28),0.571429,0.428571,14
 image (29),0.494359,0.505641,15
 image (3),0.533333,0.466667,15
 image (30),0.888889,0.111111,18
 image (31),0.563333,0.436667,15
 image (4),0.566667,0.433333,15
 image (6),0.853383,0.146617,14
 image (8),0.843367,0.156633,14
 image (9),0.661905,0.338095,15
 image,0.367946,0.632054,15
 ITE55336627_PASSEPORT_Abasse_GUEYE__Completed,0.414286,0.585714,14
 ITE55340318_ID_kat__1__Completed,0.851677,0.148323,15
 ITE55340822_passeport_pascal_2028_Completed,0.857143,0.142857,14
 ITE55341271_C_I_face_Completed,0.888889,0.111111,18
 ITE55346966_17537774579547958575370370624241_Completed,0.400000,0.600000,5
 ITE55347865_Snapchat-715567440_Completed,0.803922,0.196078,17
 ITE55347866_Snapchat-1551171803_Completed,0.774928,0.225072,18
 ITE55347926_DHONDT_CNI,0.931624,0.068376,18
 ITE55348016_image_Completed,0.882353,0.117647,17
 ITE55348464_IMG-20250721-WA0000_Completed,0.583492,0.416508,15
 ITE55348878_CNI_RECTO,0.991900,0.008100,17
 ITE55349793_2_em_CNI_CHASTAING_,0.760000,0.240000,15
 ITE55350967_PLI_INCONNU_-_ADH_625302_42296806_DMS,0.723529,0.276471,17
 ITE55351087_PLI_INCONNU_-_ADH_625120_42296075_DMS,0.848366,0.151634,17
 ITE55351128_PLI_INCONNU_-_ADH_625123_42296103_DMS,0.812707,0.187293,14
 ITE55351156_Passeport_Completed,0.761364,0.238636,16
 ITE55352532_PLI_INCONNU_-_ADH_625374_42297046_DMS,0.428571,0.571429,14
 ITE55352984_CNI_Completed,0.793939,0.206061,15
 ITE55354550_20250728_113359_Completed,0.843712,0.156288,14
 ITE55354891_image_Completed,0.882353,0.117647,17
--- a/data/task/cni/Human/human-label.csv
+++ b/data/task/cni/Human/human-label.csv
@@ -0,0 +1,82 @@
 imagefile,coherence
 ITE55336627_PASSEPORT_Abasse_GUEYE__Completed_0.png,4
 ITE55340318_ID_kat__1__Completed_0.png,3
 ITE55340822_passeport_pascal_2028_Completed_0.png,3
 ITE55341271_C_I_face_Completed_0.png,4
 ITE55341677_2023_-_CNI_Recto-Verso_Hassina_0.png,4
 ITE55343297_20250729_172626_Completed_0.png,4
 ITE55343716_CNI_Completed_0.png,2
 ITE55343716_CNI_Completed_1.png,2
 ITE55346966_17537774579547958575370370624241_Completed_0.png,3
 ITE55347865_Snapchat-715567440_Completed_0.png,3
 ITE55347866_Snapchat-1551171803_Completed_0.png,2
 ITE55347926_DHONDT_CNI_0.png,1
 ITE55348016_image_Completed_0.png,4
 ITE55348464_IMG-20250721-WA0000_Completed_0.png,2
 ITE55348878_CNI_RECTO_0.png,2
 ITE55348904_cni_verso_0.png,3
 ITE55349793_2_em_CNI_CHASTAING__0.png,3
 ITE55350967_PLI_INCONNU_-_ADH_625302_42296806_DMS_0.png,3
 ITE55351087_PLI_INCONNU_-_ADH_625120_42296075_DMS_0.png,2
 ITE55351128_PLI_INCONNU_-_ADH_625123_42296103_DMS_0.png,2
 ITE55351128_PLI_INCONNU_-_ADH_625123_42296103_DMS_1.png,1
 ITE55351156_Passeport_Completed_0.png,5
 ITE55351372_RÉPUBLIQUE_FRANÇAISE_Completed_0.png,5
 ITE55351372_RÉPUBLIQUE_FRANÇAISE_Completed_1.png,4
 ITE55352532_PLI_INCONNU_-_ADH_625374_42297046_DMS_0.png,2
 ITE55352984_CNI_Completed_0.png,2
 ITE55352984_CNI_Completed_1.png,2
 ITE55354550_20250728_113359_Completed_0.png,4
 ITE55354891_image_Completed_0.png,4
 image (1)_0.png,3
 image (1)_1.png,2
 image (10)_0.png,2
 image (10)_1.png,1
 image (11)_0.png,2
 image (11)_1.png,2
 image (12)_0.png,3
 image (12)_1.png,3
 image (13)_0.png,2
 image (13)_1.png,2
 image (14)_0.png,3
 image (15)_0.png,4
 image (16)_0.png,2
 image (17)_0.png,3
 image (17)_1.png,2
 image (18)_0.png,5
 image (18)_1.png,4
 image (18)_2.png,1
 image (18)_3.png,1
 image (19)_0.png,2
 image (19)_1.png,2
 image (2)_0.png,4
 image (21)_0.png,3
 image (22)_0.png,3
 image (22)_1.png,2
 image (22)_2.png,5
 image (22)_3.png,3
 image (23)_0.png,3
 image (24)_0.png,2
 image (25)_0.png,3
 image (26)_0.png,3
 image (27)_0.png,4
 image (28)_0.png,3
 image (28)_1.png,2
 image (29)_0.png,5
 image (3)_0.png,5
 image (30)_0.png,5
 image (31)_0.png,3
 image (4)_0.png,4
 image (5)_0.png,3
 image (6)_0.png,5
 image (6)_1.png,4
 image (6)_2.png,4
 image (6)_3.png,4
 image (6)_4.png,4
 image (6)_5.png,3
 image (6)_6.png,5
 image (7)_0.png,5
 image (7)_1.png,4
 image (8)_0.png,4
 image (9)_0.png,5
 image_0_0.png,3
--- a/data/task/cni/inference_out/README.md
+++ b/data/task/cni/inference_out/README.md
@@ -0,0 +1,161 @@
 # Document Field Extraction Evaluation Results
 ## Overview
 This document presents the evaluation results for document field extraction using different preprocessing approaches. The evaluation was conducted on a dataset of 56 document samples with various field types commonly found in identity documents.
 ## Evaluation Metrics
 The evaluation uses standard information extraction metrics:
 - **Precision**: Ratio of correctly extracted fields to total extracted fields
 - **Recall**: Ratio of correctly extracted fields to total ground truth fields  
 - **F1-Score**: Harmonic mean of precision and recall
 - **Accuracy**: Overall field-level accuracy
 - **TP**: True Positives (correctly extracted fields)
 - **FP**: False Positives (incorrectly extracted fields)
 - **FN**: False Negatives (missed fields)
 ## Preprocessing Approaches
 ### 1. No Preprocessing (Baseline)
 - **Configuration**: Raw images without any preprocessing
 - **Performance**: 
  - Micro Precision: 79.0%
  - Micro Recall: 68.7%
  - Micro F1: 73.5%
  - Field Accuracy: 68.7%
 ### 2. Crop
 - **Configuration**: Content-aware cropping (no shadow removal)
 - **Performance**:
  - Micro Precision: 94.8%
  - Micro Recall: 89.9%
  - Micro F1: 92.3% (+18.8% improvement)
  - Field Accuracy: 89.9%
 ### 3. Crop + PaddleOCR + Shadow Removal
 - **Configuration**: Cropping with PaddleOCR document processing and shadow removal
 - **Performance**:
  - Micro Precision: 93.6%
  - Micro Recall: 89.4%
  - Micro F1: 91.5% (+18.0% improvement)
  - Field Accuracy: 89.4%
 ### 4. Crop + PaddleOCR + Shadow Removal + Cache
 - **Configuration**: Cropping with PaddleOCR, shadow removal, and caching
 - **Performance**:
  - Micro Precision: 92.5%
  - Micro Recall: 88.3%
  - Micro F1: 90.3% (+16.8% improvement)
  - Field Accuracy: 88.3%
 ### 5. Crop + Shadow Removal + Cache
 - **Configuration**: Cropping with shadow removal and caching
 - **Performance**:
  - Micro Precision: 93.6%
  - Micro Recall: 88.5%
  - Micro F1: 91.0% (+17.5% improvement)
  - Field Accuracy: 88.5%
 ## Field-Level Performance Analysis
 ### High-Performance Fields
 Fields that consistently perform well across all approaches:
 | Field | Best F1 | Best Approach | Performance Trend |
 |-------|----------|---------------|-------------------|
 | **Gender** | 85.1% | Crop + PaddleOCR | Consistent improvement |
 | **Birth Date** | 80.5% | Crop + PaddleOCR | Strong improvement |
 | **Document Type** | 85.4% | Crop + PaddleOCR | Significant improvement |
 | **Surname** | 82.9% | Crop + PaddleOCR | Consistent improvement |
 ### Medium-Performance Fields
 Fields with moderate improvement:
 | Field | Best F1 | Best Approach | Performance Trend |
 |-------|----------|---------------|-------------------|
 | **Birth Place** | 83.4% | Crop Only | Good improvement |
 | **Expiry Date** | 78.5% | Crop + PaddleOCR | Moderate improvement |
 | **Issue Date** | 69.3% | Crop + Shadow + Cache | Variable performance |
 | **Address** | 44.4% | Crop + PaddleOCR | Limited improvement |
 ### Low-Performance Fields
 Fields that remain challenging:
 | Field | Best F1 | Best Approach | Notes |
 |-------|----------|---------------|-------|
 | **MRZ Lines** | 41.8% | Crop + Shadow + Cache | Complex OCR patterns |
 | **Personal Number** | 40.0% | Crop + PaddleOCR + Cache | Small text, variable format |
 | **Issue Place** | 50.0% | Crop + PaddleOCR + Cache | Handwritten text challenges |
 ### Zero-Performance Fields
 Fields that consistently fail across all approaches:
 - **Recto/Verso**: Document side detection
 - **Code**: Encoded information
 - **Height**: Physical measurements
 - **Type**: Document classification
 ## Key Findings
 ### 1. Preprocessing Impact
 - **Cropping alone** delivers the strongest overall boost (+18.8 F1 pts vs. baseline)
 - **PaddleOCR + Shadow Removal** is highly competitive (up to +18.0 F1 pts)
 - **Caching** has minimal to moderate impact on accuracy
 ### 2. Field Type Sensitivity
 - **Structured fields** (dates, numbers) benefit most from preprocessing
 - **Text fields** (names, addresses) show moderate improvement
 - **Complex fields** (MRZ, codes) remain challenging
 ### 3. Processing Pipeline Efficiency
 - **Crop** currently provides the best overall F1 in this evaluation
 - **Crop + PaddleOCR + Shadow Removal** is close and benefits some fields
 - **Caching** shows minimal gains; use for speed, not accuracy
 ## Recommendations
 ### For Production Use
 1. **Use Crop** as the primary preprocessing step
 2. **Focus optimization** on high-value fields (dates, document types, names)
 3. **Consider field-specific** preprocessing strategies for challenging fields
 ### For Further Research
 1. **Investigate MRZ line** extraction techniques
 2. **Explore advanced OCR** methods for handwritten text
 3. **Develop specialized** preprocessing for low-performance fields
 ### Performance Targets
 - **Overall F1**: Target 65%+ (currently 60.7% best)
 - **Field Accuracy**: Target 50%+ (currently 43.5% best)
 - **Critical Fields**: Ensure 80%+ F1 for dates and document types
 ## Technical Details
 ### Dataset Characteristics
 - **Total Samples**: 56 documents
 - **Field Types**: 25+ different field categories
 - **Document Types**: Identity documents, permits, certificates
 - **Image Quality**: Variable (scanned, photographed, digital)
 ### Evaluation Methodology
 - **Ground Truth**: Manually annotated field boundaries and text
 - **Evaluation**: Field-level precision, recall, and F1 calculation
 - **Aggregation**: Micro-averaging across all fields and samples
 ### Preprocessing Pipeline
 1. **Image Input**: Raw document images
 2. **Cropping**: Content area detection and extraction
 3. **Document Processing**: PaddleOCR unwarping and orientation
 4. **Shadow Removal**: Optional DocShadow processing
 5. **Field Extraction**: OCR-based text extraction
 6. **Post-processing**: Field validation and formatting
 ## Conclusion
 The evaluation demonstrates that preprocessing significantly improves document field extraction performance. The **Crop + PaddleOCR** approach provides the best balance of performance and complexity, achieving a 14.1% improvement in F1-score over the baseline. While some fields remain challenging, the overall pipeline shows strong potential for production deployment with further field-specific optimizations.
 ---
 *Last Updated: August 2024*  
 *Evaluation Dataset: 56 document samples*  
 *Total Fields Evaluated: 900+ field instances*
--- a/data/task/cni/inference_out/SUMMARY.md
+++ b/data/task/cni/inference_out/SUMMARY.md
@@ -0,0 +1,37 @@
 # Evaluation Results Summary
 ## Quick Overview
 - **Dataset**: 56 document samples
 - **Best Approach**: Crop (No Shadow Removal)
 - **Performance Gain**: +14.1% F1-score improvement over baseline
 ## Performance Comparison (Ranked from Lowest to Highest)
 | Approach | Precision | Recall | F1-Score | Field Accuracy | Improvement vs. Baseline |
 |----------|-----------|--------|----------|----------------|---------------------------|
 | **No Preprocessing** | 79.0% | 68.7% | 73.5% | 68.7% | Baseline |
 | **Crop + PaddleOCR + Shadow Removal + Cache** | 92.5% | 88.3% | 90.3% | 88.3% | +16.8% |
 | **Crop + Shadow Removal + Cache** | 93.6% | 88.5% | 91.0% | 88.5% | +17.5% |
 | **Crop + PaddleOCR + Shadow Removal** | 93.6% | 89.4% | 91.5% | 89.4% | +18.0% |
 | **Crop** | 94.8% | 89.9% | 92.3% | 89.9% | +18.8% |
 ## Top Performing Fields
 - **Gender**: 85.1% F1 (Crop + PaddleOCR + Shadow Removal)
 - **Birth Date**: 80.5% F1 (Crop + PaddleOCR + Shadow Removal)
 - **Document Type**: 85.4% F1 (Crop + PaddleOCR + Shadow Removal)
 - **Surname**: 82.9% F1 (Crop + PaddleOCR + Shadow Removal)
 ## Key Insights
 1. **Cropping** provides the biggest performance boost
 2. **PaddleOCR + Shadow Removal** adds small but consistent improvement
 3. **Shadow removal** shows mixed results depending on field type
 4. **Caching** has minimal impact on accuracy
 ## Recommendations
 - Use **Crop + PaddleOCR + Shadow Removal** for production
 - Focus on optimizing high-value fields
 - Investigate MRZ line extraction further
 - Target 65%+ overall F1-score
 ---
 *See README.md for detailed analysis*
--- a/data/task/cni/inference_out/combined_anls/combined_per_image_anls.csv
+++ b/data/task/cni/inference_out/combined_anls/combined_per_image_anls.csv
@@ -0,0 +1,241 @@
 image,method,anls,hallucination_score,num_fields
 ITE55336627_PASSEPORT_Abasse_GUEYE__Completed_,crop,1.0,0.0,12
 ITE55340318_ID_kat__1__Completed_,crop,0.9200562169312168,0.0799437830687831,12
 ITE55340822_passeport_pascal_2028_Completed_,crop,1.0,0.0,12
 ITE55341271_C_I_face_Completed_,crop,0.802415458937198,0.197584541062802,10
 ITE55346966_17537774579547958575370370624241_Completed_,crop,0.8177777777777777,0.1822222222222222,5
 ITE55347865_Snapchat-715567440_Completed_,crop,0.9861111111111112,0.0138888888888888,8
 ITE55347866_Snapchat-1551171803_Completed_,crop,0.938471760797342,0.0615282392026579,5
 ITE55347926_DHONDT_CNI_,crop,0.9012345679012346,0.0987654320987654,9
 ITE55348016_image_Completed_,crop,0.975,0.025,8
 ITE55348464_IMG-20250721-WA0000_Completed_,crop,0.833044733044733,0.1669552669552669,11
 ITE55348878_CNI_RECTO_,crop,0.9637408088235294,0.0362591911764705,8
 ITE55349793_2_em_CNI_CHASTAING__,crop,0.988095238095238,0.0119047619047618,12
 ITE55350967_PLI_INCONNU_-_ADH_625302_42296806_DMS_,crop,0.9857142857142858,0.0142857142857142,7
 ITE55351087_PLI_INCONNU_-_ADH_625120_42296075_DMS_,crop,0.9611111111111112,0.0388888888888888,8
 ITE55351128_PLI_INCONNU_-_ADH_625123_42296103_DMS_,crop,0.9886574074074074,0.0113425925925926,12
 ITE55351156_Passeport_Completed_,crop,0.7908496732026145,0.2091503267973855,11
 ITE55352532_PLI_INCONNU_-_ADH_625374_42297046_DMS_,crop,1.0,0.0,6
 ITE55352984_CNI_Completed_,crop,0.9977477477477475,0.0022522522522523,12
 ITE55354550_20250728_113359_Completed_,crop,0.8884615384615384,0.1115384615384615,8
 ITE55354891_image_Completed_,crop,0.975,0.025,8
 image (1)_,crop,0.9979166666666668,0.0020833333333333,12
 image (11)_,crop,0.9194444444444444,0.0805555555555556,12
 image (12)_,crop,0.8828703703703704,0.1171296296296295,12
 image (13)_,crop,0.9275462962962964,0.0724537037037036,12
 image (14)_,crop,0.9,0.0999999999999999,4
 image (15)_,crop,0.9012345679012346,0.0987654320987654,9
 image (16)_,crop,0.8765432098765432,0.1234567901234567,9
 image (17)_,crop,0.9882154882154882,0.0117845117845117,12
 image (18)_,crop,0.9170777542870566,0.0829222457129433,11
 image (19)_,crop,0.9027777777777778,0.0972222222222222,12
 image (2)_,crop,0.0625,0.9375,1
 image (21)_,crop,0.8056140350877193,0.1943859649122806,5
 image (22)_,crop,0.995246913580247,0.004753086419753,12
 image (23)_,crop,0.7727842809364549,0.2272157190635451,8
 image (24)_,crop,0.0461538461538461,0.953846153846154,5
 image (25)_,crop,1.0,0.0,8
 image (26)_,crop,0.8111111111111111,0.1888888888888888,10
 image (27)_,crop,0.8666666666666667,0.1333333333333333,9
 image (28)_,crop,0.9666666666666668,0.0333333333333333,8
 image (29)_,crop,0.9122007722007724,0.0877992277992276,10
 image (3)_,crop,0.7485260770975058,0.2514739229024941,7
 image (30)_,crop,0.8637037037037036,0.1362962962962963,10
 image (31)_,crop,0.8541666666666666,0.1458333333333333,12
 image (4)_,crop,0.8035714285714286,0.1964285714285714,8
 image (6)_,crop,0.0,0.999999999,12
 image (8)_,crop,0.9818452380952382,0.0181547619047618,12
 image (9)_,crop,0.858152958152958,0.141847041847042,11
 image_,crop,0.8542229012817248,0.1457770987182751,7
 ITE55336627_PASSEPORT_Abasse_GUEYE__Completed_,crop_shadow_cache,0.9965277777777778,0.0034722222222222,12
 ITE55340318_ID_kat__1__Completed_,crop_shadow_cache,0.994498556998557,0.0055014430014429,12
 ITE55340822_passeport_pascal_2028_Completed_,crop_shadow_cache,0.8776094276094275,0.1223905723905725,12
 ITE55341271_C_I_face_Completed_,crop_shadow_cache,0.802415458937198,0.197584541062802,10
 ITE55346966_17537774579547958575370370624241_Completed_,crop_shadow_cache,0.6485470085470085,0.3514529914529914,5
 ITE55347865_Snapchat-715567440_Completed_,crop_shadow_cache,0.7700191570881226,0.2299808429118773,8
 ITE55347866_Snapchat-1551171803_Completed_,crop_shadow_cache,0.689240991566573,0.3107590084334269,5
 ITE55347926_DHONDT_CNI_,crop_shadow_cache,0.9012345679012346,0.0987654320987654,9
 ITE55348016_image_Completed_,crop_shadow_cache,0.975,0.025,8
 ITE55348464_IMG-20250721-WA0000_Completed_,crop_shadow_cache,0.8239538239538239,0.1760461760461761,11
 ITE55348878_CNI_RECTO_,crop_shadow_cache,0.9637408088235294,0.0362591911764705,8
 ITE55349793_2_em_CNI_CHASTAING__,crop_shadow_cache,0.988095238095238,0.0119047619047618,12
 ITE55350967_PLI_INCONNU_-_ADH_625302_42296806_DMS_,crop_shadow_cache,0.8428571428571429,0.1571428571428571,7
 ITE55351087_PLI_INCONNU_-_ADH_625120_42296075_DMS_,crop_shadow_cache,0.961111111111111,0.0388888888888889,8
 ITE55351128_PLI_INCONNU_-_ADH_625123_42296103_DMS_,crop_shadow_cache,0.9861111111111112,0.0138888888888888,12
 ITE55351156_Passeport_Completed_,crop_shadow_cache,0.7908496732026145,0.2091503267973855,11
 ITE55352532_PLI_INCONNU_-_ADH_625374_42297046_DMS_,crop_shadow_cache,1.0,0.0,6
 ITE55352984_CNI_Completed_,crop_shadow_cache,0.9886363636363636,0.0113636363636363,12
 ITE55354550_20250728_113359_Completed_,crop_shadow_cache,0.975,0.025,8
 ITE55354891_image_Completed_,crop_shadow_cache,0.975,0.025,8
 image (1)_,crop_shadow_cache,0.9979166666666668,0.0020833333333333,12
 image (11)_,crop_shadow_cache,0.9244623655913978,0.0755376344086021,12
 image (12)_,crop_shadow_cache,0.949537037037037,0.050462962962963,12
 image (13)_,crop_shadow_cache,0.980324074074074,0.0196759259259259,12
 image (14)_,crop_shadow_cache,0.9,0.0999999999999999,4
 image (15)_,crop_shadow_cache,0.9012345679012346,0.0987654320987654,9
 image (16)_,crop_shadow_cache,0.8765432098765432,0.1234567901234567,9
 image (17)_,crop_shadow_cache,0.9856902356902356,0.0143097643097642,12
 image (18)_,crop_shadow_cache,0.9170777542870566,0.0829222457129433,11
 image (19)_,crop_shadow_cache,0.9907407407407408,0.0092592592592591,12
 image (2)_,crop_shadow_cache,0.0625,0.9375,1
 image (21)_,crop_shadow_cache,0.5623423423423424,0.4376576576576576,5
 image (22)_,crop_shadow_cache,0.995246913580247,0.004753086419753,12
 image (23)_,crop_shadow_cache,0.7727842809364549,0.2272157190635451,8
 image (24)_,crop_shadow_cache,0.0461538461538461,0.953846153846154,5
 image (25)_,crop_shadow_cache,1.0,0.0,8
 image (26)_,crop_shadow_cache,0.7252287581699346,0.2747712418300654,10
 image (27)_,crop_shadow_cache,0.8666666666666667,0.1333333333333333,9
 image (28)_,crop_shadow_cache,0.7444444444444445,0.2555555555555555,8
 image (29)_,crop_shadow_cache,0.8726839826839827,0.1273160173160172,10
 image (3)_,crop_shadow_cache,0.7539682539682538,0.2460317460317461,7
 image (30)_,crop_shadow_cache,0.8637037037037036,0.1362962962962963,10
 image (31)_,crop_shadow_cache,0.8631205673758865,0.1368794326241135,12
 image (4)_,crop_shadow_cache,0.8427083333333334,0.1572916666666666,8
 image (6)_,crop_shadow_cache,0.0,0.999999999,12
 image (8)_,crop_shadow_cache,0.9339285714285714,0.0660714285714285,12
 image (9)_,crop_shadow_cache,0.858152958152958,0.141847041847042,11
 image_,crop_shadow_cache,0.8381574852163087,0.1618425147836912,7
 ITE55336627_PASSEPORT_Abasse_GUEYE__Completed_,crop_shadow_paddle,1.0,0.0,12
 ITE55340318_ID_kat__1__Completed_,crop_shadow_paddle,0.9858630952380952,0.0141369047619047,12
 ITE55340822_passeport_pascal_2028_Completed_,crop_shadow_paddle,0.9924242424242424,0.0075757575757576,12
 ITE55341271_C_I_face_Completed_,crop_shadow_paddle,0.8111111111111111,0.1888888888888888,10
 ITE55346966_17537774579547958575370370624241_Completed_,crop_shadow_paddle,0.9155555555555556,0.0844444444444445,5
 ITE55347865_Snapchat-715567440_Completed_,crop_shadow_paddle,0.8394636015325672,0.1605363984674328,8
 ITE55347866_Snapchat-1551171803_Completed_,crop_shadow_paddle,0.8124916943521594,0.1875083056478406,5
 ITE55347926_DHONDT_CNI_,crop_shadow_paddle,0.6790123456790123,0.3209876543209877,9
 ITE55348016_image_Completed_,crop_shadow_paddle,0.975,0.025,8
 ITE55348464_IMG-20250721-WA0000_Completed_,crop_shadow_paddle,0.833044733044733,0.1669552669552669,11
 ITE55348878_CNI_RECTO_,crop_shadow_paddle,0.8894060749299719,0.110593925070028,8
 ITE55349793_2_em_CNI_CHASTAING__,crop_shadow_paddle,0.988095238095238,0.0119047619047618,12
 ITE55350967_PLI_INCONNU_-_ADH_625302_42296806_DMS_,crop_shadow_paddle,0.9000000000000001,0.0999999999999998,7
 ITE55351087_PLI_INCONNU_-_ADH_625120_42296075_DMS_,crop_shadow_paddle,0.9861111111111112,0.0138888888888888,8
 ITE55351128_PLI_INCONNU_-_ADH_625123_42296103_DMS_,crop_shadow_paddle,0.9619522084195998,0.0380477915804001,12
 ITE55351156_Passeport_Completed_,crop_shadow_paddle,0.7908496732026145,0.2091503267973855,11
 ITE55352532_PLI_INCONNU_-_ADH_625374_42297046_DMS_,crop_shadow_paddle,0.9166666666666666,0.0833333333333333,6
 ITE55352984_CNI_Completed_,crop_shadow_paddle,0.8507575757575757,0.1492424242424243,12
 ITE55354550_20250728_113359_Completed_,crop_shadow_paddle,0.9514957264957264,0.0485042735042735,8
 ITE55354891_image_Completed_,crop_shadow_paddle,0.975,0.025,8
 image (1)_,crop_shadow_paddle,0.9979166666666668,0.0020833333333333,12
 image (11)_,crop_shadow_paddle,0.9239583333333332,0.0760416666666666,12
 image (12)_,crop_shadow_paddle,0.8632066276803118,0.1367933723196882,12
 image (13)_,crop_shadow_paddle,0.987962962962963,0.0120370370370369,12
 image (14)_,crop_shadow_paddle,0.9,0.0999999999999999,4
 image (15)_,crop_shadow_paddle,0.9012345679012346,0.0987654320987654,9
 image (16)_,crop_shadow_paddle,0.8765432098765432,0.1234567901234567,9
 image (17)_,crop_shadow_paddle,0.9882154882154882,0.0117845117845117,12
 image (18)_,crop_shadow_paddle,0.9079868451961476,0.0920131548038525,11
 image (19)_,crop_shadow_paddle,0.9717967047930284,0.0282032952069716,12
 image (2)_,crop_shadow_paddle,0.0625,0.9375,1
 image (21)_,crop_shadow_paddle,0.670990990990991,0.329009009009009,5
 image (22)_,crop_shadow_paddle,0.9969135802469136,0.0030864197530863,12
 image (23)_,crop_shadow_paddle,0.9891304347826086,0.0108695652173913,8
 image (24)_,crop_shadow_paddle,0.5843928398645379,0.415607160135462,5
 image (25)_,crop_shadow_paddle,0.9107142857142856,0.0892857142857143,8
 image (26)_,crop_shadow_paddle,0.8711111111111111,0.1288888888888889,10
 image (27)_,crop_shadow_paddle,0.8666666666666667,0.1333333333333333,9
 image (28)_,crop_shadow_paddle,0.9916666666666668,0.0083333333333333,8
 image (29)_,crop_shadow_paddle,0.7705994897959184,0.2294005102040816,10
 image (3)_,crop_shadow_paddle,0.6199546485260772,0.3800453514739228,7
 image (30)_,crop_shadow_paddle,0.8711111111111111,0.1288888888888889,10
 image (31)_,crop_shadow_paddle,0.9249999999999998,0.0750000000000001,12
 image (4)_,crop_shadow_paddle,0.9285714285714286,0.0714285714285714,8
 image (6)_,crop_shadow_paddle,0.0,0.999999999,12
 image (8)_,crop_shadow_paddle,0.9898809523809524,0.0101190476190476,12
 image (9)_,crop_shadow_paddle,0.8595959595959596,0.1404040404040404,11
 image_,crop_shadow_paddle,0.8256514727102964,0.1743485272897036,7
 ITE55336627_PASSEPORT_Abasse_GUEYE__Completed_,crop_shadow_paddle_cache,1.0,0.0,12
 ITE55340318_ID_kat__1__Completed_,crop_shadow_paddle_cache,0.9796176046176046,0.0203823953823953,12
 ITE55340822_passeport_pascal_2028_Completed_,crop_shadow_paddle_cache,1.0,0.0,12
 ITE55341271_C_I_face_Completed_,crop_shadow_paddle_cache,0.711111111111111,0.288888888888889,10
 ITE55346966_17537774579547958575370370624241_Completed_,crop_shadow_paddle_cache,0.8177777777777777,0.1822222222222222,5
 ITE55347865_Snapchat-715567440_Completed_,crop_shadow_paddle_cache,0.8255747126436782,0.1744252873563218,8
 ITE55347866_Snapchat-1551171803_Completed_,crop_shadow_paddle_cache,0.8631229235880399,0.1368770764119601,5
 ITE55347926_DHONDT_CNI_,crop_shadow_paddle_cache,0.7364197530864198,0.2635802469135802,9
 ITE55348016_image_Completed_,crop_shadow_paddle_cache,0.975,0.025,8
 ITE55348464_IMG-20250721-WA0000_Completed_,crop_shadow_paddle_cache,0.7173082173082174,0.2826917826917826,11
 ITE55348878_CNI_RECTO_,crop_shadow_paddle_cache,0.8308648459383754,0.1691351540616246,8
 ITE55349793_2_em_CNI_CHASTAING__,crop_shadow_paddle_cache,0.988095238095238,0.0119047619047618,12
 ITE55350967_PLI_INCONNU_-_ADH_625302_42296806_DMS_,crop_shadow_paddle_cache,0.976023976023976,0.0239760239760239,7
 ITE55351087_PLI_INCONNU_-_ADH_625120_42296075_DMS_,crop_shadow_paddle_cache,1.0,0.0,8
 ITE55351128_PLI_INCONNU_-_ADH_625123_42296103_DMS_,crop_shadow_paddle_cache,0.9555555555555556,0.0444444444444444,12
 ITE55351156_Passeport_Completed_,crop_shadow_paddle_cache,0.8817587641117054,0.1182412358882946,11
 ITE55352532_PLI_INCONNU_-_ADH_625374_42297046_DMS_,crop_shadow_paddle_cache,1.0,0.0,6
 ITE55352984_CNI_Completed_,crop_shadow_paddle_cache,0.9499999999999998,0.0500000000000001,12
 ITE55354550_20250728_113359_Completed_,crop_shadow_paddle_cache,0.961111111111111,0.0388888888888889,8
 ITE55354891_image_Completed_,crop_shadow_paddle_cache,0.975,0.025,8
 image (1)_,crop_shadow_paddle_cache,0.9979166666666668,0.0020833333333333,12
 image (11)_,crop_shadow_paddle_cache,0.9046875,0.0953125,12
 image (12)_,crop_shadow_paddle_cache,0.8842592592592592,0.1157407407407408,12
 image (13)_,crop_shadow_paddle_cache,0.9275462962962964,0.0724537037037036,12
 image (14)_,crop_shadow_paddle_cache,0.9,0.0999999999999999,4
 image (15)_,crop_shadow_paddle_cache,0.9012345679012346,0.0987654320987654,9
 image (16)_,crop_shadow_paddle_cache,0.854320987654321,0.145679012345679,9
 image (17)_,crop_shadow_paddle_cache,0.9882154882154882,0.0117845117845117,12
 image (18)_,crop_shadow_paddle_cache,0.9170777542870566,0.0829222457129433,11
 image (19)_,crop_shadow_paddle_cache,0.9907407407407408,0.0092592592592591,12
 image (2)_,crop_shadow_paddle_cache,0.0625,0.9375,1
 image (21)_,crop_shadow_paddle_cache,0.670990990990991,0.329009009009009,5
 image (22)_,crop_shadow_paddle_cache,0.9969135802469136,0.0030864197530863,12
 image (23)_,crop_shadow_paddle_cache,0.9891304347826086,0.0108695652173913,8
 image (24)_,crop_shadow_paddle_cache,0.2348330914368649,0.765166908563135,5
 image (25)_,crop_shadow_paddle_cache,0.9107142857142856,0.0892857142857143,8
 image (26)_,crop_shadow_paddle_cache,0.82,0.18,10
 image (27)_,crop_shadow_paddle_cache,0.8666666666666667,0.1333333333333333,9
 image (28)_,crop_shadow_paddle_cache,1.0,0.0,8
 image (29)_,crop_shadow_paddle_cache,0.9726839826839828,0.0273160173160172,10
 image (3)_,crop_shadow_paddle_cache,0.609750566893424,0.390249433106576,7
 image (30)_,crop_shadow_paddle_cache,0.8711111111111111,0.1288888888888889,10
 image (31)_,crop_shadow_paddle_cache,0.6940972222222221,0.3059027777777778,12
 image (4)_,crop_shadow_paddle_cache,0.9035714285714286,0.0964285714285714,8
 image (6)_,crop_shadow_paddle_cache,0.0,0.999999999,12
 image (8)_,crop_shadow_paddle_cache,0.9825091575091576,0.0174908424908424,12
 image (9)_,crop_shadow_paddle_cache,0.853102453102453,0.1468975468975469,11
 image_,crop_shadow_paddle_cache,0.8408029878618114,0.1591970121381886,7
 image (22),no_preprocessing,0.82125,0.1787499999999999,12
 image (16),no_preprocessing,0.8765432098765432,0.1234567901234567,9
 image (18),no_preprocessing,0.4974404148822753,0.5025595851177247,11
 image (3),no_preprocessing,0.7913832199546486,0.2086167800453514,7
 ITE55348016_image_Completed,no_preprocessing,0.7644230769230769,0.2355769230769231,8
 ITE55348878_CNI_RECTO,no_preprocessing,0.8073553599071208,0.1926446400928791,8
 image (26),no_preprocessing,0.7659498207885305,0.2340501792114695,10
 image (28),no_preprocessing,0.7822072072072072,0.2177927927927928,8
 ITE55347926_DHONDT_CNI,no_preprocessing,0.6816239316239316,0.3183760683760683,9
 image (15),no_preprocessing,0.9012345679012346,0.0987654320987654,9
 ITE55354891_image_Completed,no_preprocessing,0.6394230769230769,0.3605769230769231,8
 image (4),no_preprocessing,0.975,0.025,8
 ITE55351156_Passeport_Completed,no_preprocessing,0.8813131313131314,0.1186868686868686,11
 ITE55347866_Snapchat-1551171803_Completed,no_preprocessing,0.5564221824686941,0.4435778175313059,5
 ITE55341271_C_I_face_Completed,no_preprocessing,0.5861416361416361,0.4138583638583639,10
 ITE55352532_PLI_INCONNU_-_ADH_625374_42297046_DMS,no_preprocessing,0.6481481481481481,0.3518518518518518,6
 image (9),no_preprocessing,0.914141414141414,0.0858585858585858,11
 image (25),no_preprocessing,0.8667929292929293,0.1332070707070707,8
 ITE55352984_CNI_Completed,no_preprocessing,0.7075757575757576,0.2924242424242423,12
 ITE55351087_PLI_INCONNU_-_ADH_625120_42296075_DMS,no_preprocessing,0.9527777777777776,0.0472222222222222,8
 ITE55349793_2_em_CNI_CHASTAING_,no_preprocessing,0.938095238095238,0.061904761904762,12
 ITE55354550_20250728_113359_Completed,no_preprocessing,0.5625,0.4375,8
 ITE55340318_ID_kat__1__Completed,no_preprocessing,0.3825320512820513,0.6174679487179486,12
 ITE55340822_passeport_pascal_2028_Completed,no_preprocessing,0.8164983164983166,0.1835016835016834,12
 ITE55351128_PLI_INCONNU_-_ADH_625123_42296103_DMS,no_preprocessing,0.6349206349206349,0.3650793650793651,12
 image (2),no_preprocessing,0.0625,0.9375,1
 image (21),no_preprocessing,0.7452631578947368,0.2547368421052631,5
 image (12),no_preprocessing,0.8540820231996702,0.1459179768003298,12
 image (24),no_preprocessing,0.0461538461538461,0.953846153846154,5
 image (31),no_preprocessing,0.8666666666666666,0.1333333333333334,12
 ITE55347865_Snapchat-715567440_Completed,no_preprocessing,0.5144230769230769,0.4855769230769231,8
 image (30),no_preprocessing,0.8037037037037036,0.1962962962962964,10
 image (27),no_preprocessing,0.8888888888888888,0.1111111111111111,9
 ITE55348464_IMG-20250721-WA0000_Completed,no_preprocessing,0.8966810966810965,0.1033189033189034,11
 image (19),no_preprocessing,0.6095515276549759,0.3904484723450241,12
 ITE55346966_17537774579547958575370370624241_Completed,no_preprocessing,0.4266365007541478,0.5733634992458522,5
 image (8),no_preprocessing,0.6822344322344321,0.3177655677655678,12
 image (13),no_preprocessing,0.9907407407407408,0.0092592592592591,12
 image (23),no_preprocessing,0.0288461538461538,0.971153846153846,8
 image (1),no_preprocessing,0.6418574481074482,0.3581425518925518,12
 image (14),no_preprocessing,1.0,0.0,4
 image_,no_preprocessing,0.7751322751322751,0.2248677248677248,7
 image (11),no_preprocessing,0.3424526862026862,0.6575473137973138,12
 image (6),no_preprocessing,0.9325913349682452,0.0674086650317549,12
 image (29),no_preprocessing,0.8412925170068026,0.1587074829931973,10
 ITE55336627_PASSEPORT_Abasse_GUEYE__Completed,no_preprocessing,1.0,0.0,12
 image (17),no_preprocessing,0.6437516187516187,0.3562483812483813,12
 ITE55350967_PLI_INCONNU_-_ADH_625302_42296806_DMS,no_preprocessing,0.1428571428571428,0.8571428571428572,7
--- a/data/task/cni/inference_out/combined_anls/pipeline/pipeline_means.csv
+++ b/data/task/cni/inference_out/combined_anls/pipeline/pipeline_means.csv
@@ -0,0 +1,6 @@
 method,raw_mean,deqa_mean,human_mean,raw_count,deqa_count,human_count
 crop,0.14311980666931925,0.113280480384857,0.10987469285842993,48,48,48
 crop_shadow_cache,0.16717154207835927,0.12110082758738305,0.11779958008485086,48,48,48
 crop_shadow_paddle,0.14359221458380786,0.09942455054893862,0.09592340843496948,48,48,48
 crop_shadow_paddle_cache,0.15229752473367092,0.11032137012597554,0.10660814625695587,48,48,48
 no_preprocessing,0.3023333345414223,0.1630725097903042,0.13890361606290755,48,48,48
--- a/data/task/cni/inference_out/crop/field_metrics.csv
+++ b/data/task/cni/inference_out/crop/field_metrics.csv
@@ -0,0 +1,16 @@
 field,tp,fp,fn,precision,recall,f1,accuracy
 address,21,4,6,0.840000,0.777778,0.807692,0.777778
 birth_date,35,3,4,0.921053,0.897436,0.909091,0.897436
 birth_place,34,0,1,1.000000,0.971429,0.985507,0.971429
 doc_number,33,2,3,0.942857,0.916667,0.929577,0.916667
 document_type,43,4,5,0.914894,0.895833,0.905263,0.895833
 expiry_date,31,3,5,0.911765,0.861111,0.885714,0.861111
 gender,37,0,1,1.000000,0.973684,0.986667,0.973684
 issue_date,26,3,5,0.896552,0.838710,0.866667,0.838710
 issue_place,18,2,4,0.900000,0.818182,0.857143,0.818182
 name,40,1,2,0.975610,0.952381,0.963855,0.952381
 nationality,39,0,1,1.000000,0.975000,0.987342,0.975000
 permit_type,1,0,0,1.000000,1.000000,1.000000,1.000000
 personal_number,1,4,5,0.200000,0.166667,0.181818,0.166667
 remarks,1,0,0,1.000000,1.000000,1.000000,1.000000
 surname,39,2,3,0.951220,0.928571,0.939759,0.928571
--- a/data/task/cni/inference_out/crop/per_image_anls.csv
+++ b/data/task/cni/inference_out/crop/per_image_anls.csv
@@ -0,0 +1,49 @@
 image,anls,hallucination_score,num_fields
 ITE55336627_PASSEPORT_Abasse_GUEYE__Completed_,1.0,0.0,12
 ITE55340318_ID_kat__1__Completed_,0.9200562169312169,0.07994378306878314,12
 ITE55340822_passeport_pascal_2028_Completed_,1.0,0.0,12
 ITE55341271_C_I_face_Completed_,0.802415458937198,0.197584541062802,10
 ITE55346966_17537774579547958575370370624241_Completed_,0.8177777777777777,0.18222222222222229,5
 ITE55347865_Snapchat-715567440_Completed_,0.9861111111111112,0.01388888888888884,8
 ITE55347866_Snapchat-1551171803_Completed_,0.9384717607973421,0.06152823920265793,5
 ITE55347926_DHONDT_CNI_,0.9012345679012346,0.09876543209876543,9
 ITE55348016_image_Completed_,0.975,0.025000000000000022,8
 ITE55348464_IMG-20250721-WA0000_Completed_,0.833044733044733,0.16695526695526697,11
 ITE55348878_CNI_RECTO_,0.9637408088235294,0.036259191176470584,8
 ITE55349793_2_em_CNI_CHASTAING__,0.9880952380952381,0.011904761904761862,12
 ITE55350967_PLI_INCONNU_-_ADH_625302_42296806_DMS_,0.9857142857142858,0.014285714285714235,7
 ITE55351087_PLI_INCONNU_-_ADH_625120_42296075_DMS_,0.9611111111111111,0.03888888888888886,8
 ITE55351128_PLI_INCONNU_-_ADH_625123_42296103_DMS_,0.9886574074074074,0.011342592592592626,12
 ITE55351156_Passeport_Completed_,0.7908496732026145,0.20915032679738554,11
 ITE55352532_PLI_INCONNU_-_ADH_625374_42297046_DMS_,1.0,0.0,6
 ITE55352984_CNI_Completed_,0.9977477477477477,0.0022522522522523403,12
 ITE55354550_20250728_113359_Completed_,0.8884615384615384,0.11153846153846159,8
 ITE55354891_image_Completed_,0.975,0.025000000000000022,8
 image (1)_,0.9979166666666667,0.002083333333333326,12
 image (11)_,0.9194444444444444,0.0805555555555556,12
 image (12)_,0.8828703703703704,0.11712962962962958,12
 image (13)_,0.9275462962962964,0.07245370370370363,12
 image (14)_,0.9,0.09999999999999998,4
 image (15)_,0.9012345679012346,0.09876543209876543,9
 image (16)_,0.8765432098765432,0.12345679012345678,9
 image (17)_,0.9882154882154882,0.011784511784511786,12
 image (18)_,0.9170777542870566,0.08292224571294338,11
 image (19)_,0.9027777777777778,0.09722222222222221,12
 image (2)_,0.0625,0.9375,1
 image (21)_,0.8056140350877193,0.19438596491228066,5
 image (22)_,0.995246913580247,0.004753086419753028,12
 image (23)_,0.7727842809364549,0.2272157190635451,8
 image (24)_,0.04615384615384614,0.9538461538461539,5
 image (25)_,1.0,0.0,8
 image (26)_,0.8111111111111111,0.18888888888888888,10
 image (27)_,0.8666666666666667,0.1333333333333333,9
 image (28)_,0.9666666666666667,0.033333333333333326,8
 image (29)_,0.9122007722007723,0.08779922779922766,10
 image (3)_,0.7485260770975058,0.25147392290249415,7
 image (30)_,0.8637037037037036,0.13629629629629636,10
 image (31)_,0.8541666666666666,0.14583333333333337,12
 image (4)_,0.8035714285714286,0.1964285714285714,8
 image (6)_,0.0,0.999999999,12
 image (8)_,0.9818452380952382,0.01815476190476184,12
 image (9)_,0.858152958152958,0.141847041847042,11
 image_,0.8542229012817248,0.14577709871827516,7
--- a/data/task/cni/inference_out/crop_shadow_cache/field_metrics.csv
+++ b/data/task/cni/inference_out/crop_shadow_cache/field_metrics.csv
@@ -0,0 +1,16 @@
 field,tp,fp,fn,precision,recall,f1,accuracy
 address,20,5,7,0.800000,0.740741,0.769231,0.740741
 birth_date,35,3,4,0.921053,0.897436,0.909091,0.897436
 birth_place,33,1,2,0.970588,0.942857,0.956522,0.942857
 doc_number,33,2,3,0.942857,0.916667,0.929577,0.916667
 document_type,41,6,7,0.872340,0.854167,0.863158,0.854167
 expiry_date,31,3,5,0.911765,0.861111,0.885714,0.861111
 gender,36,1,2,0.972973,0.947368,0.960000,0.947368
 issue_date,27,2,4,0.931034,0.870968,0.900000,0.870968
 issue_place,17,3,5,0.850000,0.772727,0.809524,0.772727
 name,41,0,1,1.000000,0.976190,0.987952,0.976190
 nationality,38,1,2,0.974359,0.950000,0.962025,0.950000
 permit_type,1,0,0,1.000000,1.000000,1.000000,1.000000
 personal_number,1,4,5,0.200000,0.166667,0.181818,0.166667
 remarks,1,0,0,1.000000,1.000000,1.000000,1.000000
 surname,38,3,4,0.926829,0.904762,0.915663,0.904762
--- a/data/task/cni/inference_out/crop_shadow_cache/per_image_anls.csv
+++ b/data/task/cni/inference_out/crop_shadow_cache/per_image_anls.csv
@@ -0,0 +1,49 @@
 image,anls,hallucination_score,num_fields
 ITE55336627_PASSEPORT_Abasse_GUEYE__Completed_,0.9965277777777778,0.00347222222222221,12
 ITE55340318_ID_kat__1__Completed_,0.994498556998557,0.005501443001442952,12
 ITE55340822_passeport_pascal_2028_Completed_,0.8776094276094275,0.1223905723905725,12
 ITE55341271_C_I_face_Completed_,0.802415458937198,0.197584541062802,10
 ITE55346966_17537774579547958575370370624241_Completed_,0.6485470085470085,0.35145299145299147,5
 ITE55347865_Snapchat-715567440_Completed_,0.7700191570881226,0.22998084291187737,8
 ITE55347866_Snapchat-1551171803_Completed_,0.689240991566573,0.31075900843342696,5
 ITE55347926_DHONDT_CNI_,0.9012345679012346,0.09876543209876543,9
 ITE55348016_image_Completed_,0.975,0.025000000000000022,8
 ITE55348464_IMG-20250721-WA0000_Completed_,0.8239538239538239,0.17604617604617612,11
 ITE55348878_CNI_RECTO_,0.9637408088235294,0.036259191176470584,8
 ITE55349793_2_em_CNI_CHASTAING__,0.9880952380952381,0.011904761904761862,12
 ITE55350967_PLI_INCONNU_-_ADH_625302_42296806_DMS_,0.8428571428571429,0.15714285714285714,7
 ITE55351087_PLI_INCONNU_-_ADH_625120_42296075_DMS_,0.961111111111111,0.03888888888888897,8
 ITE55351128_PLI_INCONNU_-_ADH_625123_42296103_DMS_,0.9861111111111112,0.01388888888888884,12
 ITE55351156_Passeport_Completed_,0.7908496732026145,0.20915032679738554,11
 ITE55352532_PLI_INCONNU_-_ADH_625374_42297046_DMS_,1.0,0.0,6
 ITE55352984_CNI_Completed_,0.9886363636363636,0.011363636363636354,12
 ITE55354550_20250728_113359_Completed_,0.975,0.025000000000000022,8
 ITE55354891_image_Completed_,0.975,0.025000000000000022,8
 image (1)_,0.9979166666666667,0.002083333333333326,12
 image (11)_,0.9244623655913978,0.07553763440860217,12
 image (12)_,0.949537037037037,0.05046296296296304,12
 image (13)_,0.9803240740740741,0.01967592592592593,12
 image (14)_,0.9,0.09999999999999998,4
 image (15)_,0.9012345679012346,0.09876543209876543,9
 image (16)_,0.8765432098765432,0.12345679012345678,9
 image (17)_,0.9856902356902357,0.014309764309764272,12
 image (18)_,0.9170777542870566,0.08292224571294338,11
 image (19)_,0.9907407407407408,0.00925925925925919,12
 image (2)_,0.0625,0.9375,1
 image (21)_,0.5623423423423424,0.43765765765765763,5
 image (22)_,0.995246913580247,0.004753086419753028,12
 image (23)_,0.7727842809364549,0.2272157190635451,8
 image (24)_,0.04615384615384614,0.9538461538461539,5
 image (25)_,1.0,0.0,8
 image (26)_,0.7252287581699346,0.2747712418300654,10
 image (27)_,0.8666666666666667,0.1333333333333333,9
 image (28)_,0.7444444444444445,0.25555555555555554,8
 image (29)_,0.8726839826839827,0.12731601731601727,10
 image (3)_,0.7539682539682538,0.24603174603174616,7
 image (30)_,0.8637037037037036,0.13629629629629636,10
 image (31)_,0.8631205673758865,0.1368794326241135,12
 image (4)_,0.8427083333333334,0.1572916666666666,8
 image (6)_,0.0,0.999999999,12
 image (8)_,0.9339285714285714,0.06607142857142856,12
 image (9)_,0.858152958152958,0.141847041847042,11
 image_,0.8381574852163087,0.16184251478369127,7
--- a/data/task/cni/inference_out/crop_shadow_paddle/field_metrics.csv
+++ b/data/task/cni/inference_out/crop_shadow_paddle/field_metrics.csv
@@ -0,0 +1,16 @@
 field,tp,fp,fn,precision,recall,f1,accuracy
 address,21,5,6,0.807692,0.777778,0.792453,0.777778
 birth_date,36,2,3,0.947368,0.923077,0.935065,0.923077
 birth_place,34,0,1,1.000000,0.971429,0.985507,0.971429
 doc_number,32,3,4,0.914286,0.888889,0.901408,0.888889
 document_type,44,3,4,0.936170,0.916667,0.926316,0.916667
 expiry_date,32,2,4,0.941176,0.888889,0.914286,0.888889
 gender,37,0,1,1.000000,0.973684,0.986667,0.973684
 issue_date,24,6,7,0.800000,0.774194,0.786885,0.774194
 issue_place,19,2,3,0.904762,0.863636,0.883721,0.863636
 name,39,2,3,0.951220,0.928571,0.939759,0.928571
 nationality,39,0,1,1.000000,0.975000,0.987342,0.975000
 permit_type,1,0,0,1.000000,1.000000,1.000000,1.000000
 personal_number,2,3,4,0.400000,0.333333,0.363636,0.333333
 remarks,0,1,1,0.000000,0.000000,0.000000,0.000000
 surname,37,4,5,0.902439,0.880952,0.891566,0.880952
--- a/data/task/cni/inference_out/crop_shadow_paddle/per_image_anls.csv
+++ b/data/task/cni/inference_out/crop_shadow_paddle/per_image_anls.csv
@@ -0,0 +1,49 @@
 image,anls,hallucination_score,num_fields
 ITE55336627_PASSEPORT_Abasse_GUEYE__Completed_,1.0,0.0,12
 ITE55340318_ID_kat__1__Completed_,0.9858630952380952,0.014136904761904767,12
 ITE55340822_passeport_pascal_2028_Completed_,0.9924242424242423,0.00757575757575768,12
 ITE55341271_C_I_face_Completed_,0.8111111111111111,0.18888888888888888,10
 ITE55346966_17537774579547958575370370624241_Completed_,0.9155555555555555,0.08444444444444454,5
 ITE55347865_Snapchat-715567440_Completed_,0.8394636015325672,0.16053639846743284,8
 ITE55347866_Snapchat-1551171803_Completed_,0.8124916943521594,0.1875083056478406,5
 ITE55347926_DHONDT_CNI_,0.6790123456790123,0.32098765432098775,9
 ITE55348016_image_Completed_,0.975,0.025000000000000022,8
 ITE55348464_IMG-20250721-WA0000_Completed_,0.833044733044733,0.16695526695526697,11
 ITE55348878_CNI_RECTO_,0.8894060749299719,0.11059392507002808,8
 ITE55349793_2_em_CNI_CHASTAING__,0.9880952380952381,0.011904761904761862,12
 ITE55350967_PLI_INCONNU_-_ADH_625302_42296806_DMS_,0.9000000000000001,0.09999999999999987,7
 ITE55351087_PLI_INCONNU_-_ADH_625120_42296075_DMS_,0.9861111111111112,0.01388888888888884,8
 ITE55351128_PLI_INCONNU_-_ADH_625123_42296103_DMS_,0.9619522084195998,0.03804779158040017,12
 ITE55351156_Passeport_Completed_,0.7908496732026145,0.20915032679738554,11
 ITE55352532_PLI_INCONNU_-_ADH_625374_42297046_DMS_,0.9166666666666666,0.08333333333333337,6
 ITE55352984_CNI_Completed_,0.8507575757575757,0.1492424242424243,12
 ITE55354550_20250728_113359_Completed_,0.9514957264957264,0.048504273504273554,8
 ITE55354891_image_Completed_,0.975,0.025000000000000022,8
 image (1)_,0.9979166666666667,0.002083333333333326,12
 image (11)_,0.9239583333333333,0.07604166666666667,12
 image (12)_,0.8632066276803118,0.13679337231968824,12
 image (13)_,0.987962962962963,0.012037037037036957,12
 image (14)_,0.9,0.09999999999999998,4
 image (15)_,0.9012345679012346,0.09876543209876543,9
 image (16)_,0.8765432098765432,0.12345679012345678,9
 image (17)_,0.9882154882154882,0.011784511784511786,12
 image (18)_,0.9079868451961475,0.09201315480385253,11
 image (19)_,0.9717967047930284,0.028203295206971624,12
 image (2)_,0.0625,0.9375,1
 image (21)_,0.670990990990991,0.329009009009009,5
 image (22)_,0.9969135802469137,0.0030864197530863224,12
 image (23)_,0.9891304347826086,0.010869565217391353,8
 image (24)_,0.5843928398645379,0.41560716013546206,5
 image (25)_,0.9107142857142857,0.0892857142857143,8
 image (26)_,0.8711111111111111,0.12888888888888894,10
 image (27)_,0.8666666666666667,0.1333333333333333,9
 image (28)_,0.9916666666666667,0.008333333333333304,8
 image (29)_,0.7705994897959184,0.2294005102040816,10
 image (3)_,0.6199546485260772,0.3800453514739228,7
 image (30)_,0.8711111111111111,0.12888888888888894,10
 image (31)_,0.9249999999999998,0.07500000000000018,12
 image (4)_,0.9285714285714286,0.0714285714285714,8
 image (6)_,0.0,0.999999999,12
 image (8)_,0.9898809523809523,0.010119047619047694,12
 image (9)_,0.8595959595959596,0.14040404040404042,11
 image_,0.8256514727102964,0.17434852728970363,7
--- a/data/task/cni/inference_out/crop_shadow_paddle_cache/field_metrics.csv
+++ b/data/task/cni/inference_out/crop_shadow_paddle_cache/field_metrics.csv
@@ -0,0 +1,16 @@
 field,tp,fp,fn,precision,recall,f1,accuracy
 address,21,5,6,0.807692,0.777778,0.792453,0.777778
 birth_date,35,3,4,0.921053,0.897436,0.909091,0.897436
 birth_place,34,0,1,1.000000,0.971429,0.985507,0.971429
 doc_number,33,2,3,0.942857,0.916667,0.929577,0.916667
 document_type,44,3,4,0.936170,0.916667,0.926316,0.916667
 expiry_date,29,5,7,0.852941,0.805556,0.828571,0.805556
 gender,37,0,1,1.000000,0.973684,0.986667,0.973684
 issue_date,25,4,6,0.862069,0.806452,0.833333,0.806452
 issue_place,17,3,5,0.850000,0.772727,0.809524,0.772727
 name,36,5,6,0.878049,0.857143,0.867470,0.857143
 nationality,39,0,1,1.000000,0.975000,0.987342,0.975000
 permit_type,1,0,0,1.000000,1.000000,1.000000,1.000000
 personal_number,3,3,3,0.500000,0.500000,0.500000,0.500000
 remarks,1,0,0,1.000000,1.000000,1.000000,1.000000
 surname,37,4,5,0.902439,0.880952,0.891566,0.880952
--- a/data/task/cni/inference_out/crop_shadow_paddle_cache/per_image_anls.csv
+++ b/data/task/cni/inference_out/crop_shadow_paddle_cache/per_image_anls.csv
@@ -0,0 +1,49 @@
 image,anls,hallucination_score,num_fields
 ITE55336627_PASSEPORT_Abasse_GUEYE__Completed_,1.0,0.0,12
 ITE55340318_ID_kat__1__Completed_,0.9796176046176046,0.02038239538239539,12
 ITE55340822_passeport_pascal_2028_Completed_,1.0,0.0,12
 ITE55341271_C_I_face_Completed_,0.711111111111111,0.288888888888889,10
 ITE55346966_17537774579547958575370370624241_Completed_,0.8177777777777777,0.18222222222222229,5
 ITE55347865_Snapchat-715567440_Completed_,0.8255747126436782,0.1744252873563218,8
 ITE55347866_Snapchat-1551171803_Completed_,0.8631229235880399,0.13687707641196012,5
 ITE55347926_DHONDT_CNI_,0.7364197530864198,0.2635802469135802,9
 ITE55348016_image_Completed_,0.975,0.025000000000000022,8
 ITE55348464_IMG-20250721-WA0000_Completed_,0.7173082173082174,0.28269178269178263,11
 ITE55348878_CNI_RECTO_,0.8308648459383754,0.1691351540616246,8
 ITE55349793_2_em_CNI_CHASTAING__,0.9880952380952381,0.011904761904761862,12
 ITE55350967_PLI_INCONNU_-_ADH_625302_42296806_DMS_,0.9760239760239761,0.02397602397602394,7
 ITE55351087_PLI_INCONNU_-_ADH_625120_42296075_DMS_,1.0,0.0,8
 ITE55351128_PLI_INCONNU_-_ADH_625123_42296103_DMS_,0.9555555555555556,0.0444444444444444,12
 ITE55351156_Passeport_Completed_,0.8817587641117054,0.1182412358882946,11
 ITE55352532_PLI_INCONNU_-_ADH_625374_42297046_DMS_,1.0,0.0,6
 ITE55352984_CNI_Completed_,0.9499999999999998,0.050000000000000155,12
 ITE55354550_20250728_113359_Completed_,0.961111111111111,0.03888888888888897,8
 ITE55354891_image_Completed_,0.975,0.025000000000000022,8
 image (1)_,0.9979166666666667,0.002083333333333326,12
 image (11)_,0.9046875,0.09531250000000002,12
 image (12)_,0.8842592592592592,0.11574074074074081,12
 image (13)_,0.9275462962962964,0.07245370370370363,12
 image (14)_,0.9,0.09999999999999998,4
 image (15)_,0.9012345679012346,0.09876543209876543,9
 image (16)_,0.854320987654321,0.14567901234567904,9
 image (17)_,0.9882154882154882,0.011784511784511786,12
 image (18)_,0.9170777542870566,0.08292224571294338,11
 image (19)_,0.9907407407407408,0.00925925925925919,12
 image (2)_,0.0625,0.9375,1
 image (21)_,0.670990990990991,0.329009009009009,5
 image (22)_,0.9969135802469137,0.0030864197530863224,12
 image (23)_,0.9891304347826086,0.010869565217391353,8
 image (24)_,0.23483309143686498,0.765166908563135,5
 image (25)_,0.9107142857142857,0.0892857142857143,8
 image (26)_,0.82,0.18000000000000005,10
 image (27)_,0.8666666666666667,0.1333333333333333,9
 image (28)_,1.0,0.0,8
 image (29)_,0.9726839826839827,0.027316017316017294,10
 image (3)_,0.609750566893424,0.390249433106576,7
 image (30)_,0.8711111111111111,0.12888888888888894,10
 image (31)_,0.6940972222222221,0.30590277777777786,12
 image (4)_,0.9035714285714286,0.09642857142857142,8
 image (6)_,0.0,0.999999999,12
 image (8)_,0.9825091575091576,0.01749084249084243,12
 image (9)_,0.853102453102453,0.14689754689754697,11
 image_,0.8408029878618114,0.1591970121381886,7
--- a/data/task/cni/inference_out/no_preprocessing/field_metrics.csv
+++ b/data/task/cni/inference_out/no_preprocessing/field_metrics.csv
@@ -0,0 +1,16 @@
 field,tp,fp,fn,precision,recall,f1,accuracy
 address,13,6,14,0.684211,0.481481,0.565217,0.481481
 birth_date,33,4,6,0.891892,0.846154,0.868421,0.846154
 birth_place,22,2,13,0.916667,0.628571,0.745763,0.628571
 doc_number,28,5,8,0.848485,0.777778,0.811594,0.777778
 document_type,26,22,22,0.541667,0.541667,0.541667,0.541667
 expiry_date,28,5,8,0.848485,0.777778,0.811594,0.777778
 gender,34,2,4,0.944444,0.894737,0.918919,0.894737
 issue_date,20,7,11,0.740741,0.645161,0.689655,0.645161
 issue_place,6,3,16,0.666667,0.272727,0.387097,0.272727
 name,28,12,14,0.700000,0.666667,0.682927,0.666667
 nationality,33,4,7,0.891892,0.825000,0.857143,0.825000
 permit_type,1,0,0,1.000000,1.000000,1.000000,1.000000
 personal_number,1,0,5,1.000000,0.166667,0.285714,0.166667
 remarks,0,1,1,0.000000,0.000000,0.000000,0.000000
 surname,32,8,10,0.800000,0.761905,0.780488,0.761905
--- a/data/task/cni/inference_out/no_preprocessing/per_image_anls.csv
+++ b/data/task/cni/inference_out/no_preprocessing/per_image_anls.csv
@@ -0,0 +1,49 @@
 image,anls,hallucination_score,num_fields
 image (22),0.82125,0.17874999999999996,12
 image (16),0.8765432098765432,0.12345679012345678,9
 image (18),0.4974404148822753,0.5025595851177247,11
 image (3),0.7913832199546486,0.20861678004535145,7
 ITE55348016_image_Completed,0.7644230769230769,0.23557692307692313,8
 ITE55348878_CNI_RECTO,0.8073553599071208,0.19264464009287918,8
 image (26),0.7659498207885305,0.23405017921146953,10
 image (28),0.7822072072072072,0.21779279279279284,8
 ITE55347926_DHONDT_CNI,0.6816239316239316,0.31837606837606836,9
 image (15),0.9012345679012346,0.09876543209876543,9
 ITE55354891_image_Completed,0.6394230769230769,0.36057692307692313,8
 image (4),0.975,0.025000000000000022,8
 ITE55351156_Passeport_Completed,0.8813131313131314,0.11868686868686862,11
 ITE55347866_Snapchat-1551171803_Completed,0.5564221824686941,0.4435778175313059,5
 ITE55341271_C_I_face_Completed,0.5861416361416361,0.41385836385836394,10
 ITE55352532_PLI_INCONNU_-_ADH_625374_42297046_DMS,0.6481481481481481,0.35185185185185186,6
 image (9),0.9141414141414141,0.08585858585858586,11
 image (25),0.8667929292929293,0.13320707070707072,8
 ITE55352984_CNI_Completed,0.7075757575757576,0.29242424242424236,12
 ITE55351087_PLI_INCONNU_-_ADH_625120_42296075_DMS,0.9527777777777777,0.047222222222222276,8
 ITE55349793_2_em_CNI_CHASTAING_,0.938095238095238,0.06190476190476202,12
 ITE55354550_20250728_113359_Completed,0.5625,0.4375,8
 ITE55340318_ID_kat__1__Completed,0.3825320512820513,0.6174679487179486,12
 ITE55340822_passeport_pascal_2028_Completed,0.8164983164983166,0.1835016835016834,12
 ITE55351128_PLI_INCONNU_-_ADH_625123_42296103_DMS,0.6349206349206349,0.3650793650793651,12
 image (2),0.0625,0.9375,1
 image (21),0.7452631578947368,0.25473684210526315,5
 image (12),0.8540820231996702,0.1459179768003298,12
 image (24),0.04615384615384614,0.9538461538461539,5
 image (31),0.8666666666666666,0.13333333333333341,12
 ITE55347865_Snapchat-715567440_Completed,0.5144230769230769,0.48557692307692313,8
 image (30),0.8037037037037036,0.1962962962962964,10
 image (27),0.8888888888888888,0.11111111111111116,9
 ITE55348464_IMG-20250721-WA0000_Completed,0.8966810966810965,0.10331890331890348,11
 image (19),0.6095515276549759,0.3904484723450241,12
 ITE55346966_17537774579547958575370370624241_Completed,0.42663650075414783,0.5733634992458522,5
 image (8),0.6822344322344321,0.31776556776556786,12
 image (13),0.9907407407407408,0.00925925925925919,12
 image (23),0.02884615384615384,0.9711538461538461,8
 image (1),0.6418574481074482,0.35814255189255184,12
 image (14),1.0,0.0,4
 image_,0.7751322751322751,0.22486772486772488,7
 image (11),0.3424526862026862,0.6575473137973138,12
 image (6),0.9325913349682451,0.06740866503175491,12
 image (29),0.8412925170068026,0.15870748299319737,10
 ITE55336627_PASSEPORT_Abasse_GUEYE__Completed,1.0,0.0,12
 image (17),0.6437516187516187,0.3562483812483813,12
 ITE55350967_PLI_INCONNU_-_ADH_625302_42296806_DMS,0.14285714285714285,0.8571428571428572,7
--- a/scripts/compute_hallucination_from_anls.py
+++ b/scripts/compute_hallucination_from_anls.py
@@ -0,0 +1,303 @@
 #!/usr/bin/env python3
 """
 Compute ANLS and hallucination scores from per-sample evaluation JSONs and plot results.
 Inputs: one or more JSON files with schema like:
 [
  {
    "image": "image (22)",
    "num_pred_fields": 10,
    "num_gt_fields": 12,
    "num_correct": 9,
    "all_correct": false,
    "fields": [
      {"field": "address", "pred": "...", "gt": "...", "correct": true},
      ...
    ]
  },
  ...
 ]
 ANLS definition: average normalized Levenshtein similarity over fields present in ground truth.
 Here we approximate per-field similarity as:
  sim = 1 - (levenshtein_distance(pred, gt) / max(len(pred), len(gt)))
  clipped into [0, 1], and treat empty max length as exact match (1.0).
 Per-image ANLS is the mean of field similarities for that image. Hallucination is 1 - ANLS.
 Outputs:
 - CSV per input JSON placed next to it: per_image_anls.csv with columns [image, anls, hallucination_score, num_fields]
 - PNG bar chart per input JSON: hallucination_per_image.png with mean line and title.
 """
 from __future__ import annotations
 import argparse
 import json
 import sys
 from pathlib import Path
 from typing import Dict, List, Tuple
 import pandas as pd
 import matplotlib.pyplot as plt
 def levenshtein_distance(a: str, b: str) -> int:
    """Compute Levenshtein distance between two strings (iterative DP)."""
    if a == b:
        return 0
    if len(a) == 0:
        return len(b)
    if len(b) == 0:
        return len(a)
    previous_row = list(range(len(b) + 1))
    for i, ca in enumerate(a, start=1):
        current_row = [i]
        for j, cb in enumerate(b, start=1):
            insertions = previous_row[j] + 1
            deletions = current_row[j - 1] + 1
            substitutions = previous_row[j - 1] + (0 if ca == cb else 1)
            current_row.append(min(insertions, deletions, substitutions))
        previous_row = current_row
    return previous_row[-1]
 def normalized_similarity(pred: str, gt: str) -> float:
    """Return 1 - normalized edit distance in [0, 1]."""
    pred = pred or ""
    gt = gt or ""
    max_len = max(len(pred), len(gt))
    if max_len == 0:
        return 1.0
    dist = levenshtein_distance(pred, gt)
    sim = 1.0 - (dist / max_len)
    if sim < 0.0:
        return 0.0
    if sim > 1.0:
        return 1.0
    return sim
 def compute_anls_for_record(record: Dict) -> Tuple[float, int]:
    """Compute ANLS and number of fields for a single record object."""
    fields = record.get("fields") or []
    if not isinstance(fields, list) or len(fields) == 0:
        return 0.0, 0
    sims: List[float] = []
    for f in fields:
        pred = str(f.get("pred", ""))
        gt = str(f.get("gt", ""))
        sims.append(normalized_similarity(pred, gt))
    anls = float(sum(sims) / len(sims)) if sims else 0.0
    return anls, len(sims)
 def process_json(json_path: Path) -> Path:
    with json_path.open("r", encoding="utf-8") as f:
        data = json.load(f)
    rows = []
    for rec in data:
        image_name = rec.get("image")
        anls, num_fields = compute_anls_for_record(rec)
        hallucination = 1.0 - anls
        rows.append({
            "image": image_name,
            "anls": anls,
            "hallucination_score": hallucination,
            "num_fields": int(num_fields),
        })
    df = pd.DataFrame(rows)
    out_csv = json_path.parent / "per_image_anls.csv"
    df.to_csv(out_csv, index=False)
    # Plot hallucination bar chart with mean line
    if len(df) > 0:
        sorted_df = df.sort_values("hallucination_score", ascending=False).reset_index(drop=True)
        plt.figure(figsize=(max(8, len(sorted_df) * 0.12), 5))
        plt.bar(range(len(sorted_df)), sorted_df["hallucination_score"].values, color="#1f77b4")
        mean_val = float(sorted_df["hallucination_score"].mean())
        plt.axhline(mean_val, color="red", linestyle="--", label=f"Mean={mean_val:.3f}")
        plt.xlabel("Image (sorted by hallucination)")
        plt.ylabel("Hallucination = 1 - ANLS")
        plt.title(f"Hallucination per image: {json_path.parent.name}")
        plt.legend()
        plt.tight_layout()
        out_png = json_path.parent / "hallucination_per_image.png"
        plt.savefig(out_png, dpi=150)
        plt.close()
    return out_csv
 def common_parent(paths: List[Path]) -> Path:
    if not paths:
        return Path.cwd()
    common = Path(Path(paths[0]).anchor)
    parts = list(Path(paths[0]).resolve().parts)
    for i in range(1, len(paths)):
        other_parts = list(Path(paths[i]).resolve().parts)
        # shrink parts to common prefix
        new_parts: List[str] = []
        for a, b in zip(parts, other_parts):
            if a == b:
                new_parts.append(a)
            else:
                break
        parts = new_parts
    if not parts:
        return Path.cwd()
    return Path(*parts)
 def main() -> None:
    parser = argparse.ArgumentParser(description="Compute ANLS and hallucination from per-sample JSONs and plot results.")
    parser.add_argument("inputs", nargs="+", help="Paths to per_sample_eval.json files")
    args = parser.parse_args()
    any_error = False
    combined_rows: List[Dict] = []
    input_paths: List[Path] = []
    for in_path_str in args.inputs:
        path = Path(in_path_str)
        if not path.exists():
            print(f"[WARN] File does not exist: {path}", file=sys.stderr)
            any_error = True
            continue
        try:
            out_csv = process_json(path)
            print(f"Processed: {path} -> {out_csv}")
            # Load just-written CSV to aggregate and tag method
            df = pd.read_csv(out_csv)
            method_name = path.parent.name
            df["method"] = method_name
            combined_rows.extend(df.to_dict(orient="records"))
            input_paths.append(path)
        except Exception as exc:
            print(f"[ERROR] Failed to process {path}: {exc}", file=sys.stderr)
            any_error = True
    # Create combined outputs if we have multiple inputs
    if combined_rows:
        combo_df = pd.DataFrame(combined_rows)
        # Reorder columns
        cols = ["image", "method", "anls", "hallucination_score", "num_fields"]
        combo_df = combo_df[cols]
        base_outdir = common_parent(input_paths)
        combined_dir = base_outdir / "combined_anls"
        combined_dir.mkdir(parents=True, exist_ok=True)
        combined_csv = combined_dir / "combined_per_image_anls.csv"
        combo_df.to_csv(combined_csv, index=False)
        # Mean hallucination per method (bar chart)
        means = combo_df.groupby("method")["hallucination_score"].mean().sort_values(ascending=False)
        stds = combo_df.groupby("method")["hallucination_score"].std().reindex(means.index)
        plt.figure(figsize=(max(6, len(means) * 1.2), 5))
        plt.bar(means.index, means.values, yerr=stds.values, capsize=4, color="#2ca02c")
        overall_mean = float(combo_df["hallucination_score"].mean())
        plt.axhline(overall_mean, color="red", linestyle="--", label=f"Overall mean={overall_mean:.3f}")
        plt.ylabel("Mean hallucination (1 - ANLS)")
        plt.title("Mean hallucination by method")
        plt.xticks(rotation=20, ha="right")
        plt.legend()
        plt.tight_layout()
        bar_png = combined_dir / "mean_hallucination_by_method.png"
        plt.savefig(bar_png, dpi=160)
        plt.close()
        # Heatmap: images x methods (hallucination)
        pivot = combo_df.pivot_table(index="image", columns="method", values="hallucination_score", aggfunc="mean")
        # Sort images by average hallucination descending for readability
        pivot = pivot.reindex(pivot.mean(axis=1).sort_values(ascending=False).index)
        plt.figure(figsize=(max(8, len(pivot.columns) * 1.0), max(6, len(pivot.index) * 0.25)))
        im = plt.imshow(pivot.values, aspect="auto", cmap="viridis")
        plt.colorbar(im, label="Hallucination (1 - ANLS)")
        plt.xticks(range(len(pivot.columns)), pivot.columns, rotation=30, ha="right")
        plt.yticks(range(len(pivot.index)), pivot.index)
        plt.title("Hallucination per image across methods")
        plt.tight_layout()
        heatmap_png = combined_dir / "hallucination_heatmap.png"
        plt.savefig(heatmap_png, dpi=160)
        plt.close()
        print(f"Combined CSV: {combined_csv}")
        print(f"Saved: {bar_png}")
        print(f"Saved: {heatmap_png}")
        # Line chart: 1 line per method over images, hide image names
        # Use same image order as pivot
        methods = list(pivot.columns)
        x = list(range(len(pivot.index)))
        plt.figure(figsize=(max(10, len(x) * 0.12), 5))
        colors = plt.rcParams['axes.prop_cycle'].by_key().get('color', ['#1f77b4','#ff7f0e','#2ca02c','#d62728','#9467bd'])
        for idx, method in enumerate(methods):
            y = pivot[method].to_numpy()
            plt.plot(x, y, label=method, linewidth=1.8, color=colors[idx % len(colors)])
        plt.ylim(0.0, 1.0)
        plt.xlabel("Images (sorted by overall hallucination)")
        plt.ylabel("Hallucination (1 - ANLS)")
        plt.title("Hallucination across images by method")
        plt.xticks([], [])  # hide image names
        # Mean note box
        mean_lines = []
        for method in methods:
            m = float(combo_df[combo_df["method"] == method]["hallucination_score"].mean())
            mean_lines.append(f"{method}: {m:.3f}")
        text = "\n".join(mean_lines)
        plt.gca().text(0.99, 0.01, text, transform=plt.gca().transAxes,
                       fontsize=9, va='bottom', ha='right',
                       bbox=dict(boxstyle='round', facecolor='white', alpha=0.8, edgecolor='gray'))
        plt.legend(loc="upper right", ncol=min(3, len(methods)))
        plt.tight_layout()
        line_png = combined_dir / "hallucination_lines_by_method.png"
        plt.savefig(line_png, dpi=160)
        plt.close()
        # Grouped-by-image interlocking line chart with image labels
        # Build a consistent x position per image, with small offsets per method
        base_x = list(range(len(pivot.index)))
        offsets = {
            m: ((i - (len(methods) - 1) / 2) * 0.12) for i, m in enumerate(methods)
        }
        # Cap width to avoid extremely long images; dynamic but limited
        width = min(16, max(10, len(base_x) * 0.12))
        plt.figure(figsize=(width, 6))
        for idx, method in enumerate(methods):
            # Fill missing values with 0 to connect lines seamlessly
            y = pivot[method].fillna(0.0).to_numpy()
            x_shifted = [bx + offsets[method] for bx in base_x]
            plt.plot(x_shifted, y, label=method, linewidth=1.8, marker='o', markersize=3,
                     color=colors[idx % len(colors)])
        plt.ylim(0.0, 1.0)
        plt.xlim(-0.5, len(base_x) - 0.5)
        # Hide image names; keep index ticks sparse for readability
        plt.xticks([], [])
        plt.xlabel("Images (index)")
        plt.ylabel("Hallucination (1 - ANLS)")
        plt.title("Hallucination by image (interlocked methods)")
        plt.grid(axis='y', linestyle='--', alpha=0.3)
        # Add box with per-method mean
        text2 = "\n".join([f"{m}: {float(combo_df[combo_df['method']==m]['hallucination_score'].mean()):.3f}" for m in methods])
        plt.gca().text(0.99, 0.01, text2, transform=plt.gca().transAxes,
                       fontsize=9, va='bottom', ha='right',
                       bbox=dict(boxstyle='round', facecolor='white', alpha=0.8, edgecolor='gray'))
        plt.legend(loc='upper right', ncol=min(3, len(methods)))
        plt.tight_layout()
        group_line_png = combined_dir / "hallucination_interlocked_by_image.png"
        plt.savefig(group_line_png, dpi=160)
        plt.close()
        print(f"Combined CSV: {combined_csv}")
        print(f"Saved: {bar_png}")
        print(f"Saved: {heatmap_png}")
        print(f"Saved: {line_png}")
        print(f"Saved: {group_line_png}")
    if any_error:
        sys.exit(1)
 if __name__ == "__main__":
    main()
--- a/scripts/pipeline_compare.py
+++ b/scripts/pipeline_compare.py
@@ -0,0 +1,254 @@
 #!/usr/bin/env python3
 """
 Compare hallucination across three pipelines over five preprocessing methods:
  1) Raw: all images
  2) DeQA-filtered: keep images with DeQA score >= threshold (default 2.6)
  3) Human-filtered: keep images labeled High in CSV labels
 Inputs:
  - One or more per_sample_eval.json files (or per_image_anls.csv already generated)
  - DeQA score file (txt): lines like "3.9  - image (9)_0.png"
  - Human labels CSV with columns: filename,label where label in {High,Low}
 Outputs:
  - Combined means CSV: method vs mean hallucination for each pipeline
  - Line chart (3 lines): hallucination mean per method across the three pipelines
 """
 from __future__ import annotations
 import argparse
 import csv
 import json
 import re
 from pathlib import Path
 from typing import Dict, List, Tuple, Optional
 import pandas as pd
 import matplotlib.pyplot as plt
 def canonical_key(name: str) -> str:
    """Map various filenames to a canonical key used by per_sample_eval 'image' field.
    Examples:
      - "image (9)_0.png" -> "image (9)"
      - "image (22)" -> "image (22)"
      - "foo/bar/image (15)_3.jpg" -> "image (15)"
      - other names -> stem without extension
    """
    if not name:
        return name
    # Keep only basename
    base = Path(name).name
    # Try pattern image (N)
    m = re.search(r"(image \(\d+\))", base, flags=re.IGNORECASE)
    if m:
        return m.group(1)
    # Fallback: remove extension
    return Path(base).stem
 def read_deqa_scores(txt_path: Path) -> Dict[str, float]:
    scores: Dict[str, float] = {}
    with txt_path.open("r", encoding="utf-8") as f:
        for line in f:
            line = line.strip()
            if not line:
                continue
            # Accept formats: "3.9  - filename" or "filename,3.9" etc.
            m = re.match(r"\s*([0-9]+(?:\.[0-9]+)?)\s*[-,:]?\s*(.+)$", line)
            if m:
                score = float(m.group(1))
                filename = m.group(2)
            else:
                parts = re.split(r"[,\t]", line)
                if len(parts) >= 2:
                    try:
                        score = float(parts[1])
                        filename = parts[0]
                    except Exception:
                        continue
                else:
                    continue
            key = canonical_key(filename)
            scores[key] = score
    return scores
 def read_human_labels(csv_path: Path) -> Dict[str, str]:
    labels: Dict[str, str] = {}
    with csv_path.open("r", encoding="utf-8") as f:
        reader = csv.DictReader(f)
        for row in reader:
            filename = (row.get("filename") or row.get("file") or "").strip()
            label = (row.get("label") or row.get("Label") or "").strip()
            if not filename:
                continue
            key = canonical_key(filename)
            if label:
                labels[key] = label
    return labels
 def levenshtein_distance(a: str, b: str) -> int:
    if a == b:
        return 0
    if len(a) == 0:
        return len(b)
    if len(b) == 0:
        return len(a)
    previous_row = list(range(len(b) + 1))
    for i, ca in enumerate(a, start=1):
        current_row = [i]
        for j, cb in enumerate(b, start=1):
            insertions = previous_row[j] + 1
            deletions = current_row[j - 1] + 1
            substitutions = previous_row[j - 1] + (0 if ca == cb else 1)
            current_row.append(min(insertions, deletions, substitutions))
        previous_row = current_row
    return previous_row[-1]
 def normalized_similarity(pred: str, gt: str) -> float:
    pred = pred or ""
    gt = gt or ""
    max_len = max(len(pred), len(gt))
    if max_len == 0:
        return 1.0
    dist = levenshtein_distance(pred, gt)
    sim = 1.0 - (dist / max_len)
    if sim < 0.0:
        return 0.0
    if sim > 1.0:
        return 1.0
    return sim
 def compute_anls_for_record(record: Dict) -> tuple[float, int]:
    fields = record.get("fields") or []
    if not isinstance(fields, list) or len(fields) == 0:
        return 0.0, 0
    sims: List[float] = []
    for f in fields:
        pred = str(f.get("pred", ""))
        gt = str(f.get("gt", ""))
        sims.append(normalized_similarity(pred, gt))
    anls = float(sum(sims) / len(sims)) if sims else 0.0
    return anls, len(sims)
 def load_per_image_anls(input_json: Path) -> pd.DataFrame:
    # Prefer existing per_image_anls.csv, otherwise compute quickly
    per_image_csv = input_json.parent / "per_image_anls.csv"
    if per_image_csv.exists():
        df = pd.read_csv(per_image_csv)
        return df
    # Fallback: compute minimal ANLS like in the other script
    with input_json.open("r", encoding="utf-8") as f:
        data = json.load(f)
    rows = []
    for rec in data:
        anls, num_fields = compute_anls_for_record(rec)
        rows.append({
            "image": rec.get("image"),
            "anls": anls,
            "hallucination_score": 1.0 - anls,
            "num_fields": int(num_fields),
        })
    return pd.DataFrame(rows)
 def main() -> None:
    p = argparse.ArgumentParser(description="Compare hallucination across raw/DeQA/Human pipelines over methods")
    p.add_argument("inputs", nargs="+", help="per_sample_eval.json files for each method")
    p.add_argument("--deqa_txt", required=True, help="Path to DeQA scores txt (e.g., cni.txt)")
    p.add_argument("--human_csv", required=True, help="Path to human labels CSV")
    p.add_argument("--deqa_threshold", type=float, default=2.6, help="DeQA threshold (>=)")
    args = p.parse_args()
    # Load filters
    deqa_scores = read_deqa_scores(Path(args.deqa_txt))
    human_labels = read_human_labels(Path(args.human_csv))
    # Aggregate per method
    method_to_df: Dict[str, pd.DataFrame] = {}
    for ip in args.inputs:
        path = Path(ip)
        df = load_per_image_anls(path)
        df["method"] = path.parent.name
        df["image_key"] = df["image"].apply(canonical_key)
        method_to_df[path.parent.name] = df
    # Compute means per pipeline (fair comparison: set excluded images to hallucination=0)
    records = []
    for method, df in method_to_df.items():
        raw_mean = float(df["hallucination_score"].mean()) if len(df) else float("nan")
        # DeQA filter: mark DeQA < threshold as hallucination=0, keep all images
        df_deqa = df.copy()
        mask_deqa = df_deqa["image_key"].map(lambda k: deqa_scores.get(k, None))
        # Set hallucination=0 for images with DeQA < threshold (or missing DeQA)
        df_deqa.loc[mask_deqa.isna() | (mask_deqa < args.deqa_threshold), "hallucination_score"] = 0.0
        deqa_mean = float(df_deqa["hallucination_score"].mean()) if len(df_deqa) else float("nan")
        # Human filter: mark Low labels as hallucination=0, keep all images
        df_human = df.copy()
        mask_human = df_human["image_key"].map(lambda k: human_labels.get(k, "").lower())
        # Set hallucination=0 for images labeled Low (or missing label)
        df_human.loc[mask_human != "high", "hallucination_score"] = 0.0
        human_mean = float(df_human["hallucination_score"].mean()) if len(df_human) else float("nan")
        records.append({
            "method": method,
            "raw_mean": raw_mean,
            "deqa_mean": deqa_mean,
            "human_mean": human_mean,
            "raw_count": int(len(df)),
            "deqa_count": int(len(df_deqa)),  # Now equal to raw_count
            "human_count": int(len(df_human)),  # Now equal to raw_count
        })
    outdir = Path(args.inputs[0]).parent.parent / "combined_anls" / "pipeline"
    outdir.mkdir(parents=True, exist_ok=True)
    out_csv = outdir / "pipeline_means.csv"
    means_df = pd.DataFrame(records).sort_values("method")
    means_df.to_csv(out_csv, index=False)
    # 3-line comparison plot over methods (narrower with score annotations)
    x = range(len(means_df))
    plt.figure(figsize=(7, 5))
    # Plot lines and add score annotations
    raw_vals = means_df["raw_mean"].values
    deqa_vals = means_df["deqa_mean"].values
    human_vals = means_df["human_mean"].values
    plt.plot(x, raw_vals, marker="o", label="Raw", linewidth=2, markersize=6)
    plt.plot(x, deqa_vals, marker="s", label=f"DeQA >= {args.deqa_threshold}", linewidth=2, markersize=6)
    plt.plot(x, human_vals, marker="^", label="Human High", linewidth=2, markersize=6)
    # Annotate each point with its score
    for i, (r, d, h) in enumerate(zip(raw_vals, deqa_vals, human_vals)):
        plt.annotate(f"{r:.3f}", (i, r), textcoords="offset points", xytext=(0,8), ha='center', fontsize=8)
        plt.annotate(f"{d:.3f}", (i, d), textcoords="offset points", xytext=(0,8), ha='center', fontsize=8)
        plt.annotate(f"{h:.3f}", (i, h), textcoords="offset points", xytext=(0,8), ha='center', fontsize=8)
    plt.xticks(list(x), means_df["method"].tolist(), rotation=25, ha="right")
    plt.ylabel("Mean hallucination (1 - ANLS)")
    plt.title("Pipeline comparison over preprocessing methods")
    plt.grid(axis="y", linestyle="--", alpha=0.3)
    plt.legend()
    plt.tight_layout()
    out_png = outdir / "pipeline_comparison.png"
    plt.savefig(out_png, dpi=160)
    plt.close()
    print(f"Saved: {out_csv}")
    print(f"Saved: {out_png}")
 if __name__ == "__main__":
    main()
Author	SHA1	Message	Date
trungkienbkhn	77b9784c32	Merge branch 'main' of https://code.rizlum.com/ngpthanh15/IQA-Metric-Benchmark	2025-09-12 08:03:17 +00:00
trungkienbkhn	0d6091fc0c	update ressult	2025-09-11 09:39:02 +00:00
trungkienbkhn	3cbf74149a	update humans label	2025-09-08 07:43:21 +00:00