4-class text zone detector for Thai National ID cards. Detects field regions on raw camera frames and routes each crop to the appropriate OCR Reader expert. Trained with driver license negative mining to suppress false detections on non-ID documents. Designed for on-device inference at 30fps on mobile.

Metrics

Metric	Value
Architecture	YOLO26n
Task	detect
Classes	4
mAP50	0.986
mAP50-95	0.927
Precision	0.948
Recall	0.958
Parameters	~2.5M
GFLOPs	5.8 (fused)
Input	640x640
Experiment	EXP-015

Training Data

Synth (850) + real (138) + deduped v10i (165) + DL neg (45) + CPFGS neg (200) = 1398 train / 206 val

Mixed synthetic and real images. Synthetic cards rendered with augmentations (perspective warp, brightness/contrast, noise, blur, JPEG compression) on random backgrounds. Real images from Roboflow with manual 4-class annotations. Driver license images included as background negatives (empty labels) to suppress false field detections on non-ID-card documents.

Negative Mining Results

Metric	v1 (no neg)	v2 (45 DL neg)	v3 (245 neg)
DL False Positive Rate	97.8%	2.2%	0.0%
Phantom Detections (45 DL images)	739	1	0
Max FP Confidence	0.901	0.336	0.000

Sample Results

Confusion Matrix (Normalized)

Precision-Recall Curve

Training Curves

F1-Confidence Curve

Label Distribution

Quick Start

python

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

model_path = hf_hub_download("chayuto/thai-id-ocr-yolo26n-card-field-detector", "best.pt")
model = YOLO(model_path)
results = model("path/to/id_card_photo.jpg")

for box in results[0].boxes:
    cls_id = int(box.cls)
    conf = float(box.conf)
    x1, y1, x2, y2 = box.xyxy[0].tolist()
    print(f"Class {cls_id} ({conf:.2f}): [{x1:.0f}, {y1:.0f}, {x2:.0f}, {y2:.0f}]")

Classes

ID	Name	Routes to
0	`num_id_zone`	Numeric OCR Expert (13-digit CID)
1	`num_dob_zone`	Numeric OCR Expert (DD/MM/YYYY Buddhist Era)
2	`text_eng_zone`	English OCR Expert (romanized Thai names)
3	`text_thai_zone`	Thai OCR Expert (names, addresses, dates, religion)

Fields are grouped by OCR engine type (3 experts), not semantic meaning (10+ fields). Card boundary detection is handled separately by Pipeline B (YOLO26n-Pose with 4-corner keypoints).

Architecture

YOLO26n — NMS-free, DFL-free single-pass detector.

Input: [1, 3, 640, 640]
  → YOLO26n backbone + neck
  → 4-class bbox detection head
Output: [1, 300, 8]  (NMS-free, top-300 detections)
  → [x1, y1, x2, y2, conf, class_0, class_1, class_2, class_3]

Pipeline Context

This model is the Finder stage of the Thai ID Nano OCR pipeline:

Camera Frame (640x640)
  → YOLO26n Finder (this model) — 4-class text zone detection
      → num_id_zone, num_dob_zone  → Numeric Reader (~98.6% EM)
      → text_eng_zone              → English Reader (~97.6% EM)
      → text_thai_zone             → Thai Reader (~95.3% EM)
  → YOLO26n-Pose (Pipeline B) — card boundary + 4-corner keypoints
  → Validator (CID Mod11 checksum, date logic)

Total pipeline: <15 MB, 30fps on mobile.

TFLite Export

python

from ultralytics import YOLO
model = YOLO("best.pt")
model.export(format="tflite", int8=True, nms=False, imgsz=640)

Files

File	Description
`best.pt`	Ultralytics YOLO26n weights (~5.1 MB)
`config.json`	Architecture params, training metadata, per-class metrics
`confusion_matrix_normalized.png`	Normalized confusion matrix
`BoxPR_curve.png`	Precision-Recall curve (all classes)
`BoxF1_curve.png`	F1-Confidence curve
`results.png`	Training loss and metric curves
`labels.jpg`	Dataset label distribution and bbox statistics

License

MIT