PP-OCRv6 One Piece Bubble Line Recognition

Browser-ready OCR package for Projet Poneglyph. The pipeline detects text lines inside a manga bubble with YOLO26n, stitches the detected lines into one horizontal crop, decodes it with fine-tuned PP-OCRv6, then applies a conservative train-derived spacing/case postprocess from onnx/browser_manifest.json.

Current browser result

• Hugging Face repo: Remidesbois/pp-ocrv6-one-piece-bubble-line-rec
• Validation CER: 0.017155317360235393 (1.7155%)
• Validation exact match: 0.7514693534844669 (75.15%)
• Test CER: 0.014505395907584179 (1.4505%)
• Test exact match: 0.7596390484003281 (75.96%)
• Exported bubble images: 7821
• Bubbles without detected lines: 22
• Duplicate line-box bubbles: 0

These metrics are computed from the official 2026-06-29 release prediction CSVs after applying the browser postprocess rules trained only from the train split. The raw 2026-06-29 release baseline was validation 1.9204% CER / 71.62% exact and test 1.7097% CER / 71.21% exact.

Model and detector evidence

• YOLO mAP50: 0.98917488275913
• YOLO mAP50-95: 0.8715335950608416
• Line detector mAP50: 0.995
• Line detector mAP50-95: 0.8926965538059898
• ONNX parity text match: true

Files

• onnx/bubble_line_detector_yolo26n.onnx
• onnx/ppocrv6_bubble_line_rec.onnx
• onnx/ppocrv6_bubble_line_rec_webgpu.onnx
• onnx/browser_manifest.json
• onnx/pipeline_manifest.json
• onnx/ppocrv6_postprocess_rules.json
• postprocess_official_metrics.json
• training_metrics.json
• yolo_metrics.json
• validation_predictions.csv
• test_predictions.csv

The frontend model key is ppocrv6Line.