PP-OCRv6_manga

PP-OCRv6_manga

PP-OCRv6_manga is a lightweight pipeline for manga and anime text detection and recognition (Japanese and Chinese), small enough to run on a CPU: detection 0.9 MB + recognition 10 MB in FP16.

<p align="center"> <img src="assets/rec_by_benchmark.webp" alt="Text recognition: CER on Manga, Illustrations and JMangaBench_Read"> </p> <p align="center"> <img src="assets/rec_japanese_vs_size.webp" alt="Japanese recognition: CER vs model size"> </p>

What's new in v0.2 (2026-09-28)

New detector manga_det_v0.2 and recognizer manga_rec_v0.2, drop-in replacements for v0.1 with the same architectures, sizes and speed.

Detector. Retrained on much more manga, webtoon and illustration data. It misses fewer lines, splits or merges lines less often and gives a third fewer false positives on illustrations. With the same recognizer, end-to-end CER drops by about 11% (Table 1).

Recognizer. v0.1 had forgotten most rare kanji; v0.2 brings CER on rare characters from about 60% to about 10% and on SFX from about 45% to about 26%. The other benchmarks improve by about 20% relative CER (Table 2).

Full pipeline, end-to-end CER on the Table 1 pages:

Models

• Dict: ppocrv6_dict.txt (18,708 symbols), shared by both recognizers.
• The .pdparams files contain the inference weights (backbone + CTC branch).

---

Benchmark Results

Evaluation data

• Table 1 (842 pages): the Table 2 pages plus 286 new AnimeText pages; pages found in any training set are excluded.
• Table 2 (5,181 lines, 562 pages): manga, manhua / webtoon (~85% Chinese) and illustrations, read on the ground-truth line boxes (no detector). Boxes and texts reviewed by hand (34 lines dropped, 267 boxes fixed). Korean and Cyrillic lines excluded. CER after NFKC and folding of notation variants (・・・ = …, ♥ / ❤ = ♡, 〇 / ◯ = ○, 〝〟 dropped).
• SFX, Manga109: COO onomatopoeia annotations on Manga109 pages held out from our training, best of 4 rotations.
• SFX, Table 2 pages: the 389 Table 2 lines tagged sfx in the GT (manga 292, illustrations 97; webtoon strips not mapped), read on their GT boxes like the rest of Table 2.
• Rare kanji, synthetic: lines of rare characters in unseen fonts (plain + SFX style), character-weighted.
• Rare kanji, real lines: AnimeText tag500 test lines with a character seen < 20 times in our training labels.
• JMangaBench_Read v1.3.2 (7,789 balloon crops, RIN v1.2 normalization, composite = mean of printed / non-printed 1−NED).

---

Table 1: Text Detection Benchmark (recall, false positives, end-to-end CER)

v0.2 / v0.1: manga_det_v0.2 / manga_det_v0.1; base small / medium: the official PP-OCRv6 detectors. 960 / 1120: detector input long side (the demo app uses 960). FP: unmatched boxes; ruby boxes next to a GT line are not FP (geometric rule on the GT side, independent of the app filter). CER: end-to-end CER. Best in bold, second best underlined.

All detectors run the demo pipeline: thresh=0.15, box_thresh=0.25, unclip_ratio=1.4, area resize for long strips, learned furigana filter, block-aware reading order, then manga_rec_v0.2.

---

Table 2: Text Recognition (CER, lower is better)

v0.2 / v0.1: manga_rec_v0.2 / manga_rec_v0.1; hayai: hayai-ocr-v2.5-nova; manga-ocr nar small: manga-ocr-nar-small (muscgab, CC BY-NC-SA 4.0); base: the official PP-OCRv6 recognizers. Best in bold, second best underlined (plain numbers only, not *, ¹).

* Japanese only: manga-ocr nar small does not read Chinese or Latin, so these values are on the Japanese lines (v0.2 on the same lines: 4.23 total / 84.67 exact).

¹ Random character strings without context; see the real-lines row.

JMangaBench_Read: hayai and manga-ocr nar small read the whole crop (official leaderboard numbers, reproduced locally); PaddleOCR models run the demo pipeline on each crop. manga-ocr nar small trained on AnimeText-OCR crops, which may overlap Table 2.

---

Table 3: ONNX FP32 vs FP16 (OpenVINO, CPU)

FP16 halves the size at a cost of ~0.05 pp CER. ONNX FP32 matches Paddle (detection: 11.34% in Paddle; recognition: 1496/1500 identical predictions).

---

Directory Structure

PP-OCRv6_manga/
├── README.md
├── ppocrv6_dict.txt                   # Character dictionary (18,708 symbols)
├── examples/
│   ├── onnx_openvino.py               # ONNX / OpenVINO pipeline
│   └── paddle_inference.py            # Paddle models + PaddleOCR post-processing
├── det/
│   ├── manga_det_v0.2.pdparams        # Paddle weights (1.9 MB)   <- default
│   ├── manga_det_v0.2.onnx            # ONNX FP32 (1.73 MB)
│   ├── manga_det_v0.2_fp16.onnx       # ONNX FP16 (0.92 MB)
│   ├── manga_det_v0.1.pdparams        # previous version
│   ├── manga_det_v0.1.onnx
│   └── manga_det_v0.1_fp16.onnx
└── rec/
    ├── manga_rec_v0.2.pdparams        # Paddle inference weights (21.3 MB)   <- default
    ├── manga_rec_v0.2.onnx            # ONNX FP32 (21.2 MB)
    ├── manga_rec_v0.2_fp16.onnx       # ONNX FP16 (10.6 MB)
    ├── manga_rec_v0.1.pdparams        # previous version
    ├── manga_rec_v0.1.onnx
    └── manga_rec_v0.1_fp16.onnx

Quick Start

• ONNX / OpenVINO (CPU, ~11 MB in FP16, no PaddlePaddle): examples/onnx_openvino.py, the full pipeline (detection, DB post-processing, recognition) for one page.

  pip install openvino opencv-python numpy pyclipper huggingface_hub
  python examples/onnx_openvino.py page.jpg
  

• Paddle: examples/paddle_inference.py, model definitions and weights with PaddleOCR's DBPostProcess and CTCLabelDecode.
• Recommended DB post-processing: thresh=0.15, box_thresh=0.25, unclip_ratio=1.4, detector input long side 960 (pages longer than 2:1 get the pixel budget of a 3:4 page instead).