qefro/pp-OCRv5-6-ncnn

PP-OCR NCNN models

NCNN-converted PP-OCRv5 mobile multilingual recognition models, plus bundled PP-OCRv6 tiny/small/medium models from the PaddleOCR-ncnn-CPP archive.

Included

PP-OCRv5 multilingual rec models

Language-specific recognition models: arabic, cyrillic, devanagari, el, en, eslav, korean, latin, ta, te, th

Files:

• {lang}_PP-OCRv5_mobile_rec.param
• {lang}_PP-OCRv5_mobile_rec.bin
• ppocr_keys_{lang}.txt
• configs/config_ppocrv5_{lang}.json

PP-OCRv6 bundled models

Bundled from the PaddleOCR-ncnn-CPP release archive:

• PP_OCRv6_tiny_det/rec
• PP_OCRv6_small_det/rec
• PP_OCRv6_medium_det/rec
• ppocr_keys_v6.txt
• ppocr_keys_v6_tiny.txt
• sample configs under configs/

Notes:

• PP-OCRv6_tiny does not support Japanese.
• For Japanese, use PP_OCRv6_small or PP_OCRv6_medium.
• For Tamil / Telugu / Arabic / Devanagari / Thai / Greek / Korean multilingual work, prefer the dedicated PP-OCRv5 language-specific rec models included here.

Source

• PP-OCRv5 language-specific ONNX models: PaddlePaddle/{lang}_PP-OCRv5_mobile_rec_onnx
• PP-OCRv6 bundled assets: Avafly/PaddleOCR-ncnn-CPP release archive

Conversion details

PP-OCRv5 ONNX models were converted via pnnx with:

• inputshape=[1,3,48,320]
• inputshape2=[1,3,48,640]
• fp16=0

Blob names were patched for PaddleOCR-ncnn-CPP compatibility: in0/out0 -> input/output.