shadow-cann/hispark-modelzoo-paddleocrv4-det

PaddleOCRv4-det

PP-OCRv4检测模型在PP-OCRv3检测模型的基础上，在网络结构，训练策略，蒸馏策略三个方面做了优化。首先，PP-OCRv4检测模型使用PP-LCNetV3替换MobileNetv3，并提出并行分支融合的PFhead结构；其次，训练时动态调整shrink ratio的比例；最后，PP-OCRv4对CML的蒸馏loss进行优化，进一步提升文字检测效果。

Mirror Metadata

• Hugging Face repo: shadow-cann/hispark-modelzoo-paddleocrv4-det
• Portal model id: i9sq1n0lec00
• Created at: 2025-12-26 15:54:13
• Updated at: 2025-12-30 20:02:17
• Category: 计算机视觉

Framework

• PyTorch

Supported OS

• OpenHarmony
• Linux

Computing Power

• Hi3403V100 SVP_NNN
• Hi3403V100 NNN

Tags

• OCR

Detail Parameters

• 输入: 960x960
• 参数量: 3.476M
• 计算量: 24.83GFLOPs

Files In This Repo

• ch_ptocr_v4_det_infer.pth (源模型 / 源模型下载; 源模型 / 源模型元数据)
• ch_ptocr_v4_det_simplified.onnx (源模型 / 源模型下载; 源模型 / 源模型元数据)
• SVP_NNN_PC_V1.0.6.0.tgz (附加资源 / 附加资源)

Upstream Links

• Portal card: https://gitbubble.github.io/hisilicon-developer-portal-mirror/model-detail.html?id=i9sq1n0lec00
• Upstream repository: https://gitee.com/HiSpark/modelzoo/tree/master/samples/built-in/ocr/PaddleOCRv4-det
• License reference: https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.10/LICENSE

Notes

• This repository was mirrored from the HiSilicon Developer Portal model card and local downloads captured on 2026-03-27.
• File ownership follows the portal card mapping, not just filename similarity.
• Cover image: 1719770449772547_ocr.jpg