BADBADBADBOY / pytorchOCR

基于pytorch的ocr算法库,包括 psenet, pan, dbnet, sast , crnn
671 stars 132 forks source link
crnn dbnet ocr pannet psenet sast textdetection textrecognition

基于pytorch的OCR库

这里会有这个项目的代码详解和我的一些ocr经验和心得,我会慢慢更新,有兴趣可以看看,希望可以帮到新接触ocr的童鞋CSDN博客


最近跟新:


目前已完成:

模型 迭代次数 CUTE80 IC03_867 IC13_1015 IC13_857 IC15_1811 IC15_2077 IIIT5k_3000 SVT SVTP mean
resnet34+lstm+ctc 120000 82.98 91.92 90.93 91.59 73.10 67.98 90.16 85.16 78.29 83.56
mobilev3_large+lstm+ctc 210000 73.61 92.50 90.34 91.59 74.82 68.89 87.56 83.46 77.20 82.21
mobilev3_small+lstm+ctc 210000 66.31 90.77 88.76 91.13 73.66 69.52 88.80 84.54 72.24 80.64

检测模型效果(实验中)

训练只在ICDAR2015文本检测公开数据集上,算法效果如下: 模型 骨干网络 precision recall Hmean 下载链接
DB ResNet50_7*7 85.88% 79.10% 82.35% 下载链接(code:fxw6)
DB ResNet50_3*3 86.51% 80.59% 83.44% 下载链接(code:fxw6)
DB MobileNetV3 82.89% 75.83% 79.20% 下载链接(code:fxw6)
SAST ResNet50_7*7 85.72% 78.38% 81.89% 下载链接(code:fxw6)
SAST ResNet50_3*3 86.67% 76.74% 81.40% 下载链接(code:fxw6)
PSE ResNet50_7*7 84.10% 80.01% 82.01% 下载链接(code:fxw6)
PSE ResNet50_3*3 82.56% 78.91% 80.69% 下载链接(code:fxw6)
PAN ResNet18_7*7 81.80% 77.08% 79.37% 下载链接(code:fxw6)
PAN ResNet18_3*3 83.78% 75.15% 79.23% 下载链接(code:fxw6)

模型压缩剪枝效果

这里使用mobilev3作为backbone,在icdar2015上测试结果,未压缩模型初始大小为2.4M.

  1. 对backbone进行压缩
模型 pruned method ratio model size(M) precision recall Hmean
DB no 0 2.4 84.04% 75.34% 79.46%
DB backbone 0.5 1.9 83.74% 73.18% 78.10%
DB backbone 0.6 1.58 84.46% 69.90% 76.50%
  1. 对整个模型进行压缩
模型 pruned method ratio model size(M) precision recall Hmean
DB no 0 2.4 85.70% 74.77% 79.86%
DB total 0.6 1.42 82.97% 75.10% 78.84%
DB total 0.65 1.15 85.14% 72.84% 78.51%

模型蒸馏

模型 teacher student model size(M) precision recall Hmean improve(%)
DB no mobilev3 2.4 85.70% 74.77% 79.86% -
DB resnet50 mobilev3 2.4 86.37% 77.22% 81.54% 1.68
DB no mobilev3 1.42 82.97% 75.10% 78.84% -
DB resnet50 mobilev3 1.42 85.88% 76.16% 80.73% 1.89
DB no mobilev3 1.15 85.14% 72.84% 78.51% -
DB resnet50 mobilev3 1.15 85.60% 74.72% 79.79% 1.28

文档教程


文本检测效果


Dbnet多语种文本检测效果

生成数据集:

公开数据集:


有问题及交流加微信

微信号:-fxwispig-


参考