caffe_ocr

主流ocr算法研究实验性的项目，目前实现了CNN+BLSTM+CTC架构

Stars
1.3K
View Code on GitHub
Ecosystems: C++

caffe_ocrocrCNN+BLSTM+CTClstmwarp-ctcmulti-labelinceptionrestnetdensenetwindowslinuxcaffe

caffe

  1. data layermulti-label
  2. lstmjunhyukohlstmlstm_layer_Junhyuk.cpp/cushape(TxN)xHTxNxHctc
  3. WarpCTCLossLayersequence indicatorsCNNsequence indicator layer
  4. densenetReshapebug
  5. transpose_layerreverse_layerCNN feature maplstmshape

  1. opencv,boost,cuda,3rdpartydebuglib:http://pan.baidu.com/s/1nvIFojJ
  2. caffe-vsprojvs2015tools_bincaffe.exe
  3. dllhttp://pan.baidu.com/s/1boOiscJ

  1. caffe-vsproj/libClassification,examples\cpp_classification
  2. caffe-vsproj/ocr_testocrocrexamples/ocr/lexicon.txt

1VGG Synthetic Word Dataset 2https://pan.baidu.com/s/1TXeZUcpVOUxrtfRtWKMSGg ujjc

  • +
  • 5990
  • 10
  • 280x32
  • 3609:16

examples/ocr 3.

  • VGG Synthetic Word Dataset:
predict-CPU predict-GPU -no lexicon -lexicon-minctcloss
crnn 67.13ms 10.28ms 0.8435 0.9163 32MB
inception-bn-res-blstm 41.62ms 8.68ms 0.7353 0.8609 15MB
densenet-res-blstm N/A 6.07ms 0.7548 0.893 11MB
  • :
predict-CPU predict-GPU
inception-bn-res-blstm 65.4ms 11.2ms 0.92 26.9MB
resnet-res-blstm 64ms 10.75ms 0.91 23.2MB
densenet-res-blstm N/A 7.73ms 0.965 22.9MB
densenet-no-blstm N/A 2.4ms 0.97 5.6MB
densenet-sum-blstm-full-res-blstm N/A 7.83ms 0.9805 15.5MB
densenet-no-blstm-vertical-feature N/A 3.05ms 0.9816 9.6MB

http://pan.baidu.com/s/1i5d5zdN

  1. tricks
    1 blstm0.94-->0.965BLSTM0.9805
    2 CNNBLSTM

1blstmCNN+CTCdensenet-no-blstm aCNNb)lstm 6. 111llCNN

TODO

  1. : (https://yq.aliyun.com/articles/109555?t=t1)
  2. traintest lossAttention,STN,loss

  1. crnn(https://arxiv.org/abs/1507.05717)
  2. multi-label(http://blog.csdn.net/hubin232/article/details/50960201)
  3. junhyukohlstmhttps://github.com/junhyukoh/caffe-lstm
  4. caffe-warp-ctc(https://github.com/BVLC/caffe/pull/4681)
  5. memory-efficient densenet(https://github.com/Tongcheng/caffe/)
  6. http://www.sogou.com/labs/resource/cs.php
Related Projects
License-Plate-Detect-Recognition-via-Deep-Neural-Networks-accuracy-up-to-99.9

works in real-time with detection and recognition accuracy up to 99.8% for Chinese license plates...

29 Jun 2018 1,382