基于Caffe的HWDB手写汉字识别

基于以上模型结构, 迭代21万次(batchsize 168),约13epoch,top1 = 0.967285,top3 = 0.993488,模型大小:30.5MB,输入图像64*64(转换HWDB数据库为64*64图像,参考[6][7][8]),训练数据HWDB1.0+1.1（总共2677697个汉字,3755类）,测试数据:224419个

目前单模型识别率最高的是文献[1]提到的HCCR-CNN12Layer, 识别率:97.59%，模型大小:48.7MB，输入图像96*96，训练数据HWDB1.0+1.1

训练log:

I0608 15:14:12.734665 1176 solver.cpp:330] Iteration 210000, Testing net (#0)
I0608 15:18:56.108808 10288 data_layer.cpp:73] Restarting data prefetching from start.
I0608 15:18:57.733983 1176 solver.cpp:397] Test net output #0: accuracy_top1 = 0.967285
I0608 15:18:57.733983 1176 solver.cpp:397] Test net output #1: accuracy_top3 = 0.993488
I0608 15:18:57.733983 1176 solver.cpp:397] Test net output #2: loss = 0.161584 (* 1 = 0.161584 loss)
I0608 15:18:58.187183 1176 solver.cpp:218] Iteration 210000 (0.161847 iter/s, 308.933s/50 iters), loss = 0.102787
I0608 15:18:58.187183 1176 solver.cpp:237] Train net output #0: loss = 0.235711 (* 1 = 0.235711 loss)
I0608 15:18:58.187183 1176 sgd_solver.cpp:105] Iteration 210000, lr = 0.000559872

                                                           testAccuracy

                                                         testLoss

                                                              trainLoss

[1] HWDB 97.59%(Single) <<Building Fast and Compact Convolutional Neural Networks for Offline Handwritten Chinese Character Recognition >>
[2] https://github.com/HCIILAB/Forward-Implementation-of-Fast-and-Compact-CNN-for-Offline-HCCR
[3] HWDB 97.37%(Single) <<online and Offline Handwritten Chinese Character Recognition A Comprehensive Study and New Benchmark>>
[4] HWDB 手写汉字数据库说明: Offline Database 
[5] HWDB 手写汉字数据库下载: Download
[6] TensorFlow与中文手写汉字识别 - 知乎专栏
[7] burness/tensorflow-101
[8] TensorFlow练习22: 手写汉字识别