search

watsonyanghx / CNN_LSTM_CTC_Tensorflow

CNN+LSTM+CTC based OCR implemented using tensorflow.
MIT License
362 stars 210 forks source link

测试集精度问题。 #29

Open boris-lb opened 6 years ago

boris-lb commented 6 years ago

你好，我使用自己的数据在测试集上精度还不错，但是检查了一下错误的都是重叠的字符的漏检，下面是部分测试集结果： 38387_1077.jpg 107 38388_100,005.jpg 100,05 38389_1077.jpg 107 38393_100,005.jpg 10,05 38394_1077.jpg 107 38640_131,005.jpg 131,05 38797_61,051.jpg 61,0651 39128_44,438.jpg 4,438 39545_157,333.jpg 157,33 4876_314,554.jpg 314,54 4878_268,866.jpg 268,86 5223_111,055.jpg 111,05 5276_211,904.jpg 21,904 546_32,772.jpg 32,72 571_128,883.jpg 128,83 664_148,733.jpg 148,73 672_144,150.jpg 14,150 7218_102,332.jpg 102,32 7221_100,267.jpg 10,2657 7654_77,132.jpg 7,132 7731_215,223.jpg 215,23 7787_111,702.jpg 11,702 7791_104,773.jpg 104,73 请问针对这个问题有什么好点的解决方案吗？应该调整哪里？谢谢

jihadbourassi commented 5 years ago

@boris-lb Hi can you please tell me how did you get those results on infer mode? i tried to infer on my own dataset and when i run the command nothing happens

boris-lb commented 5 years ago

@boris-lb Hi can you please tell me how did you get those results on infer mode? i tried to infer on my own dataset and when i run the command nothing happens

this is my test.sh file： CUDA_VISIBLE_DEVICES=3 python ./main.py --infer_dir=./data/zhongdeng1/infer/ \ --checkpoint_dir=./checkpoint/checkpoint_zhongdeng1/ \ --image_height=60 \ --image_width=200 \ --num_hidden=256 \ --num_gpus=1 \ --mode=infer if you run the command and nothing happened，please check you path of model and infer data。