zfxxfeng/cnn_lstm_ctc_ocr_for_icpr

Forked from weinman/cnn_lstm_ctc_ocr for the ICPR MTWI 2018 challenge 1

cnn_lstm_ctc_ocr_for_icpr's Issues

denseNet

Maybe the model denseNet is not convergent. I think denseNet need some improvement.

raw data set

链接: https://pan.baidu.com/s/1lvSVyXQv6y2YXxG0HJXBhw 密码: mck3

validate.py使用问题

请问当初你是如何使用validate.py对整个测试集进行遍历呢?因为我看到输入图像要求维度是(31,?,1)而显然需要对测试数据进行resize,那么请问你用什么方法对测试数据进行resize呢?我直接用了image.resize((image.size[0], 31),最终效果并不好,请问你是用什么方法呢？

crop出来的文本图片不正常

你好，你的生成文本图片的函数，生成的文本图片感觉都是做了镜像？
本来水平文本都镜像了？
这样回影响训练吗？

How many steps do you run when training denseNet on IPCR dataset?

Much thanks to your work!

I haved trained the network for 70k steps and the effect is still bad. As your uploaded model's step is 1405532, I really need a long way to go, right?

Another question is, model.ckpt-1405532 will predict 5530 characters, but the word_dict.py contains 16389 characters. Where's the corresponding word_dict for the demo model?

如何利用jpg文件和txt文件生成tfrecord?

我下载了标注文件，有jpg图片和txt标注，怎么生成tfrecord呢？用mjsynth-tfrecord.py的话，里面需要的annotation_val.txt文件是什么内容，我看后续处理是文件名 xxx，这后面xxx是什么iamge_texts?是干嘛的

How do you handle vertical image?

RT.

target_cut.py 可能有点问题

首先感谢分享.

Image_list = os.listdir(Img_dataset_dir)

Label_list = os.listdir(Label_dataset_dir)
直接读取文件名可能会用顺序不一样的问题, macos python3下以及centos python3下都会有, windows下好像不会

然后处理结果都是镜像翻转正常么...

version of tf and python?

version of tf and python?

where is target_cut.py

验证自己训练的model的时候出错

我用了你的预训练模型训练了大概30万次
然后 我输入python validate.py出现下面错误

何解？

There is some problems in demo, which some weights in checkpoint not found.

NotFoundError (see above for traceback): Key rnn/bdrnn2/fw/lstm_cell/weights not found in checkpoint
[[Node: save/RestoreV2_39 = RestoreV2[dtypes=[DT_FLOAT], _device="/job:localhost/replica:0/task:0/cpu:0"](_recv_save/Const_0, save/RestoreV2_39/tensor_names, save/RestoreV2_39/shape_and_slices)]]

Can you know how to solve it?