rcnn网络tensorflow实现

# OCR_TF_CRNN_CTC This software implements the Convolutional Recurrent Neural Network (CRNN), a combination of CNN, RNN and CTC loss for image-based sequence recognition tasks, such as scene text recognition and OCR. "An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition" : https://arxiv.org/abs/1507.05717 More details for CRNN and CTC loss (in chinese): https://zhuanlan.zhihu.com/p/43534801 # Dependencies All dependencies should be installed are as follow: * tensorflow==1.8.0 * opencv-python * numpy Required packages can be installed with ```bash pip install -r requirements.txt ``` Note: This software cannot run in the tensorflow lastest version r1.11.0 since it's modified the tf.contrib.rnn API. # Run demo Asume your current work directory is OCR_TF_CRNN_CTC： ```bash cd path/to/your/OCR_TF_CRNN_CTC/ ``` Dowload pretrained model and extract it to your disc: [GoogleDrive](https://drive.google.com/file/d/1A3V7o3SKSiL3IHcTqc1jP4w58DuC8F9o/view?usp=sharing) . Export current work directory path into PYTHONPATH: ```bash export PYTHONPATH=$PYTHONPATH:./ ``` Run inference demo: ```bash python tools/inference_crnn_ctc.py \ --image_dir ./test_data/images/ --image_list ./test_data/image_list.txt \ --model_dir /path/to/your/bs_synth90k_model/ ``` Result is: ``` Predict 1_AFTERSHAVE_1509.jpg image as: aftershave ```  ``` Predict 2_LARIAT_43420.jpg image as: lariat ```  # Train a new model ### Data Preparation * Firstly you need to download [Synth90k](http://www.robots.ox.ac.uk/~vgg/data/text/) dataset and extract it into a folder. * Secondly supply a txt file to specify the relative path to the image data dir and it's corresponding text label. For example: image_list.txt ```bash 90kDICT32px/1/2/373_coley_14845.jpg coley 90kDICT32px/17/5/176_Nevadans_51437.jpg nevadans ``` * Then you are suppose to convert your dataset into tensorflow records which can be done by ```bash python tools/create_crnn_ctc_tfrecord.py \ --image_dir path/to/90kDICT32px/ --anno_file path/to/image_list.txt --data_dir ./tfrecords/ \ --validation_split_fraction 0.1 ``` Note: make sure that images can be read from the path you specificed, such as: ```bash path/to/90kDICT32px/1/2/373_coley_14845.jpg path/to/90kDICT32px/17/5/176_Nevadans_51437.jpg ....... ``` All training image will be scaled into height 32 and write to tfrecord file. The dataset will be divided into train and validation set and you can change the parameter to control the ratio of them. #### Otherwise you can use the dowload_synth90k_and_create_tfrecord.sh script automatically create tfrecord: ``` cd ./data sh dowload_synth90k_and_create_tfrecord.sh ``` ### Train model ```bash python tools/train_crnn_ctc.py --data_dir ./tfrecords/ --model_dir ./model/ --batch_size 32 ``` After several times of iteration you can check the output in terminal as follow:  During my experiment the loss drops as follow:  ### Evaluate model ```bash python tools/eval_crnn_ctc.py --data_dir ./tfrecords/ --model_dir ./model/ ``` 为了将特征输入到Recurrent Layers，做如下处理： 首先会将图像缩放到 32*W*3 大小 然后经过CNN后变为 1* (W/4)*512 接着针对LSTM，设置 T=(W/4) ， D=512 ，即可将特征输入LSTM。 所以在处理输入图像的时候，建议在保持长宽比的情况下将高缩放到 32，这样能够尽量不破坏图像中的文本细节。当然也，也可以将输入图像缩放到固定宽度，但是这样肯定会造成性能下降。 字符转义，"\"" : vlaue. warning: tensorflow/core/util/ctc/ctc_loss_calculator.cc:144] No valid path found. It turns out that the ctc_loss requires that the label lengths be shorter than the input lengths. If the label lengths are too long, the loss calculator cannot unroll completely and therefore cannot compute the los. 输入的序列长度必须 >= label 的长度，否则无法计算 CTC loss，换句话说，识别出的字符长度可以少于输入的序列长度但是 不能比它长。 需要在char_map/char_mao.json 中添加英文或中文符号，添加格式"&" : 56