we change the LSTM model to conv1D that more than 5 times faster than the original version on training base our dateset.half faster on training in ICDAR2013. conv1D with kernal 7 is almost same with lstm recall in ICDAR2013. Thanks for WeiyangWang and Beibei Hu's help.
text detection mainly based on ctpn (connectionist text proposal network). It is implemented in tensorflow. I use id card detect as an example to demonstrate the results, but it should be noticing that this model can be used in almost every horizontal scene text detection task. The origin paper can be found here. Also, the origin repo in caffe can be found in here. For more detail about the paper and code, see this blog. If you got any questions, check the issue first, if the problem persists, open a new issue.
- freeze the graph for convenient inference
- pure python, cython nms and cuda nms
- loss function as referred in paper
- oriented text connector
- for a quick demo,you don't have to build the library, simpely use demo_pb.py for inference.
- first, git clone firstname.lastname@example.org:eragonruan/text-detection-ctpn.git --depth=1
- then, download the pb file from release
- put ctpn.pb in data/
- put your images in data/demo, the results will be saved in data/results, and run demo in the root
there are some parameters you may need to modify according to your requirement, you can find them in ctpn/text.yml
- USE_GPU_NMS # whether to use nms implemented in cuda or not
- DETECT_MODE # H represents horizontal mode, O represents oriented mode, default is H
- checkpoints_path # the model I provided is in checkpoints/, if you train the model by yourself,it will be saved in output/
- requirements: python2.7, tensorflow1.3, cython0.24, opencv-python, easydict,(recommend to install Anaconda)
- if you do not have a gpu device,follow here to setup
- if you have a gpu device, build the library by
cd lib/utils chmod +x make.sh ./make.sh
- First, download the pre-trained model of VGG net and put it in data/pretrain/VGG_imagenet.npy. you can download it from google drive or baidu yun.
- Second, prepare the training data as referred in paper, or you can download the data I prepared from google drive or baidu yun. Or you can prepare your own data according to the following steps.
- Modify the path and gt_path in prepare_training_data/split_label.py according to your dataset. And run
cd lib/prepare_training_data python split_label.py
- it will generate the prepared data in current folder, and then run
- to convert the prepared training data into voc format. It will generate a folder named TEXTVOC. move this folder to data/ and then run
cd ../../data ln -s TEXTVOC VOCdevkit2007
- you can modify some hyper parameters in ctpn/text.yml, or just used the parameters I set.
- The model I provided in checkpoints is trained on GTX1070 for 50k iters.
- If you are using cuda nms, it takes about 0.2s per iter. So it will takes about 2.5 hours to finished 50k iterations.
oriented text connector
- oriented text connector has been implemented, i's working, but still need futher improvement.
- left figure is the result for DETECT_MODE H, right figure for DETECT_MODE O