PyTorch re-implementation of Real-time Scene Text Detection with Differentiable Binarization

1. Use dice loss instead of BCE(binary cross-entropy) loss.

2. Use normal convolution rather than deformable convolution in the backbone network.

3. The architecture of the backbone network is a simple FPN.

4. Have not implement OHEM.

5. The ground truth of the threshold map is constant 1 rather than 'the distance to the closest segment'.

thanks to these project:

• https://github.com/WenmuZhou/PAN.pytorch

The features are summarized blow:

• Use resnet18/resnet50/shufflenetV2 as backbone.

1. Installation
2. Download
3. Train
4. Predict
5. Eval
6. Demo

1. pytorch 1.1.0

1. ShuffleNet_V2 Models trained on ICDAR 2013+2015 (training set) https://pan.baidu.com/s/1Um0wzbTFjJC0jdJ703GR7Q

1. modify genText.py to generate txt list file for training/testing data

2. modify config.json

3. run

python train.py

1. run

python predict.py

run

python eval.py

• MobileNet backbone

• Deformable convolution

• tensorboard support

• FPN --> Architecture in the thesis

• Dice Loss --> BCE Loss

• threshold map gt use 1 --> threshold map gt use distance （Use 1 will accelerate the label generation）

• OHEM

• OpenCV_DNN inference API for CPU machine

• Caffe version (for deploying with MNN/NCNN)

• ICDAR13 / ICDAR15 / CTW1500 / MLT2017 / Total-Text