Authors: Gaoyang Jia, Jinlun Huang, Zehao Huang, Man Wang, Huaner Song, Donggen Xiong, Zongfei Li
Description: Since the development of text detection(or target detection), the method based on full convolution network and pixel segmentation keeps refreshing records in various competitions.We have also done a lot of research in this area, and have some applications in some specific occasions.
We find that it is not difficult to classify objects at the pixel level. The difficulty lies in how to deal with the relationships and boundaries of different objects. The PSEnet undoubtedly provides us with a new way of thinking. We adopted this method decisively and made some adjustments based on our own experience:
1. Replace the feature network.
2. Modify the network structure after the feature network.
3. Change the loss function, and expand the scope of mask.
4. Some other changes in data processing and training methods.