Datasets of degraded document image are small, making the network unable to be fully trained or easily over-fitting. And single-convolution network has poor generalization ability. These factors lead to an unsatisfactory binarization performance. This paper proposes a degraded document image binarization method based on U-Net and transfer learning to solve these problems. U-Net is used as our model’s backbone for its good performance in small datasets. The common transfer learning network models ResNet is utilized as the pre-training encoder to improve the generalization ability of our model. Then we establish different decoder network structures for the characteristics of different encoders. In addition, different from conventional U-Net, the convolutional layer output of downsampling is made as the skip connection object to be superimposed with the input of upsampling in our models, so the upsampling layers can better restore the details of document images. In this way, we improve the convergence and generalization ability to get a better binarization performance.
|