Image Restoration Using Multi-Stage Progressive Encoder-Decoder Network With Attention and Transfer Learning (MSP-ATL)
Bingcai Wei, Di Wang, Zhuang Wang, Liye Zhang, Cong Liu · 2022
Convolutional neural network (CNN) is widely used in the field of image restoration, however, most existing CNN based image restoration methods only focus on a part of image restoration without considering the relationship between image deblurring, image de-raining and image denoising. In addition, training a neural network model for different image restoration tasks requires a large amount of training data, plenty of hardware overheads and a great deal of time. In order to reduce resource consumption and improve the generality of the model, this paper proposes an image restoration algorithm using multistage progressive encoder-decoder network with attention and transfer learning (MSP-ATL). First of all, we design a multi-stage progressive encoder-decoder network with attention mechanisms, which does not require down-sampling or fragmentation of the input, and can retain the overall information of the image. Secondly, following the idea of coarse-to-fine which is widely used in the field of image restoration, we propose a multistage progressive loss function to recover images from coarse to fine by cooperating with the multi-stage network structure. Finally, using transfer learning, the obtained deblurring model can be transferred to image de-raining and image denoising tasks with less training data and simple training process. Extensive experimental results on commonly used datasets demonstrate the efficiency and effectiveness of the proposed method.