Multi-Scale and Attention Residual Network for Single Image Dehazing
Jingyuan Zhou, Chak Tou Leong, Congduan Li · 2021
Image dehazing has always been a thorny ill-posed problem in low-level vision. In this paper, we propose a new model which combines feature pyramid network and attention based resblock, which is called the pyramid attention residual feature extraction network (PARFE-Net). The architecture of PARFE-Net can be summarized as follows:(I)The feature pyramid network module is used to extract features from the input image. Multi-scale can extract different features on different scales. The extracted features have great semantic differences in different depths.(II) Innovatively integrate channel-wise attention and pixel-wise attention and apply them to residual block. Attention mechanism can make the weight of the model higher in important places, which significantly affects the quality of the output image. (III) Combine the resblocks in II to get a large residual network, which is called iterative attention resblock. The features of different dimensions extracted from I are concatenated according to the channel dimension, which is used as the input of iterative attention resblock. The experiment is based on NYU2 and RESIDE. Compared with the commonly used dehazing model, the evaluation index PSNR and SSIM of our approach have achieved considerable or even better results, with better robustness.