Multi-scale Fusion Residual Dense Dehazing Network

Abstract

A multi-scale fusion residual dense dehazing network is proposed to address the problems of high model complexity and incomplete dehazing in most existing end-to-end dehazing algorithms. In the residual dense block (DSRB), the feature information is extracted by increasing the convolutional kernel perceptual field using smooth dilation convolution, and the continuous memory mechanism is formed by means of dense jump connections to improve the characterization ability of the network. DWC and PWC convolution are used to generate channel attention and pixel attention mechanisms to fuse shallow features and deep features at each scale and focus on target features adaptively. Finally, the method of this paper is tested on the SOTS test set of RESIDE, and the results show that the model complexity of this paper's method is low, and the PSNR and SSIM metrics are improved, and the visual effect is good.