Abstract: The presence of fog severely reduces the image quality and hinders further image processing as well as the corresponding feature information extraction. An end-to-end image defogging algorithm based on a multi-scale hole residual block and multi-scale attention mechanism is proposed to address the problems of inadequate feature extraction of existing defogging algorithms. First, the shallow features of the fog map were extracted by three small-scale convolution kernels, which can obtain a large receptive field and reduce the number of parameters. Then, it was input into a plurality of network modules composed of a multi-scale residual hole convolution feature extraction module and multi-scale attention mechanism module in series, in which the multi-scale hole convolution residual feature extraction module could extract the fog map features of different receptive fields and fuse the features of different dimensions to effectively solve the problem of single feature scale, while the multi-scale attention mechanism module reasonably allocated the weights of different features and suppressed irrelevant redundant information. Finally, the fog features in the fog map were filtered to obtain the feature map of the defogging map, followed by restoring the fog-free image using convolution operation. By testing on the sots test set, better visual effects are obtained compared with other classical methods, and the performance on PSNR and SSIM is also better than other classical methods.

Key words: image defogging, residual hole convolution, attention mechanism, feature extraction, deep learning