Single-frame remote sensing image defogging network based on attention residual blocks and contrast learning

Abstract

High-resolution remote sensing image defogging can bring a lot of convenience to downstream image tasks, however, existing remote sensing image defogging algorithms have many drawbacks. In this paper, an O-type network structure for remote sensing image defogging is proposed. On the one hand, a flipped inverted attention residual block is designed to solve the shortcomings of information loss and gradient explosion of the basic residual block, so that the network can fuse features in multiple scales and directions to provide more detailed edge information and texture information for the decoding stage. On the other hand, a multi-loss function constraint network is designed to improve the shortcomings of traditional overall learning by making full use of positive and negative samples to help the network learn a kind of ‘mapping’ to achieve the transformation from fog to fog-free and adding structural similarity loss to make the output image with more detailed edges. The colour distortion is mitigated. In addition, a data set RSI_Haze is proposed for remote sensing image defogging, and fogs with different transmittance are synthesized to obtain a data set closer to the real one with higher richness to train the defogging network, which makes the parameters of the network more robust and can remove fogs of different thicknesses adaptively. Experiments show that the network has a better recovery effect in high-resolution remote sensing images than other SOTA networks, enabling remote sensing images to play more value for human activities.