Hazy Image Dehazing Algorithm Based on Two Branch Residual Feature Fusion

Aiming at the problem that the existing dehazing algorithms based on convolution neural network can not effectively remove the non-uniform distribution of haze in the real-haze image, an end-to-end image dehazing algorithm based on two branch residual feature fusion network is designed. The attention branch in context spatial domain pays attention to the high-frequency hazy region of hazy image. The spatial attention module is inserted into the multi-scale expansion convolution group to assign the weight to the pixel of hazy characteristics. The channel attention branch pays attention to the channel of high-frequency hazy features, sets ResNet self-coding structure, and introduces channel attention decoding structure to assign the weight of different channel feature maps. The feature fusion module uses adaptive weights to fuse the hazy layer feature information of pixel attention and channel attention, and outputs the uneven hazy residual layer. The original hazy image and the hazy residual layer are compared to achieve image dehazing, a discrimination network is designed to improve the visual appearance of the dehazing image. The experimental results of evaluating the algorithm with real-haze images NH-Haze show that the algorithm has good visual effect on non-uniform hazy images, and is superior to other comparison algorithms in peak signal-to-noise ratio and structural similarity evaluation.