Dehazing network based on residual global contextual attention and cross-layer feature fusion

Current deep learning-based image dehazing algorithms usually use traditional convolutional layers when extracting features, which easily cause loss of information, such as details and edges of the image, ignore the location information of the image in feature extraction, and neglect the original information of the image in feature fusion, and they thus fail to recover a high-quality dehazing image with complete and clear structure. To address this problem, a dehazing algorithm based on residual contextual attention and cross-layer feature fusion was proposed. Firstly, the residual group structure was obtained by serializing the proposed residual contextual blocks, and feature extraction was performed on the first two layers of the network, i.e. the shallow layers, to obtain rich contextual information in the shallow layers; secondly, coordinate attention was introduced to build an attention graph with location information and apply it to the residual contextual feature extraction, which was placed in the third layer of the network, i.e. the deep layer, to extract deeper semantic information; then, by fusing feature information from different resolution streams across layers in the middle layer of the network, the information exchange between the deep and shallow layers was enhanced to achieve feature enhancement; finally, the semantic information-rich features obtained from the network were combined with the original input, thus enhancing the recovery effect. Experimental results on the RESIDE dataset and the Haze4K dataset show that the proposed algorithm achieves better results in terms of visual effects and objective metrics.