Residual-and-attention-based dehazing network for non-homogeneous hazy images
Guoliang Jia · 2024
Image dehazing is a classic low-level vision task. Applying dehazing algorithms can improve image quality. Many dehazing algorithms have achieved good results on synthetic datasets, but real-world non-homogeneous dehazing is still a challenge. This paper proposes a residual-and-attention-based dehazing network (RAD) with a generative adversarial network training framework. The generator adopts an encoder-decoder network structure, including encoder, bottleneck, and decoder, with adaptive mixup operation between the encoder and decoder. Using the first three residual layers of a pre-trained ResNet model as the encoder provides the network with strong feature extraction capabilities. Attention modules combining channel attention and pixel attention are used in the bottleneck layer and decoder to enhance the dehazing network's effectiveness against non-homogeneous haze. Adaptive mixup operation is employed to connect different feature maps. Adaptive mixup operation helps the network better preserve shallow image features. Experimental results show that the proposed RAD dehazing algorithm achieves superior performance on NH-HAZE.