MADE-Net: multiscale and adaptive detail enhanced network for single image dehazing

Image dehazing is a typical task in low level of vision. Its goal is the estimation of the observed haze image has the potential to produce a haze-free image. Previous studies have confirmed that increasing convolutional depth and kernel size enhances the effectiveness of dehazing, but the exploration of multi-scale information and the learning capabilities of CNN structures remains insufficient. In the present paper, a novel architectural design is put forth, composed of multi-scale adaptive detail enhanced convolution (MADEC) and shareable attention modules. Specifically, MADEC leverages its larger receptive field to extract more diverse scale-dependent features, thereby improving model representation and generalization ability. Additionally, the shareable attention module integrates channel, spatial, and pixel-level hierarchical attention mechanisms, enabling the fusion of more useful information. A comprehensive series of experiments conducted on three distinct datasets has substantiated the efficacy of the proposed methodology.