Abstract: Image dehazing is a challenging visual task. Previous image dehazing method often depend too much on the physical model of images degraded by fog, and the current image dehazing model using convolution neural network is more complex. Therefore, a lightweight dehazing network MADNet that does not depend on physical model is proposed. The network is mainly composed of a multi-scale convolution module with attention mechanism. By viewing foggy images as composed of clear images and fog residue images, MADNet directly learns the fog residue between the target clear image and the input foggy image, and finally achieve end-to-end image fog removal. The experimental results show that the structure similarity and peak signal-to-noise ratio of the proposed method are better than those of other comparison method on SOTS and NH-HAZE datasets, and it can also achieve better fog removal in real scenes.

Key words: image dehazing, lightweight network, attention mechanism, multi-scale convolution