A lightweight attention-based network for image dehazing

In current convolution-based image dehazing networks, increasing the depth and width of convolutional layers is a common strategy to improve network performance. However, this approach significantly increases the complexity and computational cost of the dehazing network. To address this issue, this paper proposes a U-shaped multi-scale adaptive selection network (UMA-Net). Without introducing additional parameters and computational costs, the network leverages the influence of different scales of convolutional kernels on the receptive field. It combines standard and dilated convolutions into the feed-forward network (FFN) to propose a multi-scale adaptive (MA) dehazing module, further expanding the receptive field and focusing on important spatial and channel information within the FFN. To fully exploit the multi-scale features of the MA module, a lightweight channel attention-guided fusion (CAGF) module is proposed, which achieves the restoration of high-quality dehazed images from hazy images. Extensive experiments demonstrate the effectiveness of the proposed modules. On the Reside SOTS dataset, it achieves state-of-the-art performance with only 0.816M parameters and 8.794G FLOPs.