DAMNet: lightweight dual attention mixed network for efficient image deraining

A long-standing problem in computer vision (CV) is image deraining. Current deraining networks frequently fail to achieve a good balance between low system complexity and great image quality, particularly when considering spatial features and contextual feature map information. We propose the Dual Attention Mixed Network (DAMNet) as a solution, which achieves a balance between these conflicting objectives. Encoder-Decoder, Dual Attention Network (DANet), and Channel Attention Sub-Module Network (CASMNet) are the three sub-networks which jointly make up DAMNet. It has two hierarchical levels. The Encoder-Decoder and DANet operate at the first level, while CASMNet operates at the second hierarchical level. An enhanced U-Net serves as the base for the Encoder-Decoder architecture, which effectively learns contextual feature map information. The main objective of CASMNet is to preserve local information. In order to reduce network complexity, a window multi-head self-attention network is used in place of the Middle Block (MB) and multiplication is applied in place of the nonlinear activation function in the U-Net encoder-decoder. DAMNet can gradually retrieve contextual information and spatial details from degraded images through this approach. Our network performs better than many current networks on the Rain100H dataset, reaching at least 0.55 dB greater PSNR, although only using 16.3 million parameters. In comparison with state-of-the-art (SOTA) networks, extensive testing on synthetic, real-world, and raindrop datasets show that DAMNet yields promising results. Moreover, the implementation of DAMNet on a low-end edge device such as the Jetson Nano demonstrates its effectiveness in practical world environment.