Complex water surface image rain removal lightweight network based on convolution attention

Aiming at the problem that the existing algorithms are difficult to accurately identify rain streaks and water ripples and there is a loss of background information, an image de-raining lightweight network based on convolutional attention is proposed. Firstly, the ShuffleNet V2 module is used to extract features to keep the model lightweight and promote the information flow between features. Secondly, the convolutional attention CBAM is introduced to adjust the weight information of the water surface ripple and the rain ripple, and the rain ripple image and the background image are accurately separated. Finally, a detail recovery network is constructed based on the residual block to ensure the details and quality of the rain removal image and avoid excessive smoothing or loss of details. The experimental results show that compared with other algorithms, the proposed algorithm can effectively remove the rain streaks while retaining the texture information of the original image, and has good rain removal effect and generalization ability.