A lightweight underwater image and video enhancement method based on multi-scale feature fusion

Underwater image degradation resulting from light absorption and scattering in water significantly affects the visual perception of autonomous underwater vehicles (AUVs), subsea infrastructure inspections, and marine environmental monitoring. We propose a lightweight enhancement algorithm (UIVE) for underwater images and videos through a lightweight design of residual block replacement BN layer. The innovation of this study is reflected in the dynamic combination of downsampling features and upsampling process through multi-scale jump connection, which effectively retains multi-scale detail information. Replacing the batch normalization (BN) layer strategy with residual blocks effectively reduces computational complexity while maintaining model performance and significantly enhances the effectiveness of feature migration. The adaptive brightness correction module is introduced to compensate for the problem of image brightness reduction after the nonlinear mapping model. Experiments show that UIVE achieves a PSNR value of 18.57 dB and a SSIM value of 0.91 on the UIEB dataset, respectively, and especially demonstrates an obvious advantage in removing bias color. In terms of real-time processing, UIVE achieves a frame rate of 36.23 frames per second (FPS) with a video resolution of 640×480 pixels and can be deployed on embedded systems in AUV or ROV. Experimental results demonstrate that UIVE excels in color correction and detail reproduction, making it particularly well-suited for applications in underwater environments within marine engineering.