Real-time enhancement of low-light underwater images based on adaptive Gaussian filtering and adaptive GAN

Low-light underwater images face high noise and poor contrast, limiting practical applications. This study proposes an enhancement framework combining adaptive Gaussian filtering and adaptive GAN. The preprocessing step adaptively adjusts Gaussian template parameters via pixel neighborhood standard deviation, suppressing noise while preserving edges and contours. The GAN’s generative network (encoder, dilated convolution group, decoder, edge enhancement module) expands receptive fields, maintains feature map size, recovers details, and fuses multiscale features via channel-concatenated skip connections. Validated on 2000 images (950 from UIEB, 1050 network-acquired), comparative analysis with two methods using UIQM and UCIQE metrics shows the proposed method outperforms in brightness, color reproduction, and detail preservation (max UIQM=4.1, UCIQE=0.78), even for small samples. It provides an effective real-time solution meeting practical needs.