SwinCNet leveraging Swin Transformer V2 and CNN for precise color correction and detail enhancement in underwater image restoration

Abstract

Underwater image restoration confronts three major challenges: color distortion, contrast degradation, and detail blurring caused by light absorption and scattering. Current methods face difficulties in effectively balancing local detail preservation with global information integration. This study proposes SwinCNet, an innovative deep learning architecture that incorporates an enhanced Swin Transformer V2 following primary convolutional layers to achieve synergistic processing of local details and global dependencies. The architecture introduces two novel components: a dual-path feature extraction strategy and an adaptive feature fusion mechanism. These components work in tandem to preserve local structural information while strengthening cross-regional feature correlations during the encoding phase and enable precise multi-scale feature integration during decoding. Experimental results on the EUVP dataset demonstrate that SwinCNet achieves PSNR values of 24.1075 dB and 28.1944 dB on the EUVP-UI and EUVP-UD subsets, respectively. Furthermore, the model demonstrates competitive performance in reference-free evaluation metrics compared to existing methods while processing 512×512 resolution images in merely 30.32 ms—a significant efficiency improvement over conventional approaches, confirming its practical applicability in real-world underwater scenarios.