Attention-based Bi-Stream network for underwater image enhancement

Enhancing underwater images is difficult because of degradation caused to visibility, contrast, and color distortion due to scattering, light absorption in underwater environments. In this paper, we present a new multi-stream deep learning approach, UWIEn_BiAlCn to enhance underwater images. The proposed framework begins with preprocessing, then employs a Bi-Stream feature extraction networks: an Improved AlexNet with Pyramid Dilated Convolution (ImAlPDC) and an Improved CNN with Spatial Pyramid Hybrid Convolution (ImCSPHC), to capture both local and global contextual features at multiple scales. The bi-scale features generated are fused using an attention-based fusion module, ensuring optimal feature integration and relevance weighting. A subsequent decoder network reconstructs the enhanced underwater image with improved perceptual quality. Experimental results demonstrate superior performance, achieving PSNR up to 39.12, SSIM of 0.988, UIQM of 4.87, and UCIQE of 0.758 on UIEB, outperforming conventional single-stream models and existing enhancement methods across multiple benchmark datasets.