Integration of Multi-Level Attention in U-Net with Generative Adversarial Network for Enhanced Underwater Visibility

Abstract

Underwater environment offers exceptional challenges like color cast, contrast problems and haze, hindering precise analysis. To increase visual perceptual quality of underwater images, we introduce, Integrated Generative Adversarial Network (IGAN), a novel approach that uses a combination of Generative Adversarial Networks (GANs), Spatial and Channel Attention and U-Net Architecture. The integration of attention mechanisms within GAN and U-Net architecture noticeably improves capability of model to effectively learn features essential for image enhancement. Our proposed model outrivals other methods in capturing and enhancing complicated details by concentrating on pertinent spatial and channel features. Firstly, we add dualattention mechanism to U-Net encoder which accurately extracts features in different channels and spatial dimensions. Secondly, we used Avg-TopK pooling in spatial attention which averages top specified K pixels within a feature map to retain most salient features while reducing dimensionality. Thirdly, one-sided label smoothing is used for better generalization and training stability. Finally, we used Efficient Channel Attention (ECA) module which is a lightweight attention mechanism that enhances feature representation by weighting each channel’s importance using a fast 1D convolution to capture local inter-channel interactions, where kernel size k denoting number of local inter-channel interactions, is adaptively selected. Comprehensive qualitative and quantitative analysis across various datasets, which includes U45 dataset, UIEB test dataset, UID dataset and UIEB challenge dataset, demonstrate extraordinary performance of ix IGAN in comparison with current state-of-the-art methods. In UIEB test dataset, IGAN achieves outstanding PSNR (26.53) and SSIM (0.905) values, outperforming MuLA-GAN, second best model, with values of 25.59 and 0.893, respectively representing 3.65% increase in PSNR and 1.34% improvement in SSIM. Experimental results illustrate that our proposed model surpasses others in robustness, especially in enhancing underwater images with distortion, turbidity, and various color shifts. Moreover, it also achieved optimal color accuracy and contrast levels to desired level of human visual perception while improving detailed texture of the images. Our proposed model surpasses other methods in capturing and enhancing intricate details by concentrating on relevant spatial and channel features.