The scattering and absorption of light in aquatic environments often result in significant deterioration of underwater images, manifesting as color distortion, reduced contrast, and loss of information. These challenges significantly restrict the effectiveness of underwater vision systems utilized in marine exploration, robotic surveillance, and ecological research. The pursuit of visually appealing results with reliable detail recovery and balanced computational cost remains a significant challenge, despite the numerous underwater image enhancement techniques that have been proposed. This paper introduces SETAU-Net (Sobel Enhancement with Transformer Attention U-Net), a novel architecture that is specifically engineered to enable effective underwater image enhancement. SETAU-Net was trained on the LSUI, EUVP and UFO-120 datasets using a composite loss function that integrates L1 loss, L2 loss, Structural Similarity Index Measure loss, VGG-based perceptual loss, and Laplacian loss, thus exhibiting exceptional quantitative performance in various benchmarks. SETAU-Net demonstrates robust quantitative performance across many underwater benchmarks, with average scores of 25.90 PSNR (dB), 0.86 SSIM, and 0.14 LPIPS on EUVP; 28.96 PSNR (dB), 0.92 SSIM, and 0.07 LPIPS on LSUI and 27.28 PSNR (dB), 0.87 SSIM, and 0.12 LPIPS on UFO-120. SETAU-Net achieves this level of enhancement while maintaining a computational footprint of only 3.36 million parameters and 31.1 GFLOPs for a 256 × 256 input, demonstrating a notable balance between performance and efficiency. Qualitative assessments demonstrate SETAU-Net’s ability to restore natural colors, enhance contrast, and recover subtle characteristics obscured in degraded underwater environments. SETAU-Net provides a dependable and efficient method for practical applications in underwater environments ranging from real-time surveillance to resource-constrained exploration.