New underwater image enhancement technique using single scale retinex and CNN fusion

Improving the quality of underwater images remains a significant challenge for the scientific community due to high color distortion, low contrast, and limited visibility. To address these issues, our study introduces a new methodology integrating the Single Scale Retinex (SSR) algorithm with a Convolutional Neural Network (CNN) to enhance underwater image quality. Our approach employs a unique dual-input CNN architecture, utilizing both the raw image and its SSR-enhanced version as inputs. The process begins with applying the SSR algorithm to the raw images. The resulting SSR-enhanced image is then concatenated with the CNN's first and second convolutional blocks, enhancing feature representation at these stages. A fusion strategy is also introduced where the raw image input is concatenated with the output of the third convolutional block. This tactic preserves crucial low-level details while incorporating enriched representations from the SSR-enhanced images at key points in the CNN structure. We developed this methodology to validate the integration of classical image enhancement techniques with modern deep-learning methods, achieving substantial improvements. Extensive evaluations of the widely used UIEB and EUVP underwater image databases demonstrate significant improvements in contrast, illumination, and overall image quality metrics. Our model achieved a PSNR of 23.50, SSIM of 0.92, and IUQM of 0.97 with a processing time of 0.28 seconds, outperforming state-of-the-art methods such as Ucolor (PSNR 21.86, SSIM 0.89) and WaterNet (PSNR 19.81, SSIM 0.86) in terms of image quality and processing efficiency. Compared to other techniques, our proposed methodology shows promising results in enhancing underwater image quality, with potential applications across various underwater imaging scenarios.