Underwater Imaging plays a vital role in exploring the features of marine ecosystem. These images are affected by various factors like scattering of light, color distortion and inadequate light absorption which causes degradation in quality and visibility of images in aquatic environments. The dispersion and refraction of light may cause distortions in image as light penetrates from lighter to denser medium underwater imagery. To address these challenges, a color correction hybrid model which leverages the strengths of a Truncated U-Net architecture with skip connections and Shallow UNet has been employed in this work. The proposed model is named as MarineNeural Network(MarineNN) as it blends the efficiency and speed of Shallow U-Net with the robust structure and feature-preserving properties of a truncated U-Net, making it ideal for real-time underwater image enhancement with limited resources. It was trained and verified using the University of Minnesota EUVP dataset, which included underwater dark, Image-Net and scenes. Experimental results demonstrate MarineNN’s superior performance in image quality and detailed preservation by comparing to other contemporary methods like Pix2Pix, FUnIE-GAN UGAN models. The output values of MS-SSIM and Loss function L1 value attained from MarineNN model showed the significant improvement over Pix2Pix and UGAN models. The higher Mean Opinion Score(MOS) value has again assured the excellent quantitative performance of MarineNN when compared with UGAN. The qualitative metrics, no-reference metrics and the comparison with Image processing techniques SIFT and Canny Edge Detection mechanisms proves that the output images for our MarineNN model and shows a rapid progresssive enhancement when compared to the original images. The MarineNN model removes noise faster during training with quicker inference, providing a practical, versatile solution for enhancing the underwater images and has a potential to improve visual perception in underwater.
Kundeti et al. (Thu,) studied this question.