A Cascade-Based Composite Neural Network for Underwater Image Enhancement

Abstract

Underwater images are often affected by various degradation phenomena, such as low contrast, blurred details, color distortion, poor clarity, non-uniform illumination, and limited viewing distance. To address these issues, this paper proposes a cascaded composite neural network for underwater image enhancement, which incorporates a deep learning-based routing mechanism. Three individual neural networks, namely UWCNN (UW), Deep Wave-Net (DW), and PUIE-Net (PU), are employed as core components, and a method library is constructed using pairwise superimposed serial composite enhancement models. This framework is designed to enhance degraded underwater images and investigate the performance of the composite models. Experimental evaluations are conducted using metrics including PSNR, SSIM, UIQM, and UCIQE. The results indicate that the representative composite neural network model DW-PU achieves favorable performance with indicators of 20.495 (PSNR), 0.874 (SSIM), 3.270 (UIQM), and 0.897 (UCIQE), outperforming current mainstream underwater image enhancement models in certain aspects. Comparative analysis of images enhanced by multiple methods reveals that, in most underwater scenarios, the DW-PU model can effectively correct the color of degraded underwater images, making them more suitable for observing underwater conditions.