CCCSIN: Color Compensation and Cross-Scale Semantic Information-Guided Network for Underwater Image Enhancement

Abstract

In this article, to simultaneously address color distortion and content reconstruction pertaining to underwater image enhancement, a color compensation and cross-scale semantic information-guided network (CCCSIN) is devised. The multidense connection residual block (MCRB) with spatial and pixel attentions in a cascading pattern is developed, such that color distortion of underwater image can be effectively alleviated. By virtue of global and local semantic information derived from the semantic segmentation model, the cross-scale semantic feature enhancement module (CSFEM) is created, such that detailed image content can be accurately reconstructed. Employing the supervision of absolute error and mean square error loss functions, the entire CCCSIN scheme is constructed by organically combining with MCRB and CSFEM. Experiments demonstrate the effectiveness and superiority of the proposed CCCSIN scheme, peak signal-to-noise ratios are 3.2981, 2.2953, 2.1711, 5.4679, 7.3372, 9.5632, 4.1839, and 2.4253 higher than those of contrast limited adaptive histogram equalization, integrated colour model (ICM), unsupervised correction method (UCM), dark channel prior (DCP), image blurring and light absorption, underwater DCP, underwater light attenuation prior, and fast underwater image enhancement with generative adversarial networks (FunIE-GAN) frameworks, respectively, demonstrating compound advantages in detail recovery, texture preservation, color correction, contrast improvement, and clarity enhancement.