HCMPE-Net: An unsupervised network for underwater image restoration with multi-parameter estimation based on homology constraint

Abstract

Underwater images suffer from severe degradation due to light absorption and scattering. Although deep learning-based methods have demonstrated impressive performance in underwater image restoration, their dependence on paired datasets limits their applicability. Therefore, an unsupervised network for underwater image restoration with multi-parameter estimation based on homology constraint is proposed to address this challenge. This method eliminates the dependency on real labels, thereby broadening its applicability across various scenarios. Meanwhile, compared with traditional unsupervised networks, this paper designed the optimized parameter estimation modules that not only improve restoration accuracy but also ensure real-time processing performance. Specifically, a contextual attention-based residual network is employed to estimate scene radiance. This module integrates global and local features to achieve accurate estimation using a contextual attention mechanism. Additionally, an adaptive cross-channel interaction network and a quadtree-based Gaussian blur module are established for precise estimation of the transmission map and background light. The adaptive cross-channel interaction network dynamically adjusts the interaction between RGB channels to enhance detail fidelity and local transmission estimation accuracy. The background light estimation module integrates quad-tree and Gaussian blur strategies to effectively mitigate background light bias in complex lighting environments. During the training phase, a color loss function based on three color spaces is defined. This function imposes joint constraints on the image in different color spaces, accurately capturing color deviations and optimizing the overall color restoration performance. Extensive experimental results demonstrate that our method achieves superior restoration accuracy and real-time performance across multiple real-world underwater image datasets, significantly outperforming existing state-of-the-art methods.