基于改进CycleGAN的水下图像颜色校正与增强

针对水下观测图像的颜色失真和散射模糊问题, 提出一种基于改进循环一致性生成对抗网络(Cycle-consistent generative adversarial networks, CycleGAN)的水下图像颜色校正与增强算法. 为了利用CycleGAN学习水下降质图像到空气中图像的映射关系, 对传统CycleGAN的损失函数进行了改进, 提出了基于图像强边缘结构相似度(Strong edge and structure similarity, SESS)损失函数的SESS-CycleGAN, SESS-CycleGAN可以在保留原水下图像的边缘结构信息的前提下实现水下降质图像的颜色校正和对比度增强. 为了确保增强后图像和真实脱水图像颜色的一致性, 建立了SESS-CycleGAN和正向生成网络G相结合的网络结构; 并提出了两阶段学习策略, 即先利用非成对训练集以弱监督方式进行SESS-CycleGAN学习, 然后再利用少量成对训练集以强监督方式进行正向生成网络G的监督式学习. 实验结果表明: 本文算法在校正水下图像颜色失真的同时还增强了图像对比度, 且较好地实现了增强后图像和真实脱水图像视觉颜色的一致性.

Underwater Image Color Correction and Enhancement Based on Improved Cycle-consistent Generative Adversarial Networks

Aiming at the problem of color distortion and scattering blurry effect of underwater observation images, an color correction and enhancement algorithm of underwater images is proposed based on improved cycle-consistent generative adversarial networks (CycleGAN). Firstly, in order to learn the mapping relationship between underwater degraded images and in-air images using CycleGAN, the CycleGAN is improved by introducing a new loss function based on strong edge and structure similarity (SESS). The new SESS-CycleGAN can realize color correction and contrast enhancement of underwater degraded images while retaining their edge and structure information. Secondly, for ensuring the color consistency between the enhanced images and the their corresponding in-air images, a dual-level network structure combining SESS-CycleGAN with a forward generation network G is constructed, and a two-stage learning strategy is proposed. In the two-stage learning process, the SESS-CycleGAN is first trained in a weakly supervised manner with a unpaired training set composed of underwater images and in-air images, and then the forward generation network G of SESS-CycleGAN is trained again in a strong supervised manner with a small number of paired training set. Experimental results show that the proposed algorithm can enhance image contrast while correcting color distortion of underwater images, and has better visual color consistency with real images in air.