CBBGAN: a color-blur balanced generation adversarial network for underwater image enhancement

Underwater optical image processing has garnered significant attention in various underwater applications. However, the presence of particles and the attenuation characteristics of optics in underwater environments lead to color distortion, low contrast, and blurring in optical images. Image enhancement techniques play a crucial role in improving the effectiveness of underwater image processing. In this study, we propose a color-blur balanced generation adversarial network (CBBGAN) for enhancing underwater optical images. CBBGAN aims to address color distortion and blurring issues. To address color distortion, we introduce the fusion-based Color Compensation module to mitigate color variations in the images. Then, the Multi-stage Residual based Generator is proposed to enhance the generative capacity of the Generative Adversarial Network (GAN), enabling the extraction of multi-dimensional features from the images. Furthermore, we propose a Structural Similarity based Joint Loss Function during the training phase, which is used to guide network training. We conducted qualitative analysis on different algorithms on three public datasets, which intuitively demonstrates that the proposed method effectively removes color deviation and blurring issues in images. In the quantitative experiment of EUVP, compared with the most advanced algorithms in PSNR and SSIM, CBBGAN has improved by 2 and 0.03, respectively. In addition, various indicators on the UFO and UIEB datasets also demonstrate the excellent performance of the CBBGAN algorithm.