Underwater optical images are widely used in marine exploration. Due to the weak light problem caused by water depth, underwater images generally have the characteristics of background noise, dark brightness, strong blue‒green background color, and blurred images. These characteristics bring great inconvenience to marine exploration tasks. In this way, the study of underwater image enhancement has important application value. Most of the existing underwater image enhancement methods mainly solve the problem of the overall denoising and brightness enhancement of the underwater image while ignoring the partial denoising of the image. To solve these problems, this paper proposes an improved generation adversarial network (GAN) to achieve clear processing of underwater images. The main improvements include three aspects. First, a portion denoising module is added to the generator to weaken the image noise produced by the generator in a detailed manner. Second, the acceleration module is introduced into the discriminator to accelerate the training process of the GAN network. Third, the sum of squares of confrontation loss, contrast loss and color loss is used as a loss function to make the training of the GAN network stable. Extensive experimental results show that the proposed model is superior to the comparison method in both quantitative and qualitative experiments, and the visualization results of the results are natural.

水下光学图像广泛应用于海洋勘探。由于水深导致的弱光问题，水下图像普遍存在背景噪声大、亮度暗、蓝绿色底色浓、图像模糊等特点。这些特点给海洋勘探任务带来了极大的不便。这样，水下图像增强的研究具有重要的应用价值。现有的水下图像增强方法大多主要解决水下图像的整体去噪和亮度增强问题，而忽略了图像的局部去噪。针对这些问题，本文提出了一种改进的生成对抗网络（GAN）来实现水下图像的清晰处理。主要改进包括三个方面。第一的，在生成器中加入了部分去噪模块，对生成器产生的图像噪声进行了细致的削弱。其次，在判别器中引入加速模块，加速GAN网络的训练过程。第三，对抗损失、对比度损失和颜色损失的平方和作为损失函数，使GAN网络的训练稳定。大量的实验结果表明，所提出的模型在定量和定性实验中均优于对比方法，结果的可视化效果自然。使用对抗损失、对比度损失和颜色损失的平方和作为损失函数，使GAN网络的训练稳定。大量的实验结果表明，所提出的模型在定量和定性实验中均优于对比方法，结果的可视化效果自然。使用对抗损失、对比度损失和颜色损失的平方和作为损失函数，使GAN网络的训练稳定。大量的实验结果表明，所提出的模型在定量和定性实验中均优于对比方法，结果的可视化效果自然。