基于物理模型的深度学习水下图像恢复方法  

Physical model based deep learning method for underwater image restoration

作　　者：李苇杭 杨鸿波[1] 张洋 LI Weihang;YANG Hongbo;ZHANG Yang(School of Automation,Beijing Information Science&Technology University,Beijing 100192,China)

机构地区：[1]北京信息科技大学自动化学院,北京100192

出　　处：《北京信息科技大学学报（自然科学版）》2024年第3期53-59,共7页Journal of Beijing Information Science and Technology University

基　　金：国家自然科学基金项目(62001035)。

摘　　要：针对复杂水下环境中图像颜色失真、细节模糊和对比度降质等问题,提出了一种基于物理模型的深度学习水下图像恢复方法,使用水下光学成像的物理模型约束生成对抗网络(generative adversarial network, GAN),为深度学习方法增加理论支撑,同时降低深度学习恢复效果对训练数据集的依赖。首先,估计物理模型的关键参数,基于视觉显著性原理简化传统透射图计算方法,并利用GAN的生成器获取的雾度图计算环境光;其次,通过物理模型计算水下恢复图像,利用GAN的鉴别器区分恢复图像和参考图像;通过不断学习优化网络参数,最终实现水下图像恢复。测试集上的实验结果表明,图像恢复效果在主客观多种评价指标上均得到有效提升。A deep learning method based on physical model for underwater image restoration was proposed to address color distortion,detail blurring,and contrast degradation of the images in complex underwater environments.The physical model of underwater optical imaging was used to constrain the generative adversarial network(GAN),adding theoretical support to the deep learning method while lowering reliance of the deep learning restoration effect on the training dataset.Initially,crucial parameters of the physical model were estimated,simplifying traditional methods for computing transmission maps based on visual saliency principles.Then,the haze map obtained by the generator of GAN was used to calculate ambient light.Subsequently,the underwater restoration images were computed through the physical model,and the discriminator of GAN was used to distinguish between the restoration and reference images.Through iterative optimization of network parameters,underwater image restoration was achieved.Experimental results on the test set show that the image restoration effect has been effectively improved in various subjective and objective evaluation indicators.

关 键 词：水下图像恢复 深度学习 水下成像物理模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]