基于深度估计和梯度下降的水下图像恢复与增强

Underwater image restoration and enhancement using depth estimation and gradient descent

由于水下介质散射和吸收等固有特性,水下图像面临图像模糊、低对比度和颜色失真等多重降质问题,严重影响视觉感知性能。针对上述问题,提出基于深度估计和梯度下降策略的水下图像恢复与增强框架(UIRENet)。首先,借助卷积、非线性激活函数模块,构建深度感知网络,实现对不同退化区域的场景深度感知,克服场景深度依赖的退化;其次,提出梯度优化策略,优化卷积网络参数,提升深度网络增强性能;最后,结合感知损失、边缘损失和水下色彩恒常损失,形成水下图像增强网络损失函数。通过在UIEB⁃90、UIEB⁃M和EUVP数据集上开展综合测试实验,验证了UIRENet框架在降低水下图像模糊度、提升视觉效果方面均显著优于目前典型水下图像增强方法,特别在客观评价指标UIQM上,相比CLAHE、ICM、GC、IBLA、DCP、ULAP、FUnIE⁃GAN、UGAN和Uformer等方法分别提高0.3700、0.6446、0.5919、1.3081、1.3032、1.1672、0.0593、0.1329和0.0934。

Due to inherent scattering and absorption,underwater images inevitably suffer from multiple degradations arising from blurring,low contrast and color distortion,thereby seriously deteriorating visual perception.In this paper,a deep learningbased underwater image restoration and enhancement framework(UIRENet)was proposed by virtue of depth esti⁃mation and gradient descent strategy.With the aid of convolutional and nonlinear activation function modules,a deep perception network was constructed to achieve scene depth perception maps for different degradation regions,thereby overcoming the dependence of scene⁃depth degradation.A gradient optimization strategy was further proposed to optimize the parameters of convolutional networks and improve the performance of deep network enhancement.Combined with perceptual,edge and underwater color constancy losses,a comprehensive loss function for underwater image enhancement networks was rationally formed.Comprehensive experiments on the UIEB⁃90,UIEB⁃M and EUVP datasets show that the UIRENet framework significantly outperforms typical underwater image enhancement methods in terms of reducing underwater image blurriness and improving visual effects.In particular,comparing to CLAHE,ICM,GC,IBLA,DCP,ULAP,FUnIE⁃GAN,UGAN and Uformer,the objective evaluation metric UIQM can be promoted by 0.3700,0.6446,0.5919,1.3081,1.3032,1.1672,0.0593,0.1329 and 0.0934,respectively.