水下图像盲复原非局部变分方法及质量评价

Blind underwater image restoration is a kind of complex ill-posed inverse problem due to the characteristics of hazy, low contrast, blurring, noise and so on. On account of the influence of multi-factor coupling, the traditional single algorithm can only solve one specific problem, and the multi-step method is often regarded as “attend to one thing and lose another”, the simultaneous multi-factor variational model and alternating optimization are expected to be effective solutions. Our research aims to design the data item and smooth item of the unified variational model based on the underwater optical image formation model, and the point spread function and (PSF) background light are empirically estimated, and its convolution kernel function and scene depth function are solved via smooth priors. The non-local differential operator with higher order term is used to maintain the smoothness, edge, texture and other features of restoration image. Furthermore, in the process of global convex optimization, a fast Nesterov method with the iterative step size and the alternating direction multiplier method (ADMM) are introduced to accelerate the whole progress. Finally, an underwater benchmark will be established and the quality of restored image is further evaluated to guide and modify the method according to the non-specific characteristics of an image such as edge and texture. The proposed models and algorithms will not only improve the accuracy and speed of underwater image restoration, but also expand their applications in underwater robot operation, oceanographic inspection and other fields such as image understanding and target recognition.

雾化、低对比度、模糊、噪声等致使水下图像盲复原为复杂的病态反问题，由于多因素耦合影响，传统单一算法仅能解决某一特定问题，而多步法往往顾此失彼，基于综合多因素的变分模型及交替优化算法有望将其归为一类问题进行求解。本项目拟基于水下光学成像模型设计变分模型的规则项和数据项，对点扩展函数和背景光的选取采用拟合估计，对卷积核函数、场景深度函数采用光滑先验，对拟恢复图像采用非局部微分算子和高阶项描述以保持图像光滑度、边缘、纹理等特征。在全局凸优化求解过程中，将引入预估-校正迭代步长的Nesterov方法与交替方向乘子方法对所提出模型设计快速算法。最后通过建立水下图像模拟基准图库及设计面向非特定失真可评估边缘、纹理等特征信息的无参考方法来评价图像质量，修正复原效果。所提出模型和算法将提高水下图像复原的准确率和速度，并有助于探索其在水下机器人作业、海洋考察等领域更深层次如图像理解与目标识别等方面的应用。

雾化、低对比度、模糊、噪声等致使水下图像复原为复杂的病态反问题，由于多因素耦合影响，传统的单一算法仅能解决某一特定问题，而多步法往往顾此失彼。此外，在度量水下图像的视觉质量时，现有的图像质量评价方法与主观感知结果仍存在很大的差异。本项目围绕水下成像、水下视觉增强及图像质量评价等方面开展研究，以几何偏微分方程为指导，以变分法为核心建立了一套水下成像模型表达、变分模型设计和数值求解以及质量评价的水下视觉处理技术方案。在水下图像增强和复原方面，一是完善了水下成像模型；对现有的水下成像模型进行了扩展，提升了成像模型的鲁棒性。二是设计了融合成像物理模型的变分模型；在变分模型中设计了基于总变分项、曲率项、拉普拉斯项和非局部微分算子思想等非线性各向同性光滑项，解决了传统方法普遍存在通用性差、成像模型表达不完善、特征保持不理想的局限性等技术难题。在全局凸优化求解过程中，采用经典的快速交替方向乘子方法（ADMM）、软阈值公式、快速傅里叶变换等方法将原变分模型转化为一系列实施与计算简单的模型，提高了水下退化图像盲复原算法的计算效率。在图像质量评价方面，健全了现有的水下图像质量评价体系，提出了基于色度、饱和度、对比度以及结构、纹理等特征信息的无参考水下图像质量评价度量方法；同时，建立了全参考水下图像质量评价仿真图像数据集和面向真实水下图像增强的质量评价数据集，为未来水下图像质量客观评价方法以及水下图像增强算法的研究提供了参考与支持。本项目研究成果将为水下图像的盲复原及图像质量客观评价提供新方法、新技术，并有助于探索其在水下作业、渔业养殖、海洋资源探测与开发、海洋环境监测等实际工程领域中的应用。