基于视觉认知的大景深雾霾退化图像去雾技术研究

Visibility of image degraded by haze descends serverely, so designing methods to improve the visibility of haze degraded image, especailly image with deep depth, becomes a problem need to be resolved imperatively. Presently, models that single image dehaze bases on could not describe the degradation factors and process of deep depth image, and priors that model solution bases on have some constraints. Therefore, basic knowledge about deep depth haze image and haze removal technology for it are researched in this project, haze degradation model for deep depth image is proposed through combining complex illumination, turbidity medium and onflow medium degradation, and it could describe the process of deep depth image degradation much more exactly. Robust priors that model solution needs are researched based on human visual cognition mechanism and numerical evaluation. To evaluate some variants in the deep depth image's haze degradation model, some human visual mechanisms are introduced in and modeled, and optimized solutions are designed to ensure evaluating these variants rapidly and exactly. Method to objectively evaluate the haze removal algorithm's performance is proposed according to human visual cognition, which combines statistical model, visual mechanism model and machine learning theory, and the evaluation method could get almost the same score as subjective evaluation. Finally, a novel method to do deep depth image haze removal will come up, and the prospective production will have great value in theory and application in many fields, such as intelligent traffic, surveillance, intelligent weapon and so on.

可见光成像系统受雾霾天气影响，成像质量严重下降，如何改善雾霾退化图像的质量，尤其是大景深图像的质量成为当前亟需解决的问题。目前，单幅图像去雾所依赖的模型无法准确描述大景深图像的退化要素和过程，且模型求解所依赖的先验知识具有局限性。因此，本项目开展大景深图像雾霾退化的基础理论及去雾技术研究，融合复杂大气光、混浊介质和湍流介质退化提出大景深图像雾霾退化模型，更为准确的描述大景深图像的退化过程；分别通过人眼视觉认知机制和数值仿真方法研究模型求解所需的具有鲁棒性的先验知识；对多种人眼视觉机制建模，并应用于模型的变量求解，设计最优化求解方法，实现对变量的快速、准确估计；融合统计模型、视觉信息模型和机器学习理论等，提出基于特征认知的图像去雾质量评价方法，充分拟合人眼主观评价结果。最终将形成一套大景深图像雾霾去除的新方法，预期成果将在智能交通、视频监控、智能武器等领域，具有重大理论意义和实用价值。

可见光成像系统受雾霾天气影响，成像质量严重下降，如何改善雾霾退化图像的质量，尤其是大景深图像的质量成为当前亟需解决的问题。目前，单幅图像去雾所依赖的模型无法准确描述大景深图像的退化要素和过程，且模型求解所依赖的先验知识具有局限性。因此，本项目开展了大景深图像雾霾退化的基础理论及去雾技术研究，融合复杂大气光、混浊介质和湍流介质退化提出大景深图像雾霾退化模型，更为准确的描述大景深图像的退化过程；分别通过人眼视觉认知机制和数值仿真方法研究了模型求解所需的具有鲁棒性的先验知识；对多种人眼视觉机制建模，并应用于模型的变量求解，设计最优化求解方法，实现了对变量的快速、准确估计；融合统计模型、视觉信息模型和机器学习理论等，提出基于特征认知的图像去雾质量评价方法，基本达到了人眼主观评价的结果。研究的成果将在智能交通、视频监控、智能武器等领域，具有重大理论意义和实用价值。在本项目的资助下课题组在国际期刊、会议和国内核心期刊上发表51篇，其中SCI检索8篇，EI检索25篇，较好的完成了学术论文发表工作。项目共培养博士生3名，硕士生7名。