基于对抗生成网络的雾霾图像复原方法研究

Haze Removal technology has been widely used in civil and military fields. However, due to the complexity and variability of natural scenes, dehazing is still a challenging task. At present, the main problem of haze removal comes from three aspects, such as the poor universality of priors, the lack of robust feature description and the lower image quality. Therefore, we want to construct an improved Generative Adversarial Network model for solving this problem, which has recognized multi sample training and multi learning framework as a breakthrough point. On the one hand, considering self-characteristic and artificial-characteristic together, we deeply explore the establishment and representations of the complete feature set to search for the nature information. On the other hand, on the basis of learning from the successful experience, we can improve the stability and efficiency of the network by studying and analyzing the principle of the model and training skills. Finally, a dehazing algorithm with good applicability and stability is established. The expected results of our project will promote the development of the image restoration and the automation level of the intelligent transportation system, and has important scientific significance and potential applications.

图像去雾技术在民用和军事领域有着极其广泛的应用，但由于自然场景的复杂多变性，雾霾图像的质量提升依然是一个具有挑战性的任务。目前，图像去雾的主要问题来自于先验知识普适性较差、鲁棒特征描述不够准确、恢复图像质量有待提高这三个方面。因此，本申请以多域样本集训练和多学习框架相结合为突破点，构建适用于图像去雾的对抗生成网络模型，努力解决现有方法中存在的问题。一方面，深入探索完备特征集合的建立与表示，从图像自我特征出发，结合人工标定特征，充分发掘场景本质信息；另一方面，借鉴已有成功经验，通过对模型原理及训练技巧的分析学习，提升网络框架的稳定性及求解的高效性。最终建立一套适用性及稳定性都较好的图像去雾算法。本申请的开展和预期成果将在图像复原和智能交通系统的自动化水平等许多方面，具有重要的科学意义和潜在的应用前景。

雾霾图像复原，是图像处理领域反问题求解的代表内容之一，也是实现复杂气象环境下图像分析、检测、识别、跟踪的前提条件和首要步骤，对于去除杂质干扰，提升图像内在价值具有重要作用。特别是在雾霾天气日益蔓延、严重影响国计民生的今天，如何有效提升相关视觉设备的效能比，已成为破解穹顶之下国家现实需要的关键科学问题之一。本项目以剖析雾天退化模型为手段，以预估精准传递图为中心，在对景深相适应的典型几何结构特征分析下，提出了一系列围绕传递图/景深的构建思路和方法，以期初步搭建系统性与有效性并存的去雾方法体系。与之相对，为了更好的将研究成果与复杂气象环境下的遥感目标检测相匹配，项目组在深入剖析目标基本特性的基础上，以旋转特征为主要手段，可实现局部遮挡下的目标检测，从而更好的指引雾霾天气下的目标检测与识别。.在本项目的资助下，课题组相关成员在国际期刊、会议和国内核心期刊发表论文20篇，其中SCI检索8篇，EI检索12篇，较好的完成了学术论文发表工作，达到了申请结题的基本条件。项目共培养硕士生5名。