基于物理特性与特征学习的图像去霾算法研究

霾天气下的图像常因气溶胶粒子的散射作用而降质，并影响视觉效果和后续处理。针对现有方法大多将不同成因的雾霾图像等同处理的问题，本课题拟在分析霾和其他气象条件散射行为差异的基础上，深入研究：1）面向不同波长大气光的散射差异，设计基于霾环境下成像模型的图像去霾算法；2）根据霾图像远/近景差异，引入信息熵来描述图像不同位置的能量变化，设计基于图像融合的去霾算法；3）针对传输图估计误差导致复原图像的失真问题，设计基于估计传输图的多项式函数，并通过最小化其与标准传输图的差异来实现传输图优化；4）构建全卷积网络，利用“霾图像—标准无霾图像”图像对数据集学习霾对图像降质的影响，实现有霾图像的复原。重点突破基于深度神经网络进行霾图像特征学习的关键技术，实现对任意给定场景下单幅图像的去霾处理。研究成果能够进一步提升室外视觉设备对恶劣天气的适应能力，对面向公共安全的安防、水文监控及交通监测等领域有重要应用价值。

Outdoors images are always degraded by suspending particles in haze condition, which affect the visual quality and post-processing. In order to solve the problem of unsatisfactory restoration caused by dehazing in an analogous way with defogging, basing on the analysis of different scattering behavior of haze from other conditions, we plan to study deeply on: 1) designing a dehazing algorithm based on the imaging model in haze condition and different scattering behavior along with the variation of wavelength. 2) employing the theory of information entropy to describe the variation of energy in far range and close range regions in hazy image, and design a dehazing algorithm based on image fusion. 3) designing a polynomial function of the estimated transmission map to solve the distortion like halo effect, and optimize it by minimizing the difference between estimated and standard transmission map. 4) building a fully convolutional network and learning the influence of haze impacted on images with "hazy-haze free" image pairs, then restoring images with this model. We will make breakthroughs in feature learning of hazy images with the deep neural networks, and achieve single image dehazing in any given scenes. Our research can improve the adaptability of outdoor computer vision devices to bad weathers and is of significant meaning to public safety oriented security, Hydrological and traffic monitoring.