基于区域卷积神经网络稀疏正则模型的图像除雾理论与方法研究

Outdoor imaging devices usually produce degraded images with low contrast, bad clearness and color distortion when they are disturbed by fog and haze. It becomes difficulty for the outdoor visual systems to detect and identify objects efficiently. Algorithms of removing fog from images provide a useful solution. At present, estimating the atmosphere light and the atmospheric transmissivity accurately at the location where the field depths change abruptly, and evaluating the quality of defogged images are challenging the fog removal algorithms. Firstly, the theory of RCNN will be investigated. An accurate model of ambient light will be built by combining atmospheric scattering ambient light with the deep RCNN after its physical and image features are analyzed. Secondly, sparse regularization constraints of ambient light orienting to abrupt field depth will be studied. We will study the method of estimating atmospheric transmissivity accurately by applying RCNN. Then a fog removal algorithm will be proposed which will be efficient and have high quality. Finally, high order structure features and joint visual features are investigated thoroughly. Constraint structure support vector machine and RCNN will be combined to construct an image evaluation model with no reference and evaluate the defogged image quality correctly. It will provide a good support to develop adaptive algorithms. The proposed algorithm and image evaluation model will lay a foundation for outdoor visual devices to detect and recognize outdoor objects accurately and efficiently.

在雾霾条件下，室外成像设备获取的图像对比度低，清晰度差，颜色有失真，室外视觉系统目标检测和识别效果差。图像除雾算法成为解决上述问题的一种有效途径。目前，图像除雾算法主要面临着精确估计大气环境光和景深突变处的大气透射率，以及正确评估除雾后的图像质量等瓶颈。首先，研究区域卷积神经网络理论，通过分析室外大气散射环境光的物理和图像特性，结合深层区域卷积神经网络，构建精确大气环境光的模型。其次，研究面向图像景深突变处大气透射率的稀疏正则化约束条件，探究运用深层区域卷积神经网实现大气透射率准确估计的方法。在此基础上，提出了基于区域卷积神经网络稀疏正则模型的除雾算法。最后，探究图像高阶结构特性和联合视觉特征，结合约束结构支持向量机和区域卷积神经网络，构建无参图像质量评估模型，为提高除雾算法自适应性提供依据。该项目提出的除雾算法和图像质量评估模型，将为室外视觉系统准确快速检测和识别目标奠定基础。

在雾霾条件下，室外成像设备获取的图像对比度低，清晰度差，颜色有失真，室外视觉系统目标检测和识别效果差，图像除雾成为解决上述问题的一种有效途径。首先，通过分析室外大气散射环境光的物理和图像特性，结合深层区域卷积神经网络，提出了精确估计大气环境光值的方法。依据面向图像景深突变处大气透射率的稀疏正则化约束条件，探究基于变分正则化模型的大气透射率准确估计方法。然后，研究基于区域卷积神经网络的除雾方法。针对局部区域有雾霾干扰的图像，为了尽可能保留雾霾区域以外的景物信息，通过分析区域卷积神经网络的组成和结构特点，建立了基于Faster-RCNN的单幅图像雾区识别网络，采用多种深度卷积网络为主干网，比较和分析了Faster-RCNN网络的雾区识别效果。同时，探讨了稀疏正则化理论，研究了细节信息注意力机制和视觉感知机制，结合大气环境光和透射率图估计方法，提出了三种基于区域卷积神经网络的除雾方法，即基于细节信息注意的去雾方法(DMDA)、基于深度稠密残差感知的去雾方法(DDRA)、基于空间和通道感知去雾方法(IDSCAN)，构建了相应的网络模型。其次，在研究去雾后图像的视觉显著性和感知锐度特征的基础上，将不同尺度的自然图像评估值和自然图像的颜色偏离程度相结合，设计了一种无参图像的多尺度去雾图像质量评估算法，有效提高了评估图像质量的准确性。最后，开展图像实时采集和处理实验。结果表明，提出的基于区域卷积神经网络的图像去雾方法，能有效提高雾霾图像的景物清晰度和对比度，优于暗通道先验、AOD-net、MAXIM、FAMED-Net等算法。本项目提出的图像质量评估算法采用MNIQC评估值，提高了图像评估的准确性。本项目成果将为室外视觉系统准确快速检测和识别目标奠定基础。