单幅雾天图像的去雾算法研究

发布时间：2018-05-19 04:14

本文选题：图像去雾算法 + 加权引导滤波　；参考：《安徽大学》2017年硕士论文

【摘要】：雾霾天气下,由于大气中存在的悬浮颗粒对光线具有吸收、散射作用,使得户外捕捉到的图像出现对比度下降,颜色衰减等降质现象,导致物体特征难以辨别,图像的观赏性降低,影响图像的后续处理。因此,在目标跟踪、视频监控、遥感等计算机视觉方向的应用上,雾天图像的去雾技术研究有着重要的意义。随着相关计算机技术的日益发展与趋于成熟,雾天图像的去雾研究也成为了广大研究人员研究的热点。去雾算法主要分为图像增强和基于物理模型的图像复原两种。两种类型的算法各有优缺点,但近年来,基于物理模型的的去雾算法更为普遍。其中,基于中值滤波的快速去雾算法因其快速的去雾处理速度和良好的去雾效果而受到关注。但是,一方面,由中值滤波得到的大气面纱丢失了很多图像的边缘信息,进而不能真实反映景物深度的信息,导致去雾的不彻底;另一方面,天空等区域的透射率估计值较低,导致复原图像中出现色彩失真和噪声。为此,本文主要围绕大气面纱和透射率两方面来进行去雾算法的研究。首先,提出了一种基于加权引导滤波的单幅图像去雾算法。在引导滤波的基础上,借助Canny算子来自适应地设置输入图像的边缘区域和平坦区域的权值,使得输出图像保留了更多的边缘信息,平滑了更多的噪声纹理信息。根据这个特点,采用加权引导滤波代替中值滤波的方法来获取边缘信息更加丰富的大气面纱。同时,用暗原色先验代替白平衡操作来估计大气光,使得大气光值更为准确。实验结果表明,本算法有着较快的去雾处理速度,且边缘处的去雾更为彻底。其次,提出了一种基于大气面纱优化和透射率修正的单幅图像去雾算法。一方面,由于大气面纱的求解算法引入了噪声纹理信息,使得大气面纱并不能真实地反映景物深度的信息,无论是使用中值滤波,还是加权引导滤波,复原图像都存在去雾的不彻底。另一方面,天空等区域由于透射率的估计值偏低,导致复原图像存在噪声和色彩失真。为此,首先根据阈值分割得到天空等目标区域,并适当减小雾天图像的各颜色通道最小值图像相应区域的灰度值;然后通过两次不同引导图像的加权引导滤波操作,得到优化后的大气面纱,同时根据暗原色先验估计大气光;最后根据大气散射模型反演得到复原图像。实验结果表明,本算法在景物深度的不连续区域有着更为彻底的去雾效果,同时天空区域避免了噪声、色彩失真,与大气光相似的白色区域避免了颜色偏暗。
[Abstract]:In haze weather, because of the absorption and scattering of light by suspended particles in the atmosphere, the images captured outdoors are degraded in contrast and color attenuation, which leads to the difficulty of distinguishing the characteristics of objects. Image viewing is reduced, which affects the subsequent processing of the image. Therefore, in the applications of target tracking, video surveillance, remote sensing and other computer vision directions, the research on fog removal technology of fog images is of great significance. With the development and maturity of computer technology, the research of fog image defog has become a hot topic for researchers. The de-fogging algorithm is mainly divided into two kinds: image enhancement and image restoration based on physical model. The two kinds of algorithms have their own advantages and disadvantages, but in recent years, physical model-based de-fogging algorithms are more common. Among them, the fast de-fogging algorithm based on median filter is concerned because of its fast de-fogging speed and good de-fogging effect. However, on the one hand, the atmospheric veil obtained by the median filter has lost the edge information of many images, which can not truly reflect the depth of the scene, resulting in incomplete fog removal. On the other hand, the estimated transmittance of the sky and other regions is lower. Color distortion and noise appear in the restored image. Therefore, this paper mainly focuses on the atmospheric veil and transmittance to study the fog removal algorithm. Firstly, a single image de-fogging algorithm based on weighted guided filter is proposed. On the basis of guided filtering, the Canny operator is used to set the weights of the edge region and flat region of the input image adaptively, which makes the output image retain more edge information and smooth more noise texture information. According to this characteristic, the weighted guided filter is used instead of the median filter to obtain the atmospheric veil with richer edge information. At the same time, the dark primary color is used instead of the white balance operation to estimate atmospheric light, which makes the atmospheric light more accurate. The experimental results show that the proposed algorithm has a faster de-fogging speed and a more thorough de-fogging at the edge. Secondly, a single image de-fogging algorithm based on atmospheric veil optimization and transmittance correction is proposed. On the one hand, because the noise texture information is introduced into the solution algorithm of atmospheric veil, the atmospheric veil can not truly reflect the depth of the scene, whether using median filter or weighted guided filter. Restoration images are not completely foggy. On the other hand, because the estimated transmittance of the sky and other regions is low, there is noise and color distortion in the reconstructed image. Therefore, the sky and other target regions are segmented according to the threshold value, and the gray value of the corresponding region of the minimum value of each color channel of the fog image is appropriately reduced, and then the weighted guided filtering operation of two different guided images is carried out. The optimized atmospheric veil is obtained, and the atmospheric light is estimated prior to the dark primary color. Finally, the reconstructed image is obtained by inversion of the atmospheric scattering model. The experimental results show that the algorithm has a more thorough effect of removing fog in the discontinuous region of the depth of the scene, at the same time, the noise and color distortion are avoided in the sky region, and the white area similar to the atmospheric light avoids the dark color.
【学位授予单位】：安徽大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】