基于暗原色先验信息的单幅图像去雾算法研究

发布时间：2018-05-08 05:01

本文选题：单幅图像去雾 + 暗原色原理　；参考：《哈尔滨工业大学》2017年硕士论文

【摘要】：人类获取外界的信息量超过70%来自于视觉,随着工业水平的快速发展以及生活水平的提升,雾霾天气出现的频率加快、程度加深。雾霾天气使得户外成像系统获得的图像视频降质。为了从雾霾天气获得的户外图像中提取出准确、足够的信息,需要对有雾图像进行去雾处理,使得能够应用到交通道路监控、遥感监测等方面。本文主要工作基于何恺明提出的暗原色原理进行单幅图像去雾。首先,深入研究何恺明提出的去雾算法:第一步,估计图像的粗略透射率;第二步,优化粗略透射率,本文利用何恺明提出的软抠图和引导滤波两种方法实现;第三步,利用大气散射模型进行参数反演获得去雾图像。然后,分析了何恺明的方法的缺点并提出改进方案:(1)利用四分加权法进行大气光估计,避免当图像中存在大片浅色非天空区域造成的误差;(2)利用邻域相似暗原色先验估计透射率,能够解决软抠图造成的时间复杂度高问题和消除引导滤波过平滑导致的光晕现象。然后,利用大气散射模型和本文算法估计的大气光值和透射率得到去雾图像。最后,设计GUI界面用于显示输入的有雾图像和三种算法的去雾后图像,以及对三种算法的客观评价指标,包括运行时间、峰值信噪比、平均结构相似性、新增可见边之比、可见边的规范化梯度均值和饱和黑色或白色像素点的百分比。本文利用以上三种算法分别对自然图像、含人物的图像、含建筑物的图像以及车牌图像四类进行去雾并对每一幅图像进行评价,改进后的算法具有如下几个优点:(1)时间复杂度大大降低(一幅600*450的图像,本文算法相比于使用软抠图的算法和使用引导滤波的算法分别降低了89.5%和35%);(2)能够消除由引导滤波过平滑造成的光晕现象;(3)综合主观评价方法和客观评价方法评估出本文的改进算法的去雾能力,其能在保持何恺明算法效果的基础上获得更多的细节信息。同时,本文算法存在需要改进的地方:(1)去雾后的图像可能出现颜色过饱和的现象,(2)透射率估计选用的窗口大小也会影响去雾结果。
[Abstract]:With the rapid development of industry and the improvement of living standard, the frequency and degree of haze weather appear faster and deeper with the rapid development of industry and the improvement of living standard. Haze weather allows outdoor imaging systems to obtain image video degradation. In order to extract accurate and sufficient information from outdoor images obtained from haze weather it is necessary to defog the foggy images so that they can be applied to traffic road monitoring remote sensing monitoring and so on. The main work of this paper is to defog a single image based on the principle of dark primary color proposed by Ho Kaiming. Firstly, the de-fogging algorithm proposed by Ho Kaiming is studied in depth: the first step is to estimate the rough transmittance of the image; the second step is to optimize the rough transmittance. In this paper, the soft matting method and the guided filtering method proposed by Ho Kai-ming are used to realize the algorithm. The model of atmospheric scattering is used for parameter inversion to obtain the defog image. Then, the shortcomings of Ho Kai-ming 's method are analyzed and an improved scheme is put forward: 1) the atmospheric light estimation is carried out by using the quadrilateral weighting method. To avoid the error caused by large light-colored non-sky region in the image we can solve the problem of high time complexity caused by soft matting and eliminate the halo caused by over-smoothing of guided filter by using neighborhood similar dark primary color priori estimation of transmittance. Then, using the atmospheric scattering model and the atmospheric light value and transmittance estimated by the present algorithm, the defogging image is obtained. Finally, the GUI interface is designed to display the inputted foggy image and the post-fogging image of the three algorithms, and the objective evaluation indexes of the three algorithms, including running time, peak signal-to-noise ratio (PSNR), average structural similarity, and the ratio of visible edges added. The normalized gradient mean of the visible edge and the percentage of saturated black or white pixels. In this paper, the above three algorithms are used to defog and evaluate the natural image, the image of the person, the image of the building and the image of license plate, respectively. The improved algorithm has the following advantages: 1) time complexity is greatly reduced (an image of 600,450, Compared with the soft matting algorithm and the guided filter algorithm, the proposed algorithm can reduce 89. 5% and 35%, respectively.) it can eliminate the halo phenomenon caused by the guiding filter smoothing.) the comprehensive subjective evaluation method and the objective evaluation method can be evaluated. In this paper, the improved algorithm of the ability to remove fog, It can obtain more details on the basis of maintaining the effect of Ho Kaiming algorithm. At the same time, there is a need for improvement in this algorithm. (1) the phenomenon of color supersaturation may occur in the image after fog removal. (2) the window size selected for the transmission estimation will also affect the de-fogging results.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】