基于暗原色先验的快速单幅图像去雾算法

针对基于暗原色先验理论的单幅图像去雾算法中，由于某些场景下的雾天图像存在大面积明亮区域（如天空、水面或者偏白色物体等）不满足暗原色先验假设，从而导致去雾处理效果不好的问题。基于暗原色先验理论，提出了一种改进的单幅图像去雾算法。首先利用统计截断的方法估计出大气光值；然后对暗通道图进行中值滤波得到粗略估计的透射率图，并对明亮区域的透射率图进行自适应校正处理；最后将这些参数带入大气散射成像模型完成去雾处理。实验结果显示，相较于原算法而言，所提算法可以准确地选取出天空区域的像素点对大气光进行估计，有效降低明亮区域的色彩失真。通过不同算法对不同室外场景下采集的雾天图像的去雾效果的对比可知，所提算法在对明亮区域的处理上更加合理，可以较好地处理一些带有光源的图像，恢复出的图像具有很好的细节保持，视觉效果显著提高。所提算法对含有大面积明亮区域的雾天图像具有很好的增强处理效果，可以为图像分割、语义检索、智能分析等图像处理工作提供有效的预处理手段，对于交通监管、视频监控、行车视频记录、视觉导航等研究领域具有重要的意义。

Fast single image dehazing algorithm based on dark channel prior

The existing single image dehazing algorithms based on dark channel prior theory cannot achieve satisfactory performance in the cases that the image has large high light regions (i. e. sky area, water surface or white objects) which does not meet the dark channel prior hypothesis. To solve this problem, this paper proposed a novel single image dehazing algorithm based on the dark channel prior. Firstly, it estimated the atmospheric light value by a new statistical truncation process, then a median filter is applied to the dark channel prior to estimate the approximative value of the transmission map. Finally, it restored the haze-removed image by the atmospheric scattering model. Experimental result shows that compare to the original method, the proposed algorithm can pick out the pixels in the sky regions correctly then estimate the atmospheric light value accurately, which can alleviate the colour distortion effectively. Compare to other dehazing methods, the proposed algorithm has better performance in haze removal in foggy images with large high light regions, even has spot lighting sources, and the details can be reserved and the visual quality is enhanced significantly. The proposed algorithm can be applied to restore the foggy image that has large high light regions or spot lighting source. This method can serve as a useful pre-processing to image segmentation, semantic retrieval, intelligent analysis, etc. , which has great significance to the research areas of traffic supervision, video surveillance, automobile data recorder and visual guidance and so forth.