基于暗原色及入射光假设的单幅图像去雾

目的 雾是一种常见的天气状况,针对雾能使图像中的景物对比度降低、表面颜色退化的问题,提出一种基于入射光假设的单幅图像去雾方法。方法 首先利用全局暗原色进行初步去雾,从而使图像透射率处于[0,1]范围内;然后利用雾天光照均匀的特点以及Retinex的照度估计原理进行透射图的估计;最后利用透射图以及初步去雾图像得到复原图像。结果 与He算法、Fattal算法的对比实验结果显示,该算法获得的复原图像细节清晰,颜色自然。与引导滤波优化后的He去雾算法相比,本文算法速度提高了93%。结论 大量对比实验结果表明,本文算法能够显著恢复雾天降质图像,对于薄雾和浓雾同样有效,具有广泛的适用性,且算法原理简单。此外,本文算法也同样适用于灰度图。

Single-image dehazing based on dark channel and incident light assumption

Objective Fog is a common condition that reduces the contrast of an image, bleaches the surface color, and considerably reduces the value of outdoor images. To address this problem, we propose a defogging method for a single degraded image on the basis of dark channel and incident light assumption. Method We scan the image with a window to determine the window with the maximum mean brightness. We use the obtained average value as the atmosphere light. The dark channel prior assumption raised by He is not suitable to images that contain a large scene, so we weaken the assumption. We assume that a channel of a pixel whose value is zero exists. Basing on this assumption, we identify the darkest pixel value in the entire image and use the darkest pixel value as the global dark channel. We use the ratio of the grayscale of the point to the atmospheric light as the basis transmission of the image. Using this basis transmission, we conduct the initial dehazing. The transmission rate of the image will then be stretched to the [0,1] range. Images taken under a foggy weather almost have no shadow. We therefore assume that the incident light during a foggy day is uniform. We estimate the transmission by using a multi-scale approach combined with retinex theory that uses Gaussian convolution to estimate illumination. According to the haze imaging model, we can recover a high-quality, haze-free image by using this transmission map and the initial dehazing image. Result By weakening the dark channel prior assumption of He, we considerably improve its accuracy and perform the initial dehazing on the basis of the weakened assumption. Unlike in other methods, the transmission map of our algorithm does not exhibit an apparent object contour. The fuzzy transmission map is obviously reasonable according to the scattering characteristic of fog. Experimental results indicate that the algorithm can provide an accurate estimation of the transmission, and the restored images show natural colors and clear details. The algorithm also exhibits low computational complexity and almost does not need to set any parameters. Our algorithm shows good results and substantially increases the computing speed compared with haze removal theory on the basis of the dark channel prior of He. The proposed algorithm is not limited by a poor capability in processing images with thick fog, which is a key concern in Fattal's method. Conclusion This study proposes a new method to assume transmission on the basis of incident light assumption and retinex illumination estimation principle. A large number of comparative experiments show that the algorithm can significantly restore the quality of an image degraded by fog. Our method is effective for images taken under thin and thick fog, demonstrates wide applicability, and involves a simple principle. The proposed method is also applicable to grayscale.