The Research, Validating And Optimization Of Digital Image Haze Removal

With the development of computer technology, computer vision has been the hotspot in many research fields, such as information science, computer science and so on, and it has been used in systems such as military reconnaissance system, traffic safety system, satellite remote sensing system and so on. In bad weather such as haze and fog, however, the digital images gathered by optical sencers are serious degraded because of the scattering of the atmosphere. The degraded images suffer from decrease of ornamental and local contrast and the color of it will become gray and the detail of the scene objects will be fuzzy. Haze will not only decrease the visual effects of the images, but also lead to the inaccuracy on characteristic extracting and adverse effects on subsequent processing. What's more, the systems rely on the optical sencers will fail due to the images degration. Therefore, digital image haze removal has become an important direction of research on computer vision and image processing.This article discuss about the topic mainly from the following aspects:Firstly, this article sums up two kinds of image haze removal algorithms, which are the algorithms based on image enhancement and based on the physics model. After that, all the existing haze removal algorithms are fully reviewed.Secondly, the Quick Shift clusting image segmentation guided dark channel prior single image haze removal algorithm is introduced.The Quick Shift image segmentation approach is introduced to decompose the input image into some gray level consistent regions. Compared with the traditional fixed image partition schemes, better estimation of the atmospheric light can be obtain as well as to avoid the problem of halo artifacts. With the Quick Shift segmentation guided dehazing approach, the haze-free image with better visual quality can be achieved.Thirdly, this article propose a novel learning-based approach for single image haze removal. The proposed approach is mostly inspired by the observation that the color of the objects fades gradually along with the increment of the scene depth. The algorithm regard the RGB values of the pixels within the image as the important feature, and use the back propagation neural network to mine the internal link between color and depth from the training samples, which consists of the hazy images and their corresponding ground truth depth map. With the trained neural network, we can easily restore the depth information as well as the scene radiance from the hazy image. Experimental results show that the proposed approach is able to produce a high-quality haze-free image with the single hazy image and achieve the real-time requirement.Finally, several representative algorithms and research results are presented, followed by quantitative and qualitative evaluations of these techniques.