Research On Image Defogging, Effect Assessment And Application

Fog removal for degraded image is a fundamental and hot problem in computer vision, which has a wide applicate prospect. However, the reasons for foggy image degradation are very complicated and the information in foggy image itself is insufficient. Currently, the degraded process of a foggy image can not be described by any algorithms and models perfectly, and the previous algorithms are not good enough when using them in particular scene. There is still much works to improve the visual effect. Thus, it is necessary to research the fog-degraded image clearness techniques based on the analysis of image degradation mechanism.There are four sub-problems involved in automatic defogging technique:fog removal for single image, fog removal for video, objective assessment of defogging effect for color image and the application of defogging technique. Dedicated to the four sub-problems the thesis developed a deep research, explore the new theory and method for defogging technique, and laid a theoretical foundation for improving the performance of vision system in bad weather condition. The main works of the thesis in research are shown as following:Due to the high time-temporal complexity of the defogging algorithm based on dark channel prior, and the user interaction of the defogging algorithm based on atmospheric veil. A fog removal algorithm based on the transmission gradient prior is proposed in this thesis. According to the prior, only small amount of pixels need to be refined by using fast bilateral filter on the atmospheric veil, while most pixels in transmission map can be directly estimated by using dark channel prior. The proposed algorithm not only obviously reduces the computation cost, but also has no complex matrix and too many parameters. The method has been tested using a lot of foggy images, and is found to give a good results.According to the frequency spectrum of the foggy image, a fog detection method based on Fourier spectrum of the entire image is proposed. Several experiments demonstrate the effectiveness of the proposed method. Aims to the block effectiveness problem of the previous video defogging algorithm, two new video defogging algorithms based on fog theory are proposed in the thesis. One is regarding fog as the veil layer to be subtracted, and the other is taking fog as the transmission map to be separated from the original video. The former uses the luminance component image obtained by Retinex algorithm and the depth information of the original video frames to separate the veil layer. The latter applies a single transmission map obtained from the background image to a series of video frames. Experiments show that both algorithms can effectively improve the quality of the video frames. Compared with other algorithms, our algorithms restore video frames from a perspective of fog with no reference image and low computation cost. The new algorithms can remove fog effectively as well as provide a good practicability and a fast speed.A framework for defogging effect assessment is constructed in the paper. By analyzing the limitation of the previous assessment method based on contrast enhancement. We propose that for the scene where the visibility can be detected, the visibility increased-value through fog and fog removal images is compted as an important assessment index. While for the scene where the visibility can’t be detected, two new assessment methods for the clearness effect of image defogging algorithm are proposed in the thesis. One is full-reference method based on the synthetic foggy images which are obtained by environment rendering or transmission map, and the other is developing comprehensive assessment system to assess defogging effct from human vision perception. Experiment show that both methods can assess the effect effectively, and the assessment results are consistent with our subjective perception. Compared with other existing methods, our proposed methods assess defogging effect from the visibility computation, synthetic foggy image and human visual perception, respectively. The new methods can obtain an overall conclusion as well as provide a good practicability and reliability.By analyzing the feature of foggy traffic image, a defogging algorithm, targets those traffic scene is presented by introducing the visibility concept into the defogging process. The algorithm can obtain a better defogging effect by limiting the possibilities of enhancement at short distance, while strongly enhancing the long distance area that is important for drivers, which is different from the unified enhancement of the previous defogging algorithms. Meanwhile, we propose some applications of the proposed defogging algorithm:improvement of road marking features extraction, improvement of circular road signs detection and improvement of traffic police gesture recognition. Experiments performed in those application scenes demonstrate the effectiveness and the practicality of the proposed algorithm.