Research On High Dynamic Range Video Based On Multi-Exposure

Based on the requirement for gathering video information of vehicle fairing separation process, and the growing demand for space video monitoring, this paper analyzed the low dynamic range problem of the video system. The problem caused information loses of the scene. Currently studies on video systems mainly focus on compressing video data efficiently, while the low dynamic range problem is ignored. Studies on generating high dynamic range image are also mainly focus on still images. This paper mainly focuses on how to generate high dynamic range image sequence in real time under the current semiconductor technology. The generated high dynamic range image data is provided to the follow-up video compressing system as video source. The paper summarize the problems to be solved when applying multi-exposure method, including fast exposure time selection, fast registering for multi-exposure images, fast merging for multi-exposure images and ghost removing. Then these problems are analyzed.1. The possibility of covering the whole dynamic range of the scene using several exposures is analyzed. The Debvec theory of calculating the response function of an image system is analyzed in detail. Proposed a fast and robust method for exposure time selection on under and over exposure area which is based on system response function. The method utilized the monotony of the imaging system: for any pixel WP that is well exposed, the number of pixels brighter than WP and the numbers of pixels darker than WP are constant in the acquired image under any exposure time. According to this characteristic the exposure time is adjusted to an initial value to make the median value of the image equals to the middle value of the system output range; then adjust the exposure time to make the pixel value on two sides of current histogram be the middle value of the system output range. Thus three low dynamic range images are acquired. The proposed method for adjusting the initial exposure time can converge in 2 iterations which is more fast and stable than average gray control method. As to the exposure time adjusting in under and over exposed area, the proposed method can use the dynamic range of the system more efficiently than fixed exposure time method.2. Analyzed the simple and effective method for registering the multi-exposed images based on MTB (Median Threshold Bitmap). Multi-exposure images are frame samples from continuous video image sequence, and there’s short time difference between frames causing the movement of the vehicle which led the small displacement of background in the images. So registration for multi-exposure images is needed. However conventional registration methods do not apply to multi-exposure images that have grate gray scale difference. This paper analyzed Ward’s registration method based on MTB and analyzed the calculation speed enhancement of pyramid based MTB. It is also showed that the registration speed can be 50 times faster using the proposed improved pyramid decomposition.3. Summarized the current methods for fusing multi-exposure images: radiance map recovering and direct fusion. Several types of these two specific methods are then described, and the multi-scale pyramid decomposition method is focused on. This paper proposed improved methods for reconstruction, Laplacian coefficients fusing rule and Gauscian coefficients fusing rule. Computation is significantly reduced in the premise of ensuring the quality of fusion adapting to real-time video capture. A set of simulation experiments verified the stability of the improved method, fusion quality and calculation speed.4. Proposed a method for removing the Ghosting Artifacts caused by multi-exposure. There’s displacement of object in consecutive frames caused by movement of object. So the object matching between frames are required to remove the ghosting artifacts. This method predicts other exposed image according to the current reference image and detecting the object movement by comparing the difference between reference image and predicted image. However, the approach has its limitations: depending on accurate system response function and requiring system stability. Inaccurate system response function will cause the failure for motion detection and distortion of predicted image, thereby affecting the fusion results.