Real-time Flicker Correction In Image Sequence

Unable light cause flicker in image sequence of moving object when multiple camera realizing 3D reconstruction in the laboratory. Intensity flicker affects the object contour extraction, which decrease the accuracy of reconstruction model. This paper is to remove flicker in image sequence and provide images with stable intensity for object contour extraction. To process real time, the processing speed must be within 15ms/frame. In recent years, though the processing speed has improved, it still can’t process real time. In view of the real-time problem, this paper uses the background of the video making small change, GPU general computing parallel processing platform and CUDA architecture to remove flicker in multi-view image sequence real time.The flicker correction algorithm used in this paper mainly includes four parts: parameter estimation based on block, motion area detection, motion vector estimation, image correction. The linear flicker model is used in this paper and the flicker is corrected based on block. When correcting the static video, the repaired result of flicker correction algorithm reference to the first frame, previous one frame and previous multiple frames is compared in this paper and analyze the effect of block size on repaired result. When correcting dynamic video, considering the influence of motion to parameter estimation, this paper proposes that the image is divided to static area and motion area by motion diction. Correction of the static area is like correcting static video, using a method based on multiple frames. The parameter of motion area is corrected based on motion vector. This paper proposes a motion diction method based on flicker space continuity, which can determine the motion area accurately, and inter-frame difference method is realised based on block in this paper. When estimating motion vector, in order to adjust to the block motion type, improve robustness and avoid local optimum if the block moves fast, this paper use adaptive search and search stop rules to improve the cross diamond search algorithm proposed by Cheung. The improved algorithm ensures the search accuracy and improves the search speed. The parameter estimation, image correction and SAD calculation of matching point in adaptive cross diamond search is parallel implemented based on CUDA which reduces the processing time.The experiment shows that the intensity mean and standard deviation curve of repaired image sequence for static and dynamic video is smoother than the original flicker image sequence. The processing speed of flicker correction algorithm based on CUDA is 12.8ms/frame achieving real-time processing.