基于运动分割的视频去模糊

摘要

It is difficult to carry out large related motion blurs and keep both spatial and temporal coherent in deblurring results using existing video deblurring methods. In this paper, we propose a segmentation-based video deblurring method to deal with such large related motion blurs. Since the sharpness is varying on vid-eo frames, blurry frames can be restored by sharp patches. Our method firstly divides a blurring frame into several motion-based regions using optical flow between the blurring frame and its adjacent sharp frame. Then, the blur function is estimated for each region to blur sharp neighbor frame. Finally, we search the nearest patches from blurred sharp frames and restore the final results using region-based synthesis. To im-prove the efficiency, the method is implemented based on GPU acceleration. Experimental results demon-strate that our segmentation-based video deblurring method can remove blurry artifact effectively on real video sequences.%针对目前视频去模糊方法难以处理大幅度相对运动模糊, 并且很难得到在时间和空间都保持一致性的去模糊结果的问题, 提出基于运动分割的视频去模糊方法. 基于视频序列清晰程度不同的前提下, 用视频中的清晰块恢复模糊帧. 根据模糊帧与其相邻的清晰帧的光流信息, 将2帧之间存在的相对运动分割为独立的处理对象; 不同运动对象分别估计模糊函数, 并利用该模糊函数将清晰帧模糊处理; 对模糊帧中的每一块区域, 在被模糊处理后的帧中查找最相似的区域, 利用对应清晰区域替换模糊区域; 不同区域之间采用纹理融合, 重建出清晰帧, 实现对视频中大幅度相对运动去模糊. 该方法基于并行思想设计和实现, 利用 GPU 并行能力完成加速. 实验结果证明, 采用该方法不仅速度快, 而且有效地解决了视频中大幅度相对运动模糊, 并可以保持运动对象纹理结构的完整性.