Abstract: There is an irreconcilable conflict between the degree of vagueness and the capacity of denoising in conventional image filters.A new video based denoising method，combining special and temporal filters，utilizing both spatial correlation and temporal correlation，is proposed to solve this problem.At first the motion trajectory is gained by a fast adaptive motion estimation method with adaptive thresholds and different search modes.Then a window-size adaptive median filtering is performed with the two windows on the motion trajectory of current and last frame.The temporal redundancy is exploited by a scalar state 1-D Kalman filter.A novel way is proposed to estimate the variance of the state noise from the noisy frames.At last the two estimates are combined with a geometric mean take full advantage of both spatial and temporal relativities.Experimental result shows that the new method is robust than both spatial and temporal methods，especially when treating the pepper & salt noises.As the time consumption of the new method is sum of temporal filter，spatial filter and motion estimation，the computational complexity increased by combination is acceptable，which is also proved in experiments.

Key words: motion estimation, Kalman filtering, median filtering, video denoising

摘要： 传统的图像滤波器在模糊程度与去噪能力之间存在不可调和的矛盾。提出了一种基于时空域联合的方法，从视频的角度出发，同时利用信号的时域和空域相关性进行去噪，以解决这一矛盾。新方法首先通过一种基于自适应阈值、搜索方法切换的快速自适应运动估计方法获得运动轨迹，然后使用自适应窗口大小的中值滤波去除空域噪声，中值滤波中使用了运动轨迹上相邻两帧的对应窗口像素，使用一维卡尔曼滤波器行时域去噪，最后用几何均值结合两次滤波的结果，使信号的时域相关性与空域相关性都得到了充分利用。实验结果证明该算法去噪效果显著，超过了各种时域或空域方法，对椒盐噪声的处理效果尤其突出。实验也表明，由于新方法的时间消耗为时域滤波、空域滤波和运动估计时间的简单累加，因时空联合而增加的计算复杂度属于可接受的范围。

关键词: 运动估计, 卡尔曼滤波, 中值滤波, 视频去噪