The motion adaptive spatial filter(MASF) is a kind of temporal or spatio-temporal filter with spatial mask. This is possible by using the motion information under the two assumptions:1) $\underline {\it {constant translation}}$:each pixel in video signal has locally constant translational motion, 2) $\underline{\it {signal invariance}}$ : signal is not changed along the motion trajectory. Typically, the assumptions are granted to be true in most part of image data except for motion boundary. Under these assumptions, signal variation along the temporal axis can be estimated in the spatial domain by using motion vector. This is basic concept of the MASF.
The MASF has advantages compared with conventional filters in the aspect of implementation cost and filtering performance. Therefore, the MASF is expected to provide performance improvement in various applications such as subband coding, image format conversion, noise reduction, and so forth. However, there are several factors that affect the performance of the MASF.
First, there may be some error in the motion vectors estimated in real image sequence. The inaccurate motion information causes erroneous temporal data prediction in the spatial domain, and therefore, makes distortions in the MASF operation. Second, the assumption of constant translation may not be satisfied in some cases. One of the cases is due to general 3-D motion such as rotation, zoom, deformation, etc., and another one occurs in motion boundary. In this dissertation, the error analysis of MASF for these performance factors and improvement method based on the analysis results are provided.
Also,3-D extension and application of MASF are investigated. The MASF for 3-D filtering with separable spatial and temporal characteristics has been researched in previous works. In this dissertation, a generalized MASF design method, which can design 3-D MASF with arbitrary shaped passband(separable or not),is proposed. In addition, as an application of the MASF, a temporal interpolation method using the MASF is proposed. Experimental results show that the proposed method gives acceptable performance in the aspect of interpolated image quality and implementation cost.