Techniques for Fast Partitioning of Compressed and Uncompressed Video

Abstract

Video partitioning is the segmentation of a videosequence into visually independent partitions,which represent various identifiable scenes in thevideo. It is an important first step inconsidering other issues in video databasesmanagement, such as indexing and retrieval. Asvideo partitioning is a computationally intensiveprocess, effective management of digital videorequires highly efficient techniques for theprocess. In general, for compressed anduncompressed video, the basic mechanism used toreduce computation is by selective processing of asubpart of the video frames. However, so farthe choice of this proportion has been maderandomly, without any formal basis. An ad hocselection of this subpart cannot always guaranteea reduction in computation while ensuringeffective partitioning.

This paper presents formal methods for determiningthe optimal window size and the minimum thresholdswhich ensure that decisions on scene similarityare made on a reliable, effective and principledbasis. Further, we propose the use ofneighbourhood-based colour ratios, and derive theratio feature for both uncompressed and transformcoded video. The neighbourhood-based ratiofeatures account for both illumination variationand possible motion in the video, while avoidingthe computational burden of explicit motioncompensation procedures. Empirical results showing the performance of the proposed techniques are are also presented.