Identification of video transitions by multi-scale gradient analysis

The video segmentation problem consists in the identification of boundaries between two consecutive shots. These boundaries can be subdivided in cut and gradual transitions. The first one is characterized by a concatenation of two shots and the second one corresponds to the gradual transformation of a shot into another. The simplest gradual transitions are fade and dissolve. The common approach to solve the identification problem of these features in a video is based on dissimilarity measures between frames. The possibility of representing the video transitions by specific patterns in a 2D image, called visual rhythm, and to apply image processing tools on this image is an interesting alternative to cope with this problem. In this work, we consider the morphological multi-scale gradient operators to identify cut and gradual transitions at the same time on the visual rhythm representation. Three different morphological operators are considered: a multi-scale gradient proposed by Soille and two variants of this operation based on ultimate erosion and thinning concepts. Also, the Canny's filter is used to compute the gradient values.

2003