Automatic indexing of video documents



Introduction

This contract, which, in the Center of Mathematical Morphology, is part of an important activity of compression and analysis of image sequences, began two years ago and should lead to the establishment of an indexing toolbox.

Our goal is to extract as much useful information as possible, by means of image processing techniques, which we want to be automated, first to help indexing any video document, and second for information retrieval. Among all the techniques developped in this framework, many of them of course apply not only to sequences, but to single images.

Before going any further in the presentation of the methods used, let us underline that the audio part of the documents, although of great interest, was not used for this project.

The first step in the indexing process is without any doubt the extraction of the document structure. This step begins with the detection of the different transitions between the shots. We have therefore built two techniques of transition detection in sequences.

We also present some interesting results obtained when going on with the shot hierarchization, once the transitions are detected. For a higher semantic level, some of the shots can be grouped together, as belonging to a unique semantic component (a same television report for example).

Several examples of color image segmentation are then proposed, the segmentation being one essential step in the indexing of still images. Let us straight away notice that this segmentation aims at extracting large homogeneous regions in the images, while small details are let apart.

Eventually, we present some results obtained by our techniques.


Last update : 24 - 05 - 99
demarty@cmm.ensmp.fr