Foreground/background segmentation in videotelephony

摘要

The current algorithms for low bit rate image coding are still being improved but to many their performance seems to have approached asymptotic limits. Recently new research considers the use of pattern recognition and artificial intelligence techniques as a promising approach, since they can supply very useful information to improve the quality of the areas to which the observer is particularly sensitive.Videophone applications seem to be especially suited to this approach, since the speaker usually appears before an unchanging background. Here the significant areas are eyes, mouth, face and speaker contour where foreground changes into background. Algorithms detecting the significant areas are useful, since adhoc coding techniques can be used to enchance the coding efficiency thereby making the picture more pleasant to the observer.The aim of this paper is to report the results obtained by the use of techniques to detect the contour on videophone-type scenes. By means of movement detection techniques an area is located where contours have to be searched. Contours are successively detected by using derivative operators and background deletion. This method has been applied to some standard CCITT test sequences obtaining speaker contours as continuous lines.