In this paper we present a new theoretical approach to the problem of optimal Vector Quantization. We base, in fact, our method on the a priori explicit analysis of the effects on the MSE distortion introduced by an arbitrary exchange of training vectors among clusters. Even when the theoretical results corresponding to the simplest possible case are used, the proposed algorithm outperforms the GLA method to an impressive extent both in speed and in performance. Experiments on different images from the USC database have proved that the proposed algorithm is 5 to 10 times faster than the GLA method increasing the convergence Peak Signal to Noise Ratio (PSNR) of up to 1.12 dB.

No abstract

No abstract

Object‐oriented approaches to still and moving picture coding, thanks to the extensive use of efficient algorithms for image segmentation and region coding, are expected to improve significantly the rate‐distortion performances of conventional block‐based methodologies. Recent results show that most of the chances to win this challenge depend on the efficiency in image segmentation: the precise aim of our investigation and the purpose of this paper is just to show the concrete possibility of employing fast adaptive algorithms for segmenting images and video frames and allowing high quality coding with very high compression. The segmentation algorithm we propose relies on the recursive computation of the Spatial Gray Level Dependence Matrix (SGLDM) and on a fast adaptive procedure for the statistics evaluation. Subsequent clustering of the image pixels in the 5‐dimensional statistical‐spatial space MVCXY, composed of local mean (M), local variance (V), local correlation (C) and spatial coordinates (XY) is encharged of providing final image segmentation. Target applications for videophone coding of color sequences at very low bitrate are foreseen.

No abstract