This paper presents a compression method for still images, based on Kohonen's neural network. To avoid the edge degradation caused by high compression ratio, the blocks are classified into two classes: blocks with high activity (edge blocks) and blocs with law activity. The image is divided first into blocks of 16 pixels. Each block of high activity are divided again into small blocs of 4 pixels. Blocs of high and law activity are coded separately with different codebooks. We have obtained a noticeable improvement of visual quality of all the rebuild images while keeping an important compression rate.
Index Terms. Compression, Medical images, Neural Networks, Vector Quantization, Classification
Cite as: Benahmed Daho, Z., Benamrane, N., Freville, A. (2001) Medical images compression by neural networks. Proc. Models and Analysis of Vocal Emissions for Biomedical Applications (MAVEBA 2001), 169-175
@inproceedings{benahmeddaho01_maveba,
author={Z. {Benahmed Daho} and N. Benamrane and A. Freville},
title={{Medical images compression by neural networks}},
year=2001,
booktitle={Proc. Models and Analysis of Vocal Emissions for Biomedical Applications (MAVEBA 2001)},
pages={169--175}
}