Image Segmentation and Compression using Neural Networks

Kohonen [1] has developed an algorithm with self-organising properties for a network of adaptive elements. These elements receive an input signal and the signal representations are automatically mapped onto a set of output responses so that these responses acquire the same topological order as the input signal. Images can be used as input signals and the networks can adjust to extract certain topological features. Image segmentation can be performed satisfactorily. By empirical knowledge, it can be supposed that as the number of neurones increases, so does the quality of the segmentation. This paper concentrates on the relationship between the quality of segmentation and the number of neurones that constitute a Kohonen Neural Network. Several experiments were conducted and the Euclidean distance between adjacent neurones measured the quality of the segmentation, which tended to maintain constant after a certain optimum level. The amount of information of the original set of images was compared with the information of the segmented structure and results were presented. Compression rates higher than 250:1 were obtained.