A clustering algorithm for spectral image processing using Self-organizing Maps

In this paper, we propose a new clustering algorithm using unsupervised neural network for spectral image processing. The proposed algorithm classifies pixels of images after Self-organizing Maps (SOM) assigns each pixel on a two-dimensional feature map. By clustering sequential spectral images, we can detect segments of features which show a distinctive spectral reflectance in the images. K-means algorithm well known as a non-hierarchical Clustering model also classifies the image pixels based on the feature vector distance between each pixel. However, K-means algorithm requires the knowledge of the optimal number of clusters to obtain a meaningful result. The most significant advantage of the proposed clustering algorithm using SOM feature map is that it guarantees to obtain a high quality clustering output without fixing the number of clusters. The proposed algorithm regards a normal distribution formed by the pixels which have a similar feature as one cluster, and divides the feature map into the optimal clusters with the boundary line between normal distributions of several clusters. In the simulation, the proposed algorithm recognized the male and female butterflies from sequential ultra-violet spectral images.