Improving the visualization technique of SOM neural network for image

Abstract

The Self-Organizing Map (SOM) neural network generates a feature map of the data. In order to observe this map, it is necessary to use visualization techniques. The U-matrix distance matrix is one such technique and is commonly used in visualization. However, the size of the U-matrix is four times larger than the size of the map, which increases the computational complexity. This paper proposes an improved U-matrix with the same size as the SOM map but still ensures the same visual ability as U-matrix. We further propose a partitioning algorithm with a predefined number of regions for the resulting visualization matrix (improved U-matrix). Results of partitioning for improved U-matrix are tested for the image segmentation problem.