Image Language Terminal Symbols from Feature Analysis

This article presents an approach to generate the image description in the form of sequence of symbols, suitable for further processing with parsing and grammar inference tools. The algorithm assumes presence of a training set of example images. From this set a number of image patches is selected. Using sparse kernel feature analysis, the similarity functions for each symbol are prepared. These functions are used to get terminal symbol locations within analyzed image. A list of symbol locations is then reduced into a tree. Application experiments performed with a database of car images show the potential of the method to represent the structure of object images.