Evolving an Image Comparison Matrix Using Genetic Programming

This paper is to apply the genetic programming technique to evolve a matrix which provides a quantitative indicator to detect subtle difference between two apparently similar pictures. The images here refer to the PPM format images. The GP environment I used is the LIL-GP which is a C language system for developing genetic programming applications. To train the LIL-GP to generate the target matrix, I defined some, terminals and functions composing the GP programs. I also defined the fitness function which was very important to my problem. The fitness case was an array of PPM images composed of one original image followed by some lower quality version images. The ideal of the fitness function was to take the standard image comparison matrix as the reference which can guide the GP to find the target matrix by compute the distance between the reference one and the GP program. The target matrix should perform better than the reference one.