Two improvements in genetic programming for image classification

A new classification algorithm for multi-image classification in genetic programming (GP) is introduced, which is the centered dynamic class boundary determination with quick-decreasing power value of arithmetic progression. In the classifier learning process using GP for multi-image classification, different sets of power values are tested to achieve a more suitable range of margin values for the improvement of the accuracy of the classifiers. In the second development, the program size is introduced into the fitness function to control the size of program growth during the evolutionary learning process. The approach is examined on a Chinese character image data set and a grass leaves data set, both of which have four or more classes. The experimental results show that while dealing with complicated problems of multi-image classification, the new approach can be used for more accurate classification and work better than the previous algorithms of either static or dynamic class boundary determination. With the fitness function, the size of the programs in the population can be controlled effectively and shortened considerably during evolution. Thus, the readability of the programs could be seemingly improved.