Using unrestricted loops in genetic programming for image classification

Loops are an important part of classic programming techniques, but are rarely used in genetic programming. This paper presents a method of using unrestricted, i.e. nesting, loops to evolve programs for image classification tasks. Contrary to many other classification methods where pre-extracted features are typically used, we perform calculations on image regions determined by the loops. Since the loops can be nested, these regions may depend on previously computed regions, thereby allowing a simple version of conditional evaluation. The proposed GP approach with unrestricted loops is examined and compared with the canonical GP method without loops and the GP approach with restricted loops on one synthesized character recognition problem and two texture classification problems. The results suggest that unrestricted loops can have an advantage over the other two methods in certain situations for image classification.