Efficient Construction of Image Feature Extraction Programs by Using Linear Genetic Programming with Fitness Retrieval and Intermediate-Result Caching

Abstract

This chapter describes a bio-inspired approach for automatic construction of feature extraction programs (FEPs) for a given object recognition problem. The goal of the automatic construction of FEPs is to cope with the difficulties in FEP design. Linear genetic programming (LGP) [4]—a variation of evolutionary algorithms—is adopted. A population of FEPs is constructed from a set of basic image processing operations-which are used as primitive operators (POs), and their performances are optimized in the evolutionary process. Here we describe two techniques that improve the efficiency of the LGP-based program construction. One is to use fitness retrieval—to avoid wasteful evaluations of the programs discovered before. The other one is to use intermediate-result caching—to avoid evaluation of the program-parts which were recently executed. The experimental results show that much computation time of the LGP-based FEP construction can be reduced by using these two techniques.KeywordsGraphic Processing UnitCanonical FormCanonical TransformationCache SizeEvolutionary SearchThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.