This paper presents a new methodology based on Genetic Programming that aims to create novel mathematical expressions that could improve local descriptors algorithms. We introduce the RDGP-ILLUM descriptor operator that was learned with two image pairs considering rotation, scale and illumination changes during the training stage. Such descriptor operator has a similar performance to our previous RDGP descriptor proposed in [1], while outperforming the RDGP descriptor in object recognition application. A set of experimental results have been used to test our evolved descriptor against three state-of-the-art local descriptors. We conclude that genetic programming is able to synthesize image operators that outperform significantly previous human-made designs.
展开▼
