Genetic programming for improving image descriptors generated using the scale-invariant feature transform

Abstract

Object recognition is an important task in the computer vision field as it has many applications, including optical character recognition and facial recognition. However, many existing methods have demonstrated relatively poor performance in all but the most simple cases. Scale-invariant feature transform (SIFT) features attempt to alleviate issues surrounding complex examples involving variances in scale, rotation and illumination, but suffer, potentially, from the way the algorithm describes the keypoints it detects in images. Genetic programming (GP) is used for the first time in an attempt to find the optimal way of describing the image keypoints extracted by the SIFT algorithm. Training and testing results show that the fittest program from a GP search can improve on the standard SIFT descriptors after only a few generations of a small population. While early results may not yet show major improvements over standard SIFT features, they do open the door for further research and experimentation.