Using Evolution to Learn How to Perform Interest Point Detection

The performance of high-level computer vision applications is tightly coupled with the low-level vision operations that are commonly required. Thus, it is advantageous to have low-level feature extractors that are optimal with respect to a desired performance criteria. This paper presents a novel approach that uses genetic programming as a learning framework that generates a specific type of low-level feature extractor: interest point detector. The learning process is posed as an optimization problem. The optimization criterion is designed to promote the emergence of the detectors´ geometric stability under different types of image transformations and global separability between detected points. This concept is represented by the operators repeatability rate. Results prove that our approach is effective at automatically generating low-level feature extractors. This paper presents two different evolved operators: IPGP1 and IPGP2. Their performance is comparable with the Harris operator given their excellent repeatability rate. Furthermore, the learning process was able to rediscover the DET corner detector proposed by Beaudet