The increase in computer resources, associated with the advent of powerful classification methods such as AdaBoost and neural networks allow current object detection systems to reach high detection rates. However those methods require a large training database of several thousands of examples. This paper presents an object detection method that gives state-of-the-art results, while using a reduced training database. First, we present the various methods used in object detection systems, and particularly the machines learning methods involved. Then, we explain a detection system based on correlation and using a database of less than one hundred images. This system allowed us to develop a method of association of similarity measures using orthogonal edges filters obtain using a method derived from the PCA. We then show that we can develop a face detection system able to work with a small number of examples. The similarity measure based on correlation is the most limiting factor of our detection system; that's why we replaced it by a Multilayer Perceptron. We then applied the association of filtered images to the new similarity measure and showed an improvement in detection rate using a small learning database. Finally, we highlight the possible solutions able to improve the detection system speed and to decrease the number of learning examples.
Robust matching of complex visual forms, application to object detection
© TELECOM ParisTech. Personal use of this material is permitted. The definitive version of this paper was published in Thesis and is available at :
PERMALINK : https://www.eurecom.fr/publication/2991