Noname manuscript No. (will be inserted by the editor) Particle Swarm Optimisation Based AdaBoost for Object Detection Ammar Mohemmed 1 , Mark Johnston 2 , Mengjie Zhang 1 ⋆ 1 School of Engineering and Computer Science, Victoria University of Wellington, PO Box 600, Wellington, New Zealand 2 School of Mathematics, Statistics and Operations Research, Victoria University of Wellington, PO Box 600, Wellington, New Zealand xxx Abstract This paper proposes a new approach to us- ing particle swarm optimisation (PSO) within an Ad- aBoost framework for object detection. Instead of using exhaustive search for finding good features to be used for constructing weak classifiers in AdaBoost, we pro- pose two methods based on PSO. The first uses PSO to evolve and select good features only and the weak classifiers use a simple decision stump. The second uses PSO for both selecting good features and evolving weak classifiers in parallel. These two methods are examined and compared on two challenging object detection tasks in images: detection of individual pasta pieces and de- tection of a face. The experimental results suggest that both approaches perform quite well for these object de- tection problems, and that using PSO for selecting good individual features and evolving associated weak clas- sifiers in AdaBoost is more effective than for selecting features only. We also show that PSO can evolve and select meaningful features in the face detection task. Key words Particle swarm optimisation, AdaBoost, object classification, object recognition 1 Introduction Object detection attempts to determine the existence of specific objects in a set of images and, if present, to de- termine the locations, sizes and shapes of these objects. It is a challenging problem because objects can occur under different orientations, lighting conditions, back- grounds and clutter. It often utilises a trained binary classifier that can distinguish the objects of interest from the background (including objects of other classes). One of the methods that was intensively investigated to improve the performance of object classification is ⋆ mark.johnston@msor.vuw.ac.nz, mengjie.zhang@ecs.vuw.ac.nz to use an ensemble of classifiers. Instead of attempting to build a single (strong) classifier, a bundle of classi- fiers that individually are not necessarily powerful, are grouped to share the burden of the classification task. Studies have shown that the performance of the ensem- ble is better than any of its components acting alone [1]. A large number of combination schemes and ensemble methods have been proposed in literature (for a survey see [2]), which can be categorised into two approaches. The first approach is the use of an ensemble of accurate, well trained classifier members. The effectiveness of this approach depends on the accuracy and diversity of the members [3,4]. To achieve good performance, the indi- vidual members in the ensemble should exhibit low error rates and produce uncorrelated errors. The second approach to ensemble classification is to allow more tolerance to the accuracy of the individual classifiers, i.e., weak classifiers [5]. Two popular meth- ods are Bagging and Boosting, which both rely on re- sampling the features to obtain different training sets for each of the classifiers. Bagging [6] combines classi- fiers each individually trained on a bootstrap replica of the original training set. Boosting refers to a general and provably effective method of producing an accurate en- semble by combining rough and moderately inaccurate rules of thumb. Kearns and Valiant [7] proved the fact that learners, each performing only slightly better than random, can be combined to form an arbitrarily good ensemble hypothesis [8]. One fast, robust detection system, based on learning, is AdaBoost (“adaptive boost”) for object detection by Viola and Jones [9]. This system has three main charac- teristics: using the AdaBoost boosting algorithm to com- bine simple weak classifiers into a more effective strong classifier; use of an integral image to rapidly compute simple features; and using a cascade of AdaBoost clas- sifiers to quickly eliminate most negative images from consideration. Due to its robustness, it has been used in different applications including face and pedestrian de-