Comparison of Boosted Gabor Feature based Local Descriptor. (2016)

• Record Type: Journal Article
• Authors: Lefkovits, Szidónia; Lefkovits, László
• Is Part Of: Procedia technology. Volume 22(2016)
• Page Start: 913
• Page End: 921
• Publication Date: 2016
• Subjects: Gabor filters -- local descriptor -- AdaBoost -- GentleBoost -- LogitBoost.; Technology -- Congresses; Technology -- Periodicals; Engineering -- Congresses; Engineering -- Periodicals; Engineering; Technology; Conference proceedings; Periodicals; 605
• DOI: 10.1016/j.protcy.2016.01.083
• Languages: English
• ISSNs: 2212-0173

Abstract

Abstract: In the domain of computer vision boosting has become a very powerful tool. The method is used to form a strong classifier applied in several applications of pattern recognition and machine vision. Boosting is a sequential algorithm which separates the instances by the selection of weak classifiers and adds them to a final classifier, thereafter modifies the weights of different training data samples and applies the same classification algorithm in iterative way. The final decision is made by the so called final classifier, the responses of which are applied in a weighted voting decision. In this approach we start from a part-based object detection system described in previous articles [14.15]. The developed patch descriptor is based on two-dimensional Gabor wavelets. The Gabor filters describe the neighborhood of a given image pixel in two-dimensional space. From these local descriptors we created different weak classifiers which are used in the training phase of boosting. We have chosen the boosting algorithm for classification because, in the last decade, the best classification results have been obtained by this algorithm in the domain. In this paper we compare three classification methods based on the boosting approach. The first is Discrete AdaBoost, which considers discrete outputs (+1 for objects, -1 or 0 for non-objects). The second, GentleBoost, this algorithm minimizes the exponential loss and returns real values as classification responses. The third, LogitBoost is the fitting of an additive symmetric logistic regression model by log-likelihood to the training set and solves this optimization with Newton numerical method. Finally, we make a comparison of these classifiers in order to draw conclusions regarding the detection rate, false detection rate and other classification measures.