Machine learning approaches for visual object detection

Abstract

Visual object detection is a fundamental and challenging problem in computer vision and image processing. The study of visual object detection usually focuses on two aspects, proposing effective yet efficient features, and designing powerful and fast detectors. While feature extraction is a domain-specific problem for image processing that usually requires substantial knowledge, experience, and even inspiration, designing the detectors usually relies on techniques of pattern recognition and machine learning. In this thesis, we study the machine learning approaches for visual object detection. We first review several theoretical machine learning issues, in particular, the structural risk minimization learning principle. Then, several empirical loss functions and optimization methods for solving the support vector machines are discussed. Based on insights into the problem formulation and the solving techniques, we propose a generalized forward feature selection scheme that is applicable to a number of problems. For the sake of comprehensiveness, learning problems other than classification, e.g. ranking and structural prediction, are also studied for the purpose of visual object detection. The Boosting cascade detector is the most popular method in visual object detection. Boosting essentially also falls into the structural risk minimization regime, but with specialized algorithms that treats the problem as a functional optimization. We examine both the functional optimization formulation and the convex optimization formulation of the AdaBoost algorithm, and propose a mixed form that solves the totally-corrective AdaBoost but using weak classifiers derived from the functional optimization perspective. To alleviate the training cost, we propose a feature subset selection method based on the partial least square regression. In building a cascade detector, three important issues are studied, i.e. optimizing for the the asymmetric objective, selecting a proper operating point for each stage of the cascade, and recycling information from the early stages of the cascade. A novel biased-selection strategy for information recycling is proposed. For the last part of this work, we study the kernel methods. Various methods for improving the efficiency of the kernel scoring function are discussed, and are grouped into three categories, i.e. the approximations of the scoring function, the explicit feature map, and learning a sparse basis set. Inspired by the kernel methods, we propose a middle-level feature based on the similarity to exemplar instances. For the visual object detection problem, elements from multiple kemel learning and multiple instance learning can be conveniently incorporated into the similarity feature, and a learning framework using the forward feature selection technique and a coarse-to-fine scheme is proposed to learn an efficient visual object detector using the similarity features.