Objects extraction and recognition for camera-based interaction : heuristic and statistical approaches

Abstract

In this thesis, heuristic and probabilistic methods are applied to a number of problems for camera-based interactions. The goal is to provide solutions for a vision based system that is able to extract and analyze interested objects in camera images and to use that information for various interactions for mobile usage. New methods and new attempts of combination of existing methods are developed for different applications, including text extraction from complex scene images, bar code reading performed by camera phones, and face/facial feature detection and facial expression manipulation.

The application-driven problems of camera-based interaction can not be modeled by a uniform and straightforward model that has very strong simplifications of reality. The solutions we learned to be efficient were to apply heuristic but easy of implementation approaches at first to reduce the complexity of the problems and search for possible means, then use developed statistical learning approaches to deal with the remaining difficult but well-defined problems and get much better accuracy. The process can be evolved in some or all of the stages, and the combination of the approaches is problem-dependent.

Contribution of this thesis resides in two aspects: firstly, new features and approaches are proposed either as heuristics or statistical means for concrete applications; secondly engineering design combining seveal methods for system optimization is studied. Geometrical characteristics and the alignment of text, texture features of bar codes, and structures of faces can all be extracted as heuristics for object extraction and further recognition. The boosting algorithm is one of the proper choices to perform probabilistic learning and to achieve desired accuracy. New feature selection techniques are proposed for constructing the weak learner and applying the boosting output in concrete applications. Subspace methods such as manifold learning algorithms are introduced and tailored for facial expression analysis and synthesis. A modified generalized learning vector quantization method is proposed to deal with the blurring of bar code images. Efficient implementations that combine the approaches in a rational joint point are presented and the results are illustrated.