HYBRID MACHINE LEARNING APPROACH FOR OBJECT RECOGNITION: FUSION OF FEATURES AND DECISIONS

摘要

Object recognition is considered to be a predominant basic issue in computer vision. It is a challenging issue against inconsistent illumination, partial occlusion, changing background and shifting viewpoint, because considerable variations are exhibited by diversified real world patterns. The virtue of feature fusion lies in its reliability and capability for object recognition in terms of actual redundancy and complementary information. In this paper, we have developed an efficient hybrid approach using scale invariant features and machine learning techniques for object recognition. We extract the scale invariant features, namely color, shape and texture of the objects, separately with the aid of suitable feature extraction techniques. Then, we integrate the color, shape and texture features of the objects at the feature level, so as to improve the recognition performance. The fused feature set serves as a pattern for the forthcoming processes involved in the developed approach. Subsequently, we hybridize the process of object recognition by combining the pattern recognition algorithms like Support Vector Machine, Discriminant Canonical Correlation, and Locality Preserving Projections. Obviously, with three different pattern recognition algorithms employed, we are likely to get three distinct or identical results enumbered with false positives. So in order to reduce the number of false positives, we devise a decision module based on Neural Networks that takes in the match percentage from the chosen pattern recognition algorithms, and then decides the recognition result based on those match values. Our approach is evaluated on the Amsterdam Library of Object Images collection, a large collection of object images containing 1000 objects recorded under various imaging circumstances. The experimental results clearly demonstrate that our approach significantly outperforms the state-of-the-art methods for combining color, shape and texture features. The developed method is shown to be effective under a wide variety of imaging conditions. Finally, we employ empirical evaluation to evaluate our approach with the aid of an accuracy estimation method, such as k-fold cross validation.