A Rough Set Approach to Face Recognition
MetadataShow full item record
Rough decision tables were originally proposed by Pawlak. The hierarchies of probabilistic decision tables (the hierarchies), defined in the context of Variable Precision Rough Set Theory (VPRS model), generalized this notion. In recent years, the hierarchies have been successfully applied to many research areas. Compared to other facial representation and recognition techniques, the hierarchies have several unique advantages. As a classifier, the hierarchies can update their structures dynamically and meanwhile maintain prior knowledge when new face images appear. In addition, the most relevant features of facial representation and recognition can be selected heuristically, and the structure of the template images representing each category is extremely simple. The objective of this research is to explore the feasibility of applying the hierarchies of probabilistic decision tables to facial representation and recognition, under the framework of the VPRS model. As a result, a new facial representation and classification methodology is proposed in this thesis called soft-cut and probabilistic distance-based classifier (soft-cut classifier). The fundamental contribution of this research is that, for the first time, the hierarchies were applied to face recognition and representation. More specifically, the primary contributions can be summarized as follows. Firstly, a set of approaches was proposed for the application of the hierarchies, including how to represent each facial image with wavelet coefficients and heuristically reduce the redundancy in the coefficients. Secondly, a new method of how to discretize real-valued face features (named soft-cut discretization) was developed. Thirdly, a new method of heuristically selecting the most relevant features as the conditional attributes for each decision table of the hierarchies was proposed. Fourthly, a technique of pruning the hierarchies to simplify their structures was suggested and applied in the presented methodology. Fifthly, a new hybrid matching method called probabilistic distance-based matching method was proposed, classifying the unknown images into each decision category after the hierarchies have been formed. Finally, it was shown how to modify the structure of the hierarchies for the purpose of incremental learning and, for the first time, implement it in the context of the VPRS Model. In order to determine whether the hierarchies can be successfully applied, various experiments have been conducted. Based on the experimental results and a comparison to other existing methods, if the hierarchies are formed by a training set with a small number of participants, the presented methodology has a higher accuracy rate (greater than 90%). During the process of incremental learning, the hierarchies can always learn from new images and meanwhile retain prior knowledge. Following incremental learning, the performance of the hierarchies can always be improved upon by making the training set either larger or smaller.