Locality-based linear discriminant projection

This paper presents a novel feature extraction technique, called locality-based linear discriminant projection (LLDP), for multi-class discriminant tasks. The proposed method can be regarded as an improved version of the classical linear discriminant projection (LDA), which is one of the most popular feature extraction methods. LLDP has at least two distinct advantages compared with LDA. Firstly, LDA may fail to discover the potential structure in the data, while LLDP integrates both class label information and neighborhood relationships between samples, so LLDP has the ability to reveal the intrinsic structure in the data which is more helpful to solve such nonlinear problem as recognition tasks than LDA. Secondly, LDA usually can only extract at most C-1 (C is the number of sample classes) features in a real situation, which doesn´t suffice to categorized all samples, while LLDP is able to obtain much more features. Therefore, LLDP has more powerful discriminant ability than LDA. We carry out LLDP and LDA, as well as other current popular algorithms such as PCA, LPP and MFA on the ORL face database and the CENPARPARMI handwritten numerical database, and the experimental results show that LLDP can achieve much higher recognition rate.