Class Similarity Based Orthogonal Neighborhood Preserving Projections for Image Recognition

Abstract

Recent Dimensionality reduction methods like Locality Preserving Projections and Neighborhood Preserving Projections learn local neighborhood characteristics and try to preserve these characteristics in the lower dimensional space. In supervised settings, conventional Orthogonal Neighborhood Preserving Projections (ONPP) uses knowledge of class label to identify the neighbors of data points. When data points are closely placed or the classes are overlapping, such hard decision rule may not help finding a good low dimensional representation. To overcome this limitation, we are proposing a novel neighborhood selection rule, where low dimensional representation is used with Logistic Regression to find the probability of every data point to belong into a particular class. Based on these probabilities a new distance measure - Class Similarity based distance is used to define a local neighborhood of data points. It is observed that class similarity based ONPP very well represents the relationship of neighbors in low dimensions. The proposed scheme is used to recognize face images and handwritten numerals images. The proposed Class Similarity based neighborhood scheme achieves same recognition performance with significantly less number of subspace dimensions.