Feature representation selection based on Classifier Projection Space and Oracle analysis

摘要

One of the main problems in pattern recognition is obtaining the best set of features to represent the data. In recent years, several feature extraction algorithms have been proposed. However, due to the high degree of variability of the patterns, it is difficult to design a single representation that can capture the complex structure of the data. One possible solution to this problem is to use a multiple-classifier system (MCS) based on multiple feature representations. Unfortunately, still missing in the literature is a methodology for comparing and selecting feature extraction techniques based on the dissimilarity of the feature representations. In this paper, we propose a framework based on dissimilarity metrics and the intersection of errors, in order to analyze the relationships among feature representations. Each representation is used to train a classifier, and the results are compared by means of a dissimilarity metric. Then, with the aid of Multidimensional Scaling, visual representations are obtained of each of the dissimilarities and used as a guide to identify those that are either complementary or redundant. We applied the proposed framework to the problem of handwritten character and digit recognition. The analysis is followed by the use of an MCS built on the assumption that combining dissimilar feature representations can greatly improve the performance of the system. Experimental results demonstrate that a significant improvement in classification accuracy is achieved due to the complementary nature of the representations. Moreover, the proposed MCS obtained the best results to date for both the MNIST handwritten digit dataset and the Cursive Character Challenge (C-Cube) dataset.