On a New Similarity Analysis in Frequency Domain for Mining Faces within a Complex Background

Abstract

A novel similarity analysis is presented in this paper for dealing with the problem of mining faces in a complex image background. The proposed approach integrates a robust feature extraction technique based on a specific method of eigenanalysis in the frequency domain of the unique classes identified in the problem at hand, with neural network based classifiers. Such an eigenalysis aims at identifying principal characteristics in the frequency domain of the above mentioned uniquely identified classes. Each unknown image, in the testing phase, is then, analyzed through a sliding window raster scanning procedure to sliding windows identified, through a first stage neural classifier, as belonging to one of the unique classes previously mentioned. After such a sliding window labeling procedure it is reasonable for a second stage neural classifier to be applied to the testing image viewed as a sequence of such labeled sliding windows for obtaining a final decision about whether a face exists within the given test image or not. Although the proposed approach is a hierarchical procedure, its most critical stage is the similarity analysis performed through eigenanalysis in the frequency domain, since, if good identification/ labeling accuracy could be then obtained, it would facilitate final face mining.