Introducing a method for extracting features from facial images based on applying transformations to features obtained from convolutional neural networks

In pattern recognition, features are denoting some measurable characteristics of an observed phenomenon and feature extraction is the procedure of measuring these characteristics. A set of features can be expressed by a feature vector which is used as the input data of a system. An efficient feature extraction method can improve the performance of a machine learning system such as face recognition in the image field. Most of the feature extraction methods in facial images are categorized as geometric feature extractor methods, linear transformation-based methods and neural network-based methods. Geometric features include some characteristics of the face such as the distance between the eyes, the height of the nose and the width of the mouth. In the second category, a linear transformation is applied to the original data and displaces them to a new space called feature space. In the third category, the last layer in the network, which is used for categorization, is removed, and the penultimate layer output is used as the extracted features. Convolutional Neural Networks (CNNs) are one the most popular neural networks and are used in recognizing and verifying the face images, and also, extracting features. The aim of this paper is to present a new feature extraction method. The idea behind the method can be applied to any feature extraction problem. In the proposed method, the test feature vector is accompanied with the training feature vectors in each class. Afterward, a proper transform is applied on feature vectors of each class (including the added test feature vector) and a specific part of the transformed data is considered. Selection of the transform type and the other processing, such as considering the specific part of the transformed data, is in such a way that the feature vectors in the actual class are encountered with less disturbing than the other ones. To meet this goal, two transformations, Fourier and Wavelet, have been used in the proposed method. In this regard, it is more appropriate to use transformations that concentrate the energy at low frequencies. The proposed idea, intuitively, can lead to improve the true positive (TP) rate. As a realization, we use the idea in CNN-based face recognition problems as a post-processing step and final features are used in identification. The experimental results show up to 3.4% improvement over LFW dataset.