Research On Feature Extraction And Thresholds Balance Method Of Face Verification

In network and information age, with the increasing requirements of social public security and information security, research on face recognition has been paid great attention and has made great progress. Face recognition includes two modes: face identification and face verification. Face verification is just decides whether the testing face is designated user. At present, most research focuses on face identification, but study for face verification is much fewer. Face verification is a typical two-class mode. Illumination, expression and other factors make the distribution of intra-class and extra-class has great differences, therefore, feature extraction is still a core problem in face verification. Besides, the thresholds problem is another key and difficulty in face verification because negative samples set is great and thresholds are directly related to error rate.The thesis studies the feature extraction and threshold balance problem, the main works are as follows:①A feature extraction algorithm DC-DCV is proposed by combing 2D Dual-Tree Complex Wavelet Transform (DT-CWT) and Discriminative Common Vector (DCV). The algorithm first using 2D DT-CWT represents face with coefficients at different scales and different orientations to extract important local features. 2D DT-CWT overcomes many deficiencies of Gabor but the dimensionality of 2D DT-CWT features is usually high, so features are then projected into DCV subspace to reduce the dimensionality and enhance the discriminative ability, and DCV also can avoid Small Sample Size (SSS) problem. The experimental results on ORL face database and FERET subset demonstrate the effectiveness and robustness of the proposed DC-DCV algorithm.②Considering the particularity of face verification, the thesis introduces Client Specific Subspace (CSS) to DC-DCV and proposes another feature extraction algorithm, e.g. DC-DCV-CSS. DC-DCV-CSS based on DC-DCV and then applies CSS to build a subspace for every user. The client and impostor samples are different for every user, thus all samples share one subspace in face identification is not suitable for face verification. The experimental results on ORL face database and FERET subset show that DC-DCV-CSS can obtain better verification performance and it is a desirable algorithm for face verification to extract features.③Compared to face identification, face verification has a specific process, e.g. setting thresholds and the thresholds have a direct influence on verification performance. The thesis presents a detailed description of setting method of"user specific thresholds", and analyses the problem about thresholds balance, then two solutions to that problem are studied:1) Reducing the difference between positive samples and negative samples by creating virtual samples. The experimental results on ORL face database and FERET subset verify that this way of adding positive samples has some effect on solving thresholds banlance problem.2) Reducing the difference between positive samples and negative samples by preferential random sampling. The experimental results on ORL face database and IISL face database indicate that this method can banlance thresholds and reduce error rate to a certain extent.