Study On Performance Evaluation Of Lbp-based Descriptors For Face Recognition

The topic of this thesis is stems from the subject of Research on Key Technology of Big Video Data Intelligent Analysis. Local binary pattern(LBP) and its variants have recently attracted considerable attention because of their successful applications in visual detection and classification. So far, none of the work has systematically investigated the performance of the descriptors for face recognition. Thus, some critical questions are mostly unanswered: 1) How about the comprehensive performance of LBP-based descriptors in expression, illumination and age changes. 2) What is the relationship between the main variant factors and performance improvement. 3) Which are the most descriptive facial components, and what are the optional components combination in various condition changes. As to these questions, the main contributions of this thesis are as follows:1) To propose a category approach to LBP-based descriptors, which are categorized into four kinds of the variants, such as selecting neighbors, encoding, LBP feature fusion, LBP related improvement. Moreover, the detail discussion for the descriptors are presented.2) To present a systemic evaluation solution. The evaluation strategy of the the descriptors like comprehensive condition changes, and reference to FRVT and FERET evaluation protocol is performed on several benchmark face datasets. The evaluation criterions include recognition accuracy rate, CMC curve, ROC curve, training time, testing time and mode size.3) To design the evaluation system referred to VFP Evaluator, which easy makes users obtain all of the evaluation results by only configuring the parameters of datasets, algorithms.4) To provide and analyze the evaluation results according to the evaluation schema. Next, the relationship between the main variant factors and performance improvement is suggested.In this thesis, we conduct a comprehensive comparative study on LBP-based descriptors for face recognition, and analyze the evaluation results. The performance results show that most of the LBP-based descriptors are robust to expression but sensitive to illumination. LGBP has the best accuracy rate and U2 LBP has the best time performance and FPLBP has the best space performance. Encoding and LBP related improvement bring great performance improvement. In face components, eyes and forehead have the best descriptive capability.