Borut Batagelj (2007) Human Face Recognition using a Hybrid System. PhD thesis.
Face recognition is one of the most successful applications of image analysis and understanding and has gained much attention in recent years. Existing systems for face recognition generally consist of a detection and a recognition module. In addition there are systems for categorization and analysis of human faces. On the basis of a complete survey of existing methods, we present in this dissertation a modular system for recognition, classification and analyses of faces. We show the usage of this extended system in the field of smart advertising. Various methods for face recognition were proposed and research groups report different and often contradictory results when comparing them. Researchers often test their methods on different databases, even their own private databases; they use even different training sets, normalization and comparison techniques, which also has an influence on the accuracy of the recognition. One of the aims of this dissertation is to present an independent, comparative study of the four most popular methods for face recognition (PCA, LDA, ICA and Gabor wavelets) under the same working conditions regarding normalization and algorithm implementation. Tests are done on two standard face databases: FERET and AR. We systematically test different normalization processes and all possible distance metrics for features comparison. For the case of a closed-set identification, we test the face recognition methods in combination with all possible distance metrics for the best match (rank 1) and for higher ranks on the standard test sets. We conclude that the metric showing the best results at rank one did not always yield the best results at higher ranks and that no particular method-metric combination is the best across all standard test sets, however we can choose an appropriate combination for a specific task. Because the normalization preprocessing is often different and not clearly defined among different reported research tests, we test how different normalization methods influence the performance of different face recognition systems. We also test how masking different parts of the face influences the recognition to find out which part of the face is the most significant for recognition. We use the optimal combinations of methods and metrics to build the hybrid systems in a serial and a parallel manner. Hybrid systems are becoming very important also in biometrics. Most popular systems are multi-modal biometric systems which use fusion to combine more biometrics samples. In this case, we design the hybrid system to work with different recognition methods on the same biometric data. The results on the same standard databases show that with the right combination of methods we can get a better performance and lower time complexity of the face recognition system.
Actions (login required)