Multi-Feature Analysis Of Face Detection And Recognition And System Implementation

In this thesis, we propose a new Multi-Feature analysis method for face detection and recognition. An extensible real time face detection and recognition system with high performance is developed. We propose Multi-Feature analysis, new weak classifier and Dynamic Cascade learning algorithm to improve face detection performance and compare different face recognition methods. Different face detection and recognition algorithms are analyzed and implemented. Finally, a real-time face detection and recognition system that can be trained online is developed with the implementation of the improved algorithms.The main contributions of this thesis include the following:(1) We propose AdaBoost based algorithm and Multi-Feature selection using Dynamic Cascade structure for face detection, which can lead to a high recall and low false alarm real time frontal face detector. Firstly, in our Real AdaBoost training procedure, we introduce a new kind of weak classifier, called"Bayesian Stump", for training boost classifiers. It provides a solution to the selection of the number of necessary bins and its partition problem during training, resulting in more stable boost classifiers with fewer features. Secondly, we propose a new cascade algorithm called"Dynamic Cascade", which can train cascade classifiers on massive data sets and only requires a small number of training parameters. Also, it makes full use of the information from the previous cascade node so that fewer average weak classifiers are needed for classification. Thirdly, multi features are used for training in different stages. Different feature types are complement to each other while training. The combination use of these features will improve the detection performance. We evaluate the proposed detector in CMU/MIT frontal face database, showing the high performance of the algorithm. Computer simulations also show that the integrated face detection system performs well in real time.(2) Different face recognition algorithms are analyzed and their recognition performances are compared with a large number of simulations. We also proposed online traininig algorithms for face recognition. By using CAS-PEAL and FERET image database, we obtain the performances of different preprocessors for recognition and the comparison result of the different face recognition algorithms. We find that Independent Component Analysis (ICA) architecture II performs excellent in difficult illumination in CAS-PEAL data set while Local Binary Pattern (LBP) outperforms the others in accessory, aging, background, distance, and expression test set. We also find that using histogram equalization and Gaussian blur for image preprocessor will effectively improve the recognition performance.(3) Using improved face detection algorithm and best performed face recognition method we develop a fast online face detection and recognition system. Also, by using multi-thread and smooth mechanism for face recognition the system works in real time with high performance and interacts with users easily. The system is easy to extend and the interface is provided for further implementation.