A Study On Min-max Modular Networks And Facial Attribution Classification

Researches on the min-max modular network (M~3-network) and facial attribution clas-sification are the main targets of this thesis. These techniques and solutions include thepruning technique of redundant modules of M~3-network, the incorporation of prior knowl-edge into the learning of M~3-network, the expansibility of M~3-network, and their applica-tions to gender classification and age estimation. The main contributions of this thesis canbe summarized as follows:We first propose a back-searching method for pruning of redundant modules of the min-max modular network. The redundant problem of min-max modular network is very impor-tant, especially in case of linear M~3-network. Although independent modules of M~3-networkcan be run on a large-scale parallel machine, however there are still too many modules inpractice. The pruning of redundant modules can largely reduce the scale of an M~3-networkand its testing time. We propose a new way to prune the redundant modules of linear M~3-network, and examine the algorithm on several standard data sets. Finally we apply theproposed algorithm to a practical industrial image fault detection problem and we develop ademo system for demonstration.We propose local binary pattern for the feature extraction of gender classification, andobtain high classification performance. Furthermore, we propose a multi-resolution localbinary pattern for facial feature extraction. This method is applied to gender classificationand age estimation. We obtain high performance by using M~3-network.We study the importance of prior knowledge for pattern classification and machinelearning, then propose a method to incorporate the prior knowledge into the learning ofan M~3-network. In some of real world applications, such as gender classification and age es-timation, we propose to incorporate face degree information, ethnic information and genderinformation as prior knowledge into the learning of M~3-network. We compare the perfor-mance of prior knowledge based methods and non prior-knowledge based methods in manyaspects to evaluate the efficiency of the proposed method.We point out that min-max modular network has a high expansibility, and study onhow to apply this expansibility in practice. In one of our real-world applications, we learn the new added modules, with facial pose information and ethnic information, to completethe M~3-network expansion. Using expansion, the large-scale classification problem can besolved effortlessly and efficiently.Considering the attractive parallel learning, prior knowledge incorporation, and expan-sibility of M~3-network, we propose to apply M~3-network to gender classification. In thesimulations, we demonstrate that M~3-network has many advantages to deal with large-scaleclassification problem such as gender classification. The attractive priorities include: highergeneralization ability, faster training and testing time, free expansibility and prior knowledgeincorporating ability. Finally we design a gender classification demo system for demonstra-tion.We also propose to apply M~3-network to age estimation. In the simulations, we alsodemonstrate that M~3-network has many advantages in dealing with large-scale classificationproblems such as age estimation. These attractive advantages include higher generalizationability, faster training, testing time, and free prior knowledge incorporating ability. Finallywe design a age estimation demo system for demonstration.