Face Recognition In Video Based On Deep Structure Learning

With the development of science and technology, the video data in the Internethave an exponentially increase. Video data has become an important source forpeople to obtain information. People take more attention on the vast faceinformation in the field where human-computer interaction is needed. How to utilizethe face information to obtain the body identification has great application value.Face recognition includes face detection and face classification, the technologybased on complexion and dynamic threshold was used to pretreat video frame imagein our subject, which can precisely carved out skin from figure. The subject is focuson face classification, which applied deep structure method to achieve the task.There has a great significant using deep structure to classify face. The purpose of thedeep structure is to simulate the progress of recognition of human brain, which iscalled integrated full-level information network system. With progressivelayer-layer, the structure can find the relationship between every pixel, and getrelated line, get related outline and get related object in the top level. By this deeptraining progress step by step, the structure can obtain accurate feature extraction,and then make a classification over these characters by association method.The major source of design idea of our subject: in the process of video facerecognition, the deep structure of DBN can make a good distinguish for differentface image, but the weight is “directional” and “biased” when the weight isinitialized. Then after several training, the extracted feature is biased. Examine theeffect of the prejudice to the classification is our concern in the subject.Face classification based on deep structure method consists of two parts,feature extraction and classification. In the subject, feature extraction uses threestructures, DBN, deep MLP and DBM, these three structures belong to feedbackneural network in a strict sense, and they all stacked by some sub-modules. Infeature extraction phase, we train a single sub-module at first, and stack the secondsub-module on the first sub-module. We utilize the feature data extracted by theprevious sub-module as the input data of next sub-module, one by one, and we getthe feature needed. In the classification section, the subject used two structures, oneis BP feed-forward neural network, and another is RBM classification. BP neuralnetwork have used conjugate gradient optimization, which is differ from traditionalBP neural network. For RBM structure, which can not only be used in featureextraction, but also in classification, but their training process have some difference.Our subject builds two kind of neural networks for face recognition and comparedwith DBN neural network and BP neural network.