Research In Face Recognition Technology Based On Neural Network

As a specific application, face recognition technology has become a very important issue in pattern recognition, now there are a lot of achievements in this area. At the same time, face authentication and gender classification also have achieved many progresses as a branch of face recognition. As the two-classification question, the crucial technology of them has acquired extensive applications in target tracking & recognition and identity authentication system, so there is enormous research value.The scientific research of this thesis outspreads mainly based on neural network, Firstly, this article introduces developmental general situation and main methods of face recognition. Secondly, in face Authentication area, based on the theory and improved algorithms of BP neural network,the author implements the training and simulation of pyramid network making use of four arithmetic. Then based on the application theory of face authentication acheived by neural network, we apply the Pyramid Network to the Face Authentication experiment. At last, in gender classification area, this article improves the normal BP network, applys the implemention algorithm of PCA by neural network to the first layer of BP. So consider the proportion of features in feature extraction from the source. Then apply the new method to gender recognition area. Moreover, these new models have been contrasted to traditional method respectively, which are "feature extration+normal BP"and "feature extration+minimal distance". And procure good effect. Main improvement work has been listed as follows.⑴, An improved face authentication algorithm is proposed based on the neural network according to LAUSANNE protocol. And based on client subspace method, using the two-classication nature of difference value of projection image and original image, we put forward a new Face Authentication model "Cs_PCA+Pyramid network". At same time, the author does an experiment to compare to the traditional "Cs_PCA+BP" model, and analyzes the performance and advantage.The author makes necessary complement for LAUSANNE protocol, puts forward the internal application of face authentication. Then carry out the experiment once again.⑵, This thesis proposes a new classifier named "PCABP". We apply the GHA arithmetic based on PCA to training process of BP, so the first layer of BP optimizes the eigenvector subspace by GD algorithm in the process of feature extraction. Therefore think of the identification ability and classification capacity from source. The whole PCABP is trained by the Rprop algorithm, so the new network has definite advantage in speed. The author conducts gender classification experiments ground on two type of pictures in the Feret face base.①Contrast the performance of the improved PCABP and normal BP, analyze the superior of the new model.②Compare the neural network to the traditional models, explain and analyse the respective performance in applicable domain.The traditional Eigenspace Separation Transform often comes across a puzzle of higher-dimension pattern, so that the dimension of correlative difference phalanx is too big to compute easily. This article solves this problem successfully, and makes use of EST fully as an important tool of feature extraction.