Research On Video Face Image Retrieval Technology Based On Graph Embedding

With the study of machine learning and neural network,the technology of face recognition is improved continuously.However,it is still a difficult problem that how to excavate the effective low dimensional representation in the high dimensional data.Graph embedding algorithm uses the undirected weighted graph to map the image data from high dimensional map to low dimensional space,which is to find the essential structural features of the embedded image.This paper explores the application of the learning algorithm based on encoding information in the face recognition,and the unsupervised learning method based on graph embedding method is proposed.Because the technology of image recognition is relatively mature,which is far beyond the video face recognition technology.This paper applies the technology of image recognition to video images,which mainly realizes the video face detection,face recognition and face retrieval function.The main contents and innovationsin this paper are as follows:(1)Because the sample datasare video sequence,the segmentation of moving regions has certain directive function to the area of face detection.There are many methods for the segmentation of moving regions.The frame difference method is greatly influenced by the environment,which contains a lot of noise;The background difference method has to be used to assign the threshold manually in advance;and the optical flow method is used in the iterative method,which is time-consuming and complex.In this paper a combination of frame difference method and adaptive sliding average method is proposed,which can guarantee the real-time performance of the frame difference method,and also has the effect of sliding average method,and also removes the effect of light.(2)In consideration of the light and shade factors in the video image thatleads to face detection results decreased,a face detection method that the ViolaJones algorithm based on AdaBoost is proposed in the moving region.The combining of frame difference method and adaptive sliding average method is used to narrow the range of detection.The face detection is carried out in the area with the highest probability of moving characters,which can effectively reduce the Haar-like features of the backgroundand the detection time,and improve the accuracy of detection.(3)In this paper,the graph embedding algorithm is introduced into the video retrieval algorithm.In consideration of theproblems of shallow learning that the manifold algorithms only take into account the linear embedding of the image features,such of PCA and LDA,and the LLE ignores the information among the classes,and only the local features of the samples are studied.A face recognition based on deep belief network is proposed Under the framework of graph embedding.The deep belief network(DBN)thatis contained by multiple hidden layer restricted Boltzmann(RBM),automatically learns the high order characteristics of image data,whichreduces the impact of related parameters on the characteristics of learning.The abstraction features of the underlying feature combinations are represented by a high level representation,Whichis to obtain the distributed features embedded in high dimensional images,andrealize image dimension reduction.Faceretrieval is complete by using distributed feature.(4)In consideration ofthat the training of deep belief network(DBN)is slow,a new weight updating function is designed,which combines the weight update direction and the direction of the current direction.The performance of the improved deep belief network isanalyzed bythe four parameters,such as the weight coefficient,the hidden unit,the dimension and the number of cycles.It is proved that the function can improve the training speed.Rich human faces meet the requirements of the deep belief network for large data.The deep belief network is applied to the video data,and the distributed features are learned from the video face.The distributed feature has more information in the image,which is used to recognise and retrieve.The simulation is carried out using 2010 b MATLAB software on the standard face database and video database.The results show the effectiveness of the proposed method.