Research On 3D Face Recognition Technology In Multi-Camera Scene Based On Deep Learning

The increased popularity of video surveillance provides more protection for social public property and citizens' personal safety.The core technology of identification in video surveillance is face recognition.However,due to the impact of angle,light and occlusion,the existing method of face recognition based on two-dimensional images can't get enough face features for an accurate answer.Therefore,it has become an important research topic to improve the efficiency of face recognition under the monitoring scene in public places which are deployed multiple surveillance cameras.Based on the research about 3D model technology and face recognition methods,we propose a method of face recognition based on the 3D face model in multi-camera surveillance scene.We use the 3D face model which can contain more face information as the recognition feature.The proposed method is mainly divided into four stages:image preprocessing,3D model reconstruction,fusion of multi-face 3D model and 3D face recognition.In the stage of image preprocessing,we use multi-task and multi-level convolutional neural network and cascade regression tree to implement face detection and face alignment,which is the data basis for reconstructing 3D face model.In the stage of 3D model reconstruction,the deep residual network ResNet-101 is used to regress the parameters of the 3D face model directly from the 2D face image.Then we use the CAS-PEAL face database consisting of Chinese people to train network model.In the stage of multi-face 3D model fusion,we propose a method of multi-dimensional model fusion by transforming matrix to enchance the feature of face models.Finally,in the 3D face recognition stage,the 3D face model is transformed into a 2D depth map according to weights calculated from X and Y coordinate and then use the convolutional neural network to extract the feature.The face is identified by calculating the cosine distance between the extracted feature and the collected feature in database.We collect face images and build a dataset by multiple cameras.The experimental results show that our 3D face recognition method has achieved good performance on our data set.