Face Recognition Based On MPCNN

Face recognition, a non-touch identity authentication technology, is becoming a research hotspot in artificial intelligence and pattern recognition. With increasingly wide range applications, face recognition is being exposed to more disturb and challenges, and demanding higher accuracy and robustness. Traditional face recognition methods extract shallow face features by artificial methods based on priori knowledge. Lacking the capacity of information presentation, shallow features extraction methods will be easily disturbed by illumination, pose and facial expression, and result in poor generalization ability. By digging deep into abstract face feature, Convolutional Neural Networks efficiently solve face recognition potentially. However, the input face images need to be fixed size either via cropping or via warping and multi-scale information of face hasn’t be additionally considered in general CNN. The network structure of CNN need improvement for superior performance.To solve the above problems, a face recognition method based on Multi-scale Pooling Convolutional Neural Networks (MPCNN) was proposed to ensure arbitrary size of face images input to the network and improve the robustness to deformation by replacing the general pooling layer with a multi-scale pooling layer before the fully connected layer. Based on multi-scale pooling, a network structure based on fully connected layer with multilayer feature expression was proposed to fuse multilayer feature of each layer by connecting feature maps from each convolutional layers to the fully connected layer, for improvement of recognition performance. To analyze the performance of the proposed network, comparison experiments were conducted after optimizing the parameters of the network with experiments. The experiments demonstrate the effectiveness of MPCNN, boosting the accuracy of face recognition.