Temporal-Spatial Analysis On Event- Related Potential In Face Recognition And Its Application In Rapid Face Image Retrieval By Single-Trial Detection

The face recognition technology is widely used. However, the performance of computer vision based face recognition technology is limited when large illumination, pose or facial expression variations are presented. Research on cognitive neuroscience has found that human brain has extra ordinary ability to recognize faces, Some face-specific event-related potential (ERP) components have been found, which makes it possible to.incorporate human cognitive ability into face recognition. In this paper, we analyze the spatial-temporal features of ERP components evoked in the process of face recognition, and implement a rapid face recognition based on single-trial ERP detection.Preprocessing methods are used to extract the features of ERP signals induced by the face images, and the ERP signal analysis and single-trial ERP detection are implemented to test the effectiveness of the methods. Scalp topography analysis, difference wave analysis and T-test statistical test are employed to analyze the spatiotemporal features of the ERP signals induced by the target and non-target face images. The results show that the ERP components evoked by the target and non-target face images have significant differences in the N2, P3 and N4 components. In single-trial ERP detection, the features are extracted by the common special pattern (CSP) and then classified using linear discriminate analysis (LDA) and support vector machines (SVM) respectively. The average area under the receiver operating characteristic (ROC) curves(AUCs)across 12 subjects are over0.85 for the LDA and SVM, which proves the effectiveness of the method.Deep learning based convolution neural network (CNN) model is implemented to detect the target-face induced ERP signals. Since the number of the samples from single subject is relatively small for CNN training, the training samples from all the subjects are used to train CNN. Then the train data from each subject is used to further optimize the trained CNN model. The testing results show that the CNN model can greatly improve the AUC with the average of 0.927 across 12 subjects, which is 4.4% higher than the combination of CSP and SVM (the right-tail paired t-test, p=0.0021).