CNN and SVM Architecture for Face Identification in Multi-View Systems

Identification is an important part of biometric systems. However, when the system works in real-time, additional problems like pose variation or different scale of the head could occur. These adverse conditions decrease the recognition rate and could fail the identification of a person. Multi-view systems could solve this problem. Nevertheless, it is a challenge to calculate the relative position of the sensors and synchronize the system. We implemented an identification algorithm for multi-view systems that recognize individuals with different head poses. A combination of CNN and SVM was implemented for the task. Various configurations of the CNN and SVM were tested. The network was trained and evaluated on the CroppedYale B dataset and our own dataset, generated by Kinect. A recognition accuracy of 94.71% was achieved with SVM’s RBF kernel for CroppedYale and 99.87% with SVM’s linear kernel for our dataset.