| dc.description.abstract | In recent years, using biometric features to recognize people has becoming an important topic. This thesis proposed a high performance embedded face recognition system, based on geometric feature/projection feature biometric decision fusion.
We perform our face recognition system on continuous stream images. Most of the traditional research implement their face recognition algorithm on static images, therefore, the final result would depend on the image quality. We can identify a person in stream image many times, and getting a more reliable result. First, we uses the facial color filter and the ellipse mask to partition the possible facial areas, then we using the Particle Swarm Optimization to locate the facial position, After extracting the bi-modal biometrics from human face, we finally recognize human by Probabilistic Neural Network(PNN).
We compare our face detection subsystem with Viola-Jones detector, the experiment reveals a better detection rate and tolerance of bad image quality. Besides, both the operated speed and detection rate, using Particle Swarm Optimization tracking is better than performing detection on every single image. We compare our face recognition subsystem with eigenface algorithm, the former has better EER result and less database storage requirement than the latter.
In this thesis , we verify our face detection、tracking and recognition system on personal computer and embedded platform. Furthermore, according to the MIAT methodology , the system is divided into independent modules, we analyze every modules to find system bottleneck , and then implementing such modules on embedded hardware .
| en_US |