• Technical Name
  • AI Occlusion Resistant Face Detection and Recognition System
  • Operator
  • National Cheng Kung University
  • Booth
  • Online display only
  • Contact
  • 吳忠崇
  • Email
  • gg579ya@gmail.com
Technical Description The system consists of five modules, including masked faces generator module, training module, face detection module, face recognition module and verification module.
(1) The masked face generator is used to change face image from unshielded one to masked one.
(2) The training module is used to train the "face detection network" and "face recognition network" via 6.5 million face images and masked ones obtained from the masked face generator.
(3) The face detection module uses the Single Stage Headless Face Detector (SSH) with the feature pyramid to achieve accurate and fast face detection.
(4) The face recognition module is inputted the face image obtained from the detection module. Moreover, ResNet is used to extract feature which can learn more much facial information and classify more precisely. However, the classification result is just temporary result.
(5) The verification module makes the system more accurate by verifying the temporary result.
Scientific Breakthrough The system uses 6.5 million face images and masked face generator module to increase the richness of data. The face detection uses Single Stage Headless Face Detector (SSH) with the feature pyramid. The face recognition is trained with ResNet. Thus, this system can achieve fast, precise, and multi-angle face recognition. Finally, use the verification module to make the identification more accurate.
Under 50% covering rate, the accuracy of the paper, \"Robust Point Set Matching for Partial Face Recognition, IEEE 2016\" is 56.67%, \"Disguised Face Identification (DFI) with Facial KeyPoints Using Spatial Fusion Convolutional Network, IEEE 2017\" is 67%, and \"Dynamic Feature Learning for Partial Face Recognition, CVPR 201\" is 56.8%. However, this system has 95%( FAR<0.001%, FRR<5%) accuracy.
Industrial Applicability In general face recognition systems, the restrictions on face conditions are strict, or it is easy to make misjudge with. As a result, user have poor experience. This system can be applied in related fields such as access control security, various electronic appliances. Replacing manpower through this occlusion resistant face detection and recognition system not only improves accuracy, reduces manpower resources but also enhance convenience in life.