SWU-NIRPV: A Near-infrared Pose Variation Face Database And Pose Invariant Face Recognition

Active near-infrared(NIR)illumination technology can achieve imaging in all black environment.It is widely used in security monitoring.However,in the near-infrared surveillance video,the human face often presents different pose angles.For face recognition under visible light,pose variations will bring about a significant decrease in recognition rate.Although qualitative analysis shows that mid-wave infrared and long-wave infrared face recognition are less affected by pose variations,the quantitative analysis of pose variations on near infrared face recognition is hard to be carried out,due to the lack of available data sets.In addition,most of the existing near-infrared face data sets are based on frontal face,and there is no public near-infrared face data set including multiple angle pose variations.The quantitative effect of pose changing on near-infrared face recognition is unknown.The verification of pose variation insensitive near-infrared face recognition algorithm is even more difficult.Based on the problems above,we develop a near-infrared face database with multiple angle pose variations,named as Southwest University NIR Pose Variation database(SWU-NIRPV).How pose variations on near-infrared influences face recognition is discussed in this paper,and neural networks are designed to achieve pose variation insensitive near-infrared face recognizing.The main contributions ofthis paper are as follows:(1)A near infrared face database SWU-NIRPV with 57 pose variation angles under dark conditions is established,which covers the near infrared face data of 0°(front face)to ± 90°(side face)on the left and right sides at an interval of 10°.At the same time,the visual up,visual front and visual down subsets are further developed at each yaw angle.(2)The influence mechanism of pose variation on near infrared face recognition is discussed and analyzed from single and cross point of view.Near infrared face recognition is also affected by the pose variation.Only 10° difference of yaw angle is enough to make the near infrared face recognition algorithm greatly reduced.The face data of 30° yaw angle at the cross recognition is conducive to the recognition of small yaw angles(10° to 40°),while the face data of 60° yaw angle is conducive to the near infrared face recognition of large yaw angles(50° to 80°).In addition,the pitch angle change is also an important factor in near infrared face recognition under the condition of pose variations.(3)A pose variation insensitive near infrared face recognition approach is proposed.CNN-Capsule Networks(CNN-Caps)was established by capsule network module,and LBP-Softmax,CNN-Gray,Resnet50 and CNN-Caps were verified on the basis of the conclusions in(2).The recognition effect of the four models based on the training set of 0° + 90° + 30° + 60° under the cross angle is significantly improved,and the average recognition rate of Resnet50 is 98.10% in all angles.