Research On Face Recognition Technology Based On NIR And VIS Image Synthesis

With the vigorous development of deep tearning,face recognition technology has been applied to all aspects of people's daily lives.The commonly used face recognition technology in the field of visible light has surpassed humans and developed to a saturation period.However,this does not mean that the technical problems of face recognition have been completely solved.In people's daily life,face recognition technology still faces many challenges.For example,in areas with poor lighting,most devices use near-infrared cameras to collect face images,but at the same time,the existing face database is basically VIS images.Therefore,the problem of NIR-VIS heterogenous face recognition is introduced,which requires the algorithm to not only handle the problems of different expression,lighting,occlusion and posture,but also overcome the differences between the NIR and VIS domain images.Under this background,this paper studies from three aspects:face detection,image fusion and face recognition,also designs and implements a heterogeneous face recognition system based on NIR-VIS images.The main work of this article is as follows:1.Aiming at the problem of face deformation when MTCNN face detection algorithm cuts a face,a new cropping method is proposed.The test results show that it is possible to effectively cut out the face with the same aspect ratio as the original image;2.In view of the poor performance of NIR-VIS image fusion,the CycleGAN algorithm is proposed for image fusion to realize the conversion between NIR and VIS domain images.The test results show that the NIR face image converted by CycleGAN algorithm fake_NIR and real_NIR in the database Face images are very similar;3.In view of poor performance of face recognition in the NIR domain,the Insightface algorithm is proposed for face recognition of fake_NIR images and real_NIR images.The accuracy rate on the CASIA NIR-VIS 2.0 dataset reaches 99.35%.The test results show that the method proposed by this paper is better than the mainstream NIR-VIS face recognition algorithm in the field of recognition accuracy.In summary,the research work in this paper has achieved certain consequents,but this does not mean that the problem of NIR-VIS heterogeneous face recognition has been completely solved.First,the CASIA NIR-VIS 2.0 database currently used by researchers is relatively small and does not include multiple NIR-VIS heterogeneous face recognition scenarios.Second,the real scene applied by the algorithm system is far more complicated than the experimental data.Therefore,NIR-VIS cross-modal face recognition still has room and need for further research.