Cross-Spectral Face Recognition With Low Image Quality

Face recognition refers to the technology that automatically identifies or verifies the identity of subjects from images or videos using computers.Due to its low cost and non-intrusive nature,face recognition has become one of the biometric modalities that attract most attentions.As the ongoing boosting of performance and improvement,face recognition is becoming more and more pervasive in every aspect of life and society.However,there are still many difficulties of the face recognition technology at present,for instance,face recognition at nighttime or in harsh atmospheric environments.Such a scenario requires the usage of the infrared waveband,which involves the very challenge of cross-spectral face recognition.This problem is also usually coupled with quality disparity between the heterogeneous imagery: high quality visible light images and low quality IR images resulting poor recognition performance.Another difficulty is face occlusion,which also leads to a sharp decline in the face recognition system.A solution to face occlusion is to use the periocular area as the biometric modality.In this thesis,to address the problems above,we have conducted some innovative work on the topic of cross-spectral face and periocular recognition from the following perspectives:(1)A long-distance cross-spectral face recognition method based on wavelet image fusion is proposed.This method fuses low-quality and long-range infrared face images using wavelet image fusion to improve its quality,and therefore reduces the gap between the heterogeneous images.To justify our method,we first use a sharpness measurement tool ASM to estimate the quality of the fused images.The results show a significant increase in the quality of the fused images.We then uses the operator Gabor+WLD+LBP+GLBP to carry out face verification experiments,where the EER and recognition accuracy are both improved on a cross-spectral face dataset TINDERS.(2)To address the problem of performance drop due to face occlusion,this thesis uses the periocular area as the biometric modality and studies the problem of cross-spectral periocular recognition.Since the periocular area cropped from the face usually has low resolution,the lack of valid information will affect the final recognition performance.In this thesis,a deep learning based super-resolution technique is utilized to reconstruct and enlarge low-resolution periocular images.The results verifies the benefit of using such a super-resolution technique for the cross-spectral periocular recognition problem.In addition,we compare it with a traditional interpolation-based method and the results show that the super-resolution method is superior to the interpolation method in terms of PSNR and SSIM.(3)This thesis also studies the feasibility of using convolutional networks for cross-spectral periocular recognition.Traditionally,handcrafted methods are used for feature extraction in periocular recognition whist we in this thesis design convolutional networks to deal with this problem.We first expand the cross-spectral periocular dataset,and construct a convolutional neural network model,Periocular Net.Experiments show that the proposed model is better than traditional methods with lower EER and higher GAR.In addition,we design two different networks Periocular Net-11 and Periocular Net-15 to justify the usage of periocular super-resolution.Experiments show that Periocular Net-15,for the case of using superresolution has higher performance than Periocular Net-11,for the case of none superresolution.