Research On Technologies Of Video Sequence Based Iris Image Processing

Iris recognition is an important means of human identification,as well as an important branch of biometrics technology.Due to the high uniqueness and stability of iris texture structure,and the non-contact and non-invasive characteristics,the importance of iris recognition in authentication is becoming increasingly prominent.However,the quality of iris images decrease due to the interference of factors such as illumination variation,motion blur,reflection spot and eyelid occlusion,influencing the iris recognition accuracy.As a result,most perfect iris recognition can only be carried out under strict constraints,which affects the wide application of iris recognition technology.In this paper,we consider the problems of image blur and occlusion encountered in the process of iris recognition,and video sequence iris images are used instead of single ones,aiming at applying iris recognition under unconstrained conditions.The methods of iris image preprocessing,iris image feature extraction and recognition are studied.The research on this issue has great meaning both in theory and reality.We proposes an improved L0-regularized deblurring method for iris images.The sparsity of dark channel,as well as the sparsity of the image gradient and estimated kernel are regarded as the regularization constraints of blind deblurring.At the same time,the natural scene statistics based blur metric is added to the process of iris image deblurring based on video sequence,which improves the efficiency of deblurring,avoiding possible damage to the quality of images.An iris image fusion method based on Multiscale Transformation(MST)and area classification is proposed.After SIFT-based image registration,an improved MST and sparse representation based fusion method is applied to the whole area of images.And then,iris and non-iris regions are classified,the result of which is used for subsequent fusion processing.For the non-iris areas,an improved stiching method is applied,which can not only keep more useful texture,but also avoid the occurrence of artifacts.An Generative Adversarial Networks based iris data augmentation method is studied.Based on conditional Generative Adversarial Networks,the proposed iris data augmentation model take the normalized iris image as input,and random noise is added into thelow dimension feature representations.The discriminator network takes the original data,other data from the same class and the data generated by the generator as its inputs,and uses triplet loss in training.The data augmentation model improves the variability of the generated iris data,while maintaining the class consistency.In the Fully Convolutional Networks based iris feature extraction model,the augmented data and the original data are used together for model training,which improves the performance of iris recognition.