Research And System Implementation Of Identification Algorithm Based On Iris And Periocular Images

In recent years,iris recognition is considered to be the most potential biometric technology because of its uniqueness,stability,anti-counterfeiting and non-invasive characteristics.Iris recognition mainly includes three steps: image acquisition,image preprocessing,image feature extraction and matching.In the feature extraction stage,compared with filter-based manual feature extraction,deep learning-based methods can automatically extract features and exhibit superior matching performance.Although the research on iris recognition based on deep learning,especially convolutional neural networks,has made great progress,the accuracy and generalization performance of the network are still lower than the expectations of biometric recognition technology,and the large number of network parameters requires more computing resources.To solve the above problems,this thesis mainly considers the network accuracy,generalization performance and parameter quantity to further optimize the network to achieve more effective iris recognition.In addition,for unconstrained environments such as surveillance and access control systems where clear iris images cannot be obtained,we choose to identify the easily obtained periocular images.Therefore,this thesis mainly studies the iris recognition and periocular recognition algorithms based on lightweight convolutional neural networks.The specific research contents are as follows:(1)Aiming at the rich detailed features in iris images,this thesis proposes a deep learning network IrisAttenNet based on attention mechanism to achieve targeted feature extraction.The network uses the Squeeze and Excitation(SE)module to apply attention to the feature channel,that is,it adaptively assigns different weights to the feature channel during the training process.Features with more information get larger weights,and features with less information get smaller weights to suppress the expression of redundant features.The proposed network exhibits good classification performance on public datasets.(2)Compared with the iris image,the periocular image contains more feature information,so this thesis adopts the network structure AttenMidNet based on the combination of attention mechanism and mid-level features to comprehensively extract the periocular image features.The feature channels of each layer in the network are first assigned corresponding weights through the SE module,and then connected to obtain the final feature vector after the global average pooling process,which not only enhances the feature representation capability of the network,but also further reduces the amount of parameters.The open-set validation results of the network model on the public datasets of near-infrared light and visible light demonstrate the high efficiency of the network in periocular image recognition,that is,the network parameters are greatly reduced while showing superior performance.(3)Considering that the current campus access control system based on face recognition has certain security risks in the context of the new crown epidemic,this thesis builds a real-time periocular recognition verification system on the Jeston Nano platform based on the proposed periocular recognition algorithm.The system can not only realize the real-time detection and verification of the periocular image,but also can improve the recognition and verification speed of the system during the peak flow of people through auxiliary identification of the mobile phone terminal.The system includes an infrared temperature measurement sensor module,which is conducive to the prevention and control of the epidemic.(4)To further verify the performance of the proposed periocular recognition system,this thesis simultaneously conducts matching verification experiments on the self-collected periocular image dataset.The dataset contains 822 periocular images of40 subjects with multiple complicating factors such as uneven lighting,hair and glasses occlusion in real-world application environments.The experimental results show that the proposed network AttenMidNet can obtain a lower EER of 17.545%compared to other networks with EER of 20.123%,19.015% and 18.568% on this dataset.