Research On Iris Recognition Based On Double-channel Network Feature Learning

Recent developments in the computer vision field have generated renewed interest in biometrics.Iris recognition is considered to be the most accurate and reliable biometric identification method,so it has been applied in different fields,such as identification and authentication systems,intelligent key systems,digital forensics and border control,etc.Iris contains a large number of unique,constant and anti-counterfeiting features used in biometric recognition,such as complex textures and explicit structural information.In addition,iris features are stable and remain constant throughout a person's life.Therefore,iris texture plays a key role in biometric practice.Common iris recognition methods usually include the following steps: iris image acquisition,iris image pre-processing,iris segmentation,iris feature extraction,feature matching and iris recognition.Iris segmentation is a critical and challenging task in iris recognition system due to the unpredictability and irregularity of iris shape.In general,iris texture is more visible near the boundary of the pupil.If the boundary of the pupil region is not located correctly,a large amount of iris texture will be lost in the feature extraction stage.In other words,the performance and robustness of iris recognition methods largely depend on the accuracy of iris segmentation.In most cases,iris area occlusion(such as eyelids or eyelashes,glasses,poor illumination,motion blur,etc,.)can greatly affect the overall accuracy of iris segmentation algorithms.Previous studies have shown that the errors generated by the steps of iris segmentation are transferred to the subsequent stages of iris recognition.Therefore,iris boundary estimation is still a very important pre-processing stage to achieve high precision of the system.In this paper,we propose an effective integrated model based on deep learning for accurate iris segmentation and recognition.Double-channel network feature learning model involves different stages such as pre-processing,detection,segmentation and recognition.Firstly,the image is pre-processed to improve the quality of the input image,using Black Hat filtering,median filtering and gamma correction.Secondly,the Hough circle transformation model is used to locate the iris effectively.Finally,a multi-channel network model is used to identify and segment the region proposal network of iris.Specifically,the proposed multi-channel network contains multiple Convolutional Neural Network(CNN),each CNN represents a specific rotation direction,and all CNNs share the same weights.According to the output features of each CNN,multi-instance learning algorithm is used to extract the final rotation invariant features.In order to verify the results of the proposed model,experiments are carried out on several public reference data sets,such as CASIA-LRIS.The results of different indexes show that the proposed method is superior to the existing methods in rotation invariant feature extraction,especially in the case of small training samples.