Research On Iris Location And Recognition Algorithm Based On Deep Neural Network Model

Iris recognition has attracted more and more attention from academic and engineering circles because of its uniqueness,stability,non-contact,high anticounterfeiting and reliability.With the deepening of iris recognition research,iris recognition products are gradually applied in different levels of security fields such as security systems,entrance guard systems,check-in systems,etc.In the field of biometric identification,iris recognition has gradually become the most accepted and recognized identity recognition technology after fingerprint recognition and face recognition.Iris recognition generally includes steps such as image acquisition,preprocessing,feature extraction,and recognition.Pre-processing includes iris quality evaluation,iris localization,iris normalization and enhancement.In the pre-processing process,limited by the user's cooperation and external environmental influences,the captured image will be subject to motion or defocus blur,external lighting and the user's eyelids will affect the quality of the iris image.With the development of deep learning technology,its application has involved many fields.In computer vision,such as classification tasks,clustering tasks,image segmentation tasks,etc.,it has caused comprehensive changes and greatly improved applications in various fields.Performance,semantic segmentation,as a hot research topic in the field of image segmentation,has been fully applied in the fields of medical images and autonomous driving.In terms of iris image preprocessing,this paper uses semantic segmentation technology to solve the problem of iris segmentation,and proposes an encoder-decoder network REDA-Net that incorporates a residual module and attention mechanism.In the the image encoding process,the residual module combines the features of the previous layer and the current features in the network downsampling process,that is,the shallow layers with rich spatial information and the deep features with semantic information are fused to improve the expression ability of the features,and The introduction of the attention mechanism is to guide the model to learn which features are important.It is easier to capture the boundary information in the decoding stage so that the mask generation is closer to the true boundary of the iris area.Then,by extracting the semantic information of the iris area,the iris area is separated from the iris image to generate an iris mask,and the mask information is used to locate the iris and judge the effectiveness of the iris area.As for iris recognition algorithm,traditional iris recognition algorithms have strict requirements on image quality or feature extraction tools need to adjust cross-database parameters,which cannot describe the distribution of data.The current trend is to learn the distribution of data in large amounts of data.In this paper,the iris recognition task is realized by building a deep neural network model,using a residual structure based on the ResNeXt network model,which uses a separation-transform-merge strategy similar to the Inception module.This grouping strategy can make the iris features be allocated to different subspaces to learn different feature representations.The features at different levels are more conducive to the extraction of rich iris texture information.Softmax loss and center loss are used during model training and the center loss aims to constrain the distance within the category.By calculating the Euclidean distance between the vectors of iris image to be recognized and a certained category of the iris dataset,which is produced by the last layer of the fully connect layers in the network.By camparing the distance with the threshold to determine they belong to the same category or not.In this paper,part of the data of JLU-V4.0 iris data set is compared with the existing semantic segmentation algorithm on the two indicators of mPA and mIoU,and the experimental analysis is carried out to prove the availability of the REDA-Net model proposed in this paper.In terms of recognition experiments,this paper compares the traditional iris recognition methods and an existing deep learning model methods on the JLU-V4.0 and JLU-V5.0 iris data sets.The experimental results show that the recognition methods proposed in this paper has more excellent Performance.