Research On Iris Segmentation Algorithm Based On Deep Learning Network Model

Human life is closely related to the rapid development of technology,and authentication and identification have become an essential component.However,as technology continues to advance,people find that traditional authentication methods,such as password login,no longer meet their requirements.In this context,iris recognition technology has received wide attention from scholars in various fields because of its high accuracy,uniqueness,stability,ease of use and anti-counterfeiting features.Iris recognition has been gradually applied to different security-related areas of our lives,such as smart lock opening,social networking,and mobile payment.Compared to fingerprint and face recognition,iris recognition is becoming one of the most widely accepted identification technologies.The execution steps of an iris recognition system include four steps: data acquisition,pre-processing,feature extraction,and template matching.However,during the pre-processing of iris images,the external environment,the performance of the acquisition device,and the level of user cooperation can have a significant impact on the iris image quality.For example,external lighting may distort the iris image,poorly performing acquisition devices capture a lower resolution iris image,and user movement during the photo taking process can cause motion blur in the iris image.These unqualified images can affect the entire subsequent recognition process,which in turn reduces the accuracy of the iris recognition system.For iris recognition systems,an iris segmentation algorithm that can accurately distinguish the iris region from the background region is critical.However,traditional iris segmentation algorithms are complex,often require human intervention,and are ineffective when dealing with noisy iris images.With the continuous innovation and optimization of deep learning techniques,scholars have found that the use of neural network techniques can effectively solve the problems of traditional iris segmentation algorithms,and also neural network techniques can largely improve the application performance of iris segmentation algorithms.In this paper,we propose an efficient and robust iris segmentation network model,FRTIris U-Net,to address the problems of low accuracy and low robustness of traditional iris segmentation algorithms and the need for extensive manual operations.FRTIris U-Net improves on the classical semantic segmentation network,U-Net,and the main work includes the following three aspects.1)Due to the inherent limitation of CNN structure,it has limited perceptual field.Although the perceptual field can be increased by stacking CNN structures or using methods such as null convolution,this also introduces new problems,which limit the final results.Therefore,in this paper,a fusion encoder of Res Net-50 and Transformer is used on FRTIris U-Net,so that the feature information extracted by FRTIris U-Net in the encoder part contains both local and global feature information.2)The inclusion of attention mechanism can enable the network to focus on the features in the key regions and ignore the features in the background regions other than the key regions.Therefore,this paper incorporates the CBAM attention mechanism at the jump connection of FRTIris U-Net,which enables the network to focus more on the extraction of key features and suppress unnecessary features.3)In order to enhance the network’s extraction of multi-scale feature information,this paper embeds an improved Dense-ASPP module at the bottleneck between the encoder and decoder of FRTIris U-Net.The feature image is reduced to the original image size by multiple upsampling operations after this module,and finally a binary iris mask image is generated.In this paper,three iris datasets,JLU-6.0,CASIA-Interval-V4,and UBIRIS.v2,are selected as the experimental datasets for this paper,and a large number of experiments are conducted on this dataset.By evaluating the experimental results,it is found that the Acc of the FRTIris U-Net network proposed in this paper on the above three iris datasets are 98.84%,99.21%,99.38%,and m PAs of 0.9772,0.9826,and0.9842,respectively,outperforming most of the iris segmentation algorithms,proving that the FRTIris U-Net network has good efficiency and robustness.