Research On Joint Iris Segmentation And Localization Algorithm Based On Deep Learning

Iris feature information has many excellent characteristics such as security,high anti-counterfeiting,stability,and inviolability due to its complex texture structure(such as stripes,crypts,crowns,spots,etc.).With increasing attention to security issues,iris recognition has gradually become another technology that has attracted the attention of academia and industry after face recognition and fingerprint recognition.Iris recognition technology can often be seen in security products,access control measures,confidential anticounterfeiting identity verification,etc.Generally speaking,four sub-steps,namely image acquisition,image preprocessing,iris feature extraction and recognition together constitute a complete iris recognition system.As part of iris preprocessing,iris segmentation defines the image region for subsequent iris feature extraction and matching.It directly affects the overall iris recognition performance and is a critical step in the system.Despite many advances has been made by the convolutional neural network-based iris segmentation methods in recent years,many challenges remain.First,most existing deep learning-based iris segmentation methods only focus on generating iris masks,which means that the results obtained by such single-task segmentation algorithms cannot smoothly connect with subsequent rubbersheet-transform to yield normalized iris images for recognition.On the other hand,existing iris segmentation methods rely heavily on stacking multiple convolutional layers to expand the receptive field and lack the ability to explicitly capture long-distance information,while the fixed size and channel design of the convolutions also make it difficult to dynamically adapt to the input.Finally,most current segmentation networks cannot be applied to real-world scenarios such as mobile devices due to the large number of parameters and computational cost.Based on the above observations,this paper proposes two deep learning-based multi-task iris segmentation algorithms HTU-Net and Li Se Net to improve the existing problems.Unlike most previous works,the methods proposed in this paper can simultaneously obtain the iris segmentation mask and parameterized pupil and iris outer boundaries through a multi-task learning framework,thus achieving complete joint iris segmentation and iris localization.(1)In response to the problem that existing algorithms rely heavily on stacked convolutional layers to expand the receptive field and ignore the problem of capturing long-distance information,the first algorithm HTU-Net explores the application of Transformer in iris segmentation,and we propose a hybrid encoder that combines convolutional neural networks and Transformers.The encoder uses convolutional layers to extract local intensity features and Transformers to capture long-range correlation information to joint model local correlation and long-distance dependencies.In the decoding stage,HTU-Net adopts a novel multi-head dilated attention.We use multi-scale context information through the gating mechanism to refine the features extracted by the encoder at each level,thus removing irrelevant noise in the features,and making the network focus more on important areas.Inspired by the consistent class-level features of the iris,HTU-Net designed a pyramid center-aware module to capture the global structural features of the iris to further improve the model performance.(2)The second algorithm designs a lightweight and efficient segmentation model,Li Se Net.Li Se Net first proposes a multi-scale concatenate block,which cascades convolution kernels of multiple sizes in a densely connected manner,gradually reduces the dimension of feature maps,and finally uses their aggregation for image representation.Based on the multi-scale concatenate block,this paper develops a two-stage refinement encoder to aggregate discriminative features through sub-network feature reuse and sub-stage feature re-evaluation to obtain sufficient receptive fields and enhance model learning ability.To utilize object contextual information more effectively,this paper further employs grouped spatial attention to emphasize important features and suppress irrelevant noise in the decoder.Extensive experiments are performed on three widely used iris databases,UBIRIS.v2,MICHE.I,and CASIA.v4-distance to verify the effectiveness of the proposed method.HTU-Net achieves the best or competitive performance in segmentation and localization without any complicated post-processing,and Li Se Net achieves a very competitive performance with only 2.2M parameters,being14 times smaller than the previous best method,which both proves the efficiency and robustness of the two iris segmentation models proposed in this paper.