| Iris recognition has the advantages of high stability,good anticounterfeiting,non-contact,and high image uniqueness,and its technology and products have been widely used in several fields and industries.To ensure that high quality iris images can be captured,iris recognition systems usually require a high level of user cooperation.However,the varying degrees of user cooperation during the image acquisition process may introduce additional interference,resulting in the effective area of the captured iris image being obscured,causing a sharp degradation in the performance of the iris recognition system.Therefore,this paper conducts a study of iris inpainting methods to address the problem of iris occlusion causing degradation of iris recognition performance.The main research contents are as follows:(1)To address the problem of iris occlusion caused by frames,illumination,blinking,distance,and motion during iris recognition,a twoway iris inpainting model guided by a regional attention mechanism is proposed to effectively combine CNN-based local features and Transformer-based global representation to enhance representation learning.The encoder of the iris inpainting model consists of a CNN branch and a Transformer branch,and the Transformer follows the design of ViT,which is an integrated combination of local convolution blocks,selfattentive modules,and multi-layer perceptrons.In addition,for the task of iris inpainting,a region attention mechanism is proposed that will allow the network to choose to focus on more fine-grained non-obscured region features in the encoder or more semantically generative obscured region features in the decoder,depending on how differently the two features contain information about the obscured region.The addition of constraints that favor iris identity recognition allows the iris image after inpainting to improve the recognition accuracy of the iris recognition system.Extensive experiments are conducted on the CASIA iris database,and the results verify that the method in this paper effectively improves the performance of iris segmentation and iris recognition.(2)To address the problem of poor generalization performance of existing inpainting methods in the practical application scenarios of iris recognition systems with various types of iris occlusions,this paper proposes a method for iris inpainting based on the inverse fusion of diffusion models,which iteratively modifies the intermediate variables of the diffusion model generation process based on a given occlusion image until the completion of iris image inpainting.Without any additional learning,the unconditional diffusion model can be used to process any type of occlusion images.Also,an inverse fusion approach is introduced to further enhance the consistency between the content generated by the diffusion model and the existing pixels of the occluded image.Extensive experiments are conducted on the CASIA iris database and more complex and diverse occlusion types are used,which are more generalizable for image occlusion types and generate higher quality images than the previous methods,and are more suitable for realistic applications of iris recognition.(3)To address the evaluation problem of iris inpainting methods,this paper designs iris segmentation and iris recognition experiments in addition to using generic image inpainting evaluation metrics to demonstrate the effectiveness of this paper’s methods for improving iris segmentation and iris recognition performance. |
