Research On Feature Optimization Methods In Palmprint Recognition

Palmprint identification is an important research topic in the field of biometric recognition,and it has significant application value in areas such as network security,criminal investigation,and judicial authentication.In existing work,palmprint feature representation often relies on the experience and knowledge of criminal investigation experts to extract lower level and limited number of image features around palmprint minutiae,while some work directly uses deep features.Although these features have achieved preferable results in existing work,their accuracy still needs to be further improved,and in special application scenarios,using only deep features lacks connection with underlying interpretable features.This article focuses on high-resolution palmprint images and examines the extraction and representation of ridge distance features and minutiae features in palmprint images.Deep learning methods are used to optimize the aforementioned features.The main contents of this article are as follows:(1)A method for estimating the palmprint ridge distance based on attention mechanism and residual network is proposed.The method uses manually labeled ridge distance values as input to identify palmprint image blocks,improving the measurement accuracy and the ability to measure fine-grained details,which are problems with traditional palmprint ridge distance estimation methods.Palmprint ridge distance is an important texture attribute of palmprint ridges.Currently,frequency domain methods are commonly used to estimate palmprint ridge distance,but existing frequency domain methods have high requirements for palmprint image quality and integrity and cannot meet the needs of high measurement accuracy and fine-grained measurement.To address these issues,this article designs an attention mechanism and residual network,organically introducing the attention module to enhance the network’s attention to ridge structural information and improve recognition accuracy.The loss function is also improved by adding a residual mechanism to avoid gradient disappearance issues.The dataset used in this study was manually labeled,and to address the problem of uneven sample distribution in the dataset,data augmentation was used to weaken the imbalance of category distributions and a loss function for imbalanced samples was designed to optimize the model.Experimental results show that this method improves recognition accuracy by 14.2% compared to frequency domain methods on the dataset and at least 5.1% compared to directly using some classic deep learning methods.(2)A method for identifying genuine palmprint minutiae based on a lightweight convolutional neural network is proposed.The method further optimizes the minutiae features extracted by existing methods,removes false minutiae,and improves the reliability of the minutiae set in the palmprint matching stage.Traditional minutiae extraction methods result in a large number of false minutiae in the set,which affects the accuracy of palmprint recognition.Existing methods for identifying genuine minutiae still need to be improved in terms of accuracy.To address this,this article proposes a method for identifying genuine palmprint minutiae based on a lightweight convolutional neural network.A lightweight network is designed to reduce the number of model parameters,and a channel attention mechanism is introduced to enhance the network’s attention to important feature channels.Additionally,a deep mutual learning method is used to promote mutual learning between networks and improve their focus on the center minutiae features of the image.The proposed method is experimentally validated on a manually labeled dataset.The results show that the proposed method outperforms existing methods in terms of precision and recall and performs better among deep learning methods.