摘要
脱机掌纹(掌印扫描图)自动种属识别是刑侦领域中确定现场掌纹遗留部位的必要工作。由于脱机掌纹涉及个人隐私,获取受到限制,公开、可靠的脱机掌纹库较少,目前针对脱机掌纹的自动种属识别缺乏系统性研究。基于此,提出了一种融合改进注意力机制的残差网络3D IA-ResNet用于左右手脱机掌纹识别研究。首先在Kaggle猫狗数据集上验证所提出的改进注意力模块的效果,然后在自建脱机掌纹数据集上利用3D IA-ResNet实现左右手脱机掌纹识别。改进注意力模块3D IA在Kaggle猫狗识别数据集上的消融实验、模型稳定性实验中均表现优异,且融合改进注意力模块的残差网络3D IA-ResNet在左右手脱机掌纹识别数据集上精确度、召回率、F1值高达96.6%、95.7%、96.1%,相比基准模型提升0.5%、0.7%、0.6%。3D IA-ResNet经过猫狗识别任务检验后,有效地实现了左右手脱机掌纹识别,填补了相关领域的研究空白,为复杂掌纹识别任务提供了必要的技术支持。
In the field of criminal investigation,automatic species identification for offline palmprint image(scanned image of palm print)is a necessary work to determine the location of the palm from which the palm print originates.Since offline palmprint involves personal privacy and limited access,there are few open and reliable offline palmprint databases.At present,there is a lack of systematic research on automatic species identification of offline palmprint.Based on this,3D IA-ResNet is proposed for left and right palmprint recognition.Firstly,the effect of the proposed improved attention module was verified on the Kaggle cat and dog dataset,and then the 3D IA-ResNet was used to realize left and right offline palmprint recognition on the self-built offline palm-print dataset.3D IA-ResNet performed well in ablation experiments and model stability experiments on Kaggle cat and dog recognition datasets,and the accuracy,recall rate and F1 value on left and right offline palmprint recognition datasets were as high as 96.6%,95.7%and 96.1%,which were 0.5%,0.7%and 0.6%higher than the baseline model.After the test of cat and dog recognition tasks,3D IA-ResNet effectively realizes the left and right hand offline palmprint recognition task,fills the research gap in related fields,and provides necessary technical support for complex palmprint recognition tasks.
作者
闫自强
左琦
张晓梅
YAN Ziqiang;ZUO Qi;ZHANG Xiaomei(College of Forensic Sciences,Criminal Investigation Police University of China,Shenyang 110035,China;Department of Criminal Science and Technology,Henan Police College,Zhengzhou 450046,China)
出处
《中国人民公安大学学报(自然科学版)》
2024年第3期19-25,共7页
Journal of People’s Public Security University of China(Science and Technology)
基金
辽宁省高等教育教学改革研究项目(辽教办[2021]254号-713)
关键词
脱机掌纹
种属识别
迁移学习
注意力机制
残差网络
深度学习
offline palmprint
species identification
transfer learning
attention mechanism
residual network
deep learning
