摘要
为了在刑侦领域中为案情分析提供重要依据,基于足迹分析技术,对足迹图像数据,采用关系网络实现赤足足迹识别。首先基于Otsu算法和左、右足的相似性,对足迹图像分别进行图像去噪和数据增强;然后将图像送入关系网络进行训练,训练集包括63个对象的赤足数据,测试集包括另外50个对象的赤足数据;最后对四种训练方式下的模型识别率进行了对比和分析。在关系网络的训练上,采用了多任务等四种方式的训练。实验表明:基于自编码器和多任务学习的多任务方式在四种训练方式中表现最好,在左、右足迹识别上识别率分别达到了88.5%和88.3%。
In order to provide important basis for case analysis in the field of criminal investigation,based on footprint analysis technology,aiming at footprint image data,Relation Network is used to realize bare footprint recognition.Firstly,based on the Otsu algorithm and the similarity between left and right foot,image denoising and data augmentation are carried out respectively.Then,the image is trained in Relation Network.The training set included bare footprint data of 63 objects.And the data of another 50 objects is used as the test set.Finally,the recognition rate of the models under four training methods are compared and analyzed.In the training of Relation Network,including multi-task training,four training methods are adopted respectively.Experiments show that the multi-task method based on autoencoder and multi-task learning is the best of all the training methods,and the recognition rates on left and right footprints are 88.5%and 88.3%,respectively.
作者
王鹏鹏
吴洛天
汪曙光
张艳
鲁玺龙
WANG Pengpeng;WU Luotian;WANG Shuguang;ZHANG Yan;LU Xilong(School of Electronics and Information Engineering,Anhui University,Hefei 230601,China;Tsinghua University Hefei Institute for Public Safety Research,Hefei 230601,China;Institute of Forensic Science,Ministry of Public Security,Beijing 100038,China)
出处
《传感器与微系统》
CSCD
北大核心
2021年第4期126-130,共5页
Transducer and Microsystem Technologies
基金
国家重点研发计划重点专项资助项目(2018YFC0807302)
科技强警基础工作专项项目(2018GABJC15)
公安部物证鉴定中心基本科研业务费专项项目(2018JB018)
痕迹重点实验室开放项目(2017FMKFKT08)。
关键词
赤足足迹
关系网络
足迹识别
自编码器
bare footprint
Relation Network
footprint recognition
autoencoder
