摘要
以赤足足迹压力图像为研究对象,采集了40人的5230幅赤足足迹压力图像,在具有较低存储消耗的哈希算法基础上,结合深度学习方法设计了一种深度中心匹配哈希(DCMH)网络实现足迹的检索.该网络首先根据足迹压力图像的特点构建深度特征融合模块,提取反映足迹形态结构的全局特征和压力分布的局部特征,并将两种特征进行融合;然后在哈希编码模块通过全连接层将融合后的特征映射为1024维特征向量,并通过哈希层生成哈希码;在网络优化过程中通过构建深度中心匹配损失函数从而减小哈希码与对应哈希中心之间的距离.深度中心匹配损失函数通过伯努利分布生成哈希中心,设计对数中心损失函数减小同类足迹压力图像数据哈希码与哈希中心的距离,并设计相似性损失函数作为正则化项约束每个批次数据间的相似性关系.通过在40人的赤足足迹压力图像数据上进行图像检索实验,本文算法检索结果的mAP可以达到0.99,优于传统的哈希算法及常用的深度哈希算法,为足迹的进一步的现场应用提供技术支撑.
Taking the barefoot footprint pressure image as the research object,5230 barefoot pressure images of 40 people were collected.On the basis of hash algorithm with lower storage consumption,a deep center matching hash(DCMH)network was designed to retrieve footprints based on deep learning method.Firstly,according to the characteristics of footprint pressure image,the deep feature fusion module was constructed.The global features reflecting footprint morphological structure and local features of pressure distribution were extracted,and the two types of features were fused.Then,in the hash coding module,the fused features were mapped into 1024 dimensional feature vectors through the full connection layer,and the hash code were generated through the hash layer.In the process of network optimization,the distance between hash code and corresponding hash center was reduced by constructing deep center matching loss function.The deep center matching loss function generated the hash center through Bernoulli distribution.The logarithmic center loss function was designed to reduce the distance between the hash code and the hash center of similar footprint pressure image data.The similarity loss function was designed as a regularization term to constrain the similarity relationship between each batch of data.Through the image retrieval experiment on the barefoot footprint pressure image data of 40 people,the mAP of the retrieval result of this algorithm can reach 0.99,which is better than the traditional hash algorithm and the commonly used deep hash algorithm.This method can provide technical support for further field application of footprint.
作者
鲍文霞
胡伟
王年
杨先军
BAO Wenxia;HU Wei;WANG Nian;YANG Xianjun(School of Electronic and Information Engineering,Anhui University,Hefei 230601,China;Hefei Institutes of Physical Science,Chinese Academy of Sciences,Hefei 230031,China)
出处
《华中科技大学学报(自然科学版)》
EI
CAS
CSCD
北大核心
2023年第9期81-87,共7页
Journal of Huazhong University of Science and Technology(Natural Science Edition)
基金
国家重点研发计划资助项目(2020YFF0303803,2018YFC0807302)
安徽省重点研发计划资助项目(2022k07020006)
安徽高校自然科学研究重大项目(KJ2021ZD0004).