摘要
为消减预训练模型抽取的深度卷积特征冗余影响图像检索准确率,提出一种深度卷积聚合特征提取算法。该算法具体包含筛选、聚合和池化三个步骤。依据香农信息熵理论,提出一种基于熵的策略用于筛选出感兴趣卷积描述子。之后利用洪泛算法将感兴趣卷积描述子聚合为目标掩码图。最后基于目标掩码图区域,将卷积特征图中对应区域进行池化连接。通过在公用的图像数据集上定性定量的实验评价,表明所提出算法显著消减冗余,同时实现更高图像检索准确率。
To reduce the redundancy of the deep convolution features extracted from the pre-trained model affects the image retrieval accuracy,a deep convolution aggregation feature extraction algorithm is proposed.The algorithm consists of three steps:filtering,aggregation and pooling.According to Shannon's information entropy theory,an entropy-based strategy is proposed to select the convolutional descriptors of interest.Subsequently,the flooding algorithm is used to aggregate the convolutional descriptors of interest into an object mask map.Finally,based on the region of the object mask map,the corresponding region in the convolution feature map is encoded by max pooling and sum pooling.Compared with other methods on the public image datasets,the results show that the proposed algorithm has reduced redundancy obviously,and has achieved the higher image retrieval accuracy.
作者
冯庆贺
聂广华
刘荣升
迟明路
王元利
高雅昆
张建霞
FENG Qinghe;NIE Guanghua;LIU Rongsheng;CHI Minglu;WANG Yuanli;GAO Yakun;ZHANG Jianxia(School of Intelligent Engineering,Henan Institute of Technology,Xinxiang 453003,China;School of Electrical Engineering and Automation,Henan Institute of Technology,Xinxiang 453003,China)
出处
《河南工学院学报》
CAS
2022年第3期30-34,共5页
Journal of Henan Institute of Technology
基金
河南省重点研发与推广专项(科技攻关)项目(202102210084,212102310119)
河南工学院高层次人才科研启动基金(KQ1863,KQ2011,KQ2001)
河南省高等学校重点科研项目(21B413002)
教育部产学合作协同育人项目(202101187010,202102120046)
河南工学院教育教学改革研究与实践项目(2021-YB023)
河南省大学生创新创业训练计划项目(202111329014,202111329015)。
关键词
图像检索
深度卷积特征
卷积描述子
冗余
image retrieval
deep convolutional feature
convolutional descriptors
redundancy
