Journal of Applied Optics, Volume. 43, Issue 4, 669(2022)
Fine-grained target classification method based on deep clustering
Figures & Tables(8)
Fig. 2. Implementation block diagram of fine-grained target classification
Table 1. Flow chart of K-means clustering method
View tableView in ArticleTable 1. Flow chart of K-means clustering method
K-means聚类方法伪代码 输入：表征矩阵集合F=[f1，f2,…,fx]、聚类簇数K及迭代次数N 过程： 1）选取K个表征矩阵作为初始的聚类质心Ki 2）对于集合中的每一个元素，即每一个图像的表征矩阵，计算与每一个聚类质心的距离d，并将其划分到该类别中； 3）对每一类重新计算新的质心Ki’，如果d（Ki,Ki’）>δ，更新Ki；否则保持不变； 4）满足迭代次数或质心不再更新，则输出结果，否则迭代步骤2、3。 输出：分类结果和质心
Table 2. Classification accuracy rates using different classification methods
View tableView in ArticleTable 2. Classification accuracy rates using different classification methods
方法 数据集 CIFAR-10 CIFAR-100细粒度 VGG-16 93.60% 67.55% VGG-16+cluster 90.42% 64.13% ResNet-34 94.26% 70.05% ResNet-34+Cluster 92.71% 67.14%
Table 3. Influence of different numbers of cluster on classification results
View tableView in ArticleTable 3. Influence of different numbers of cluster on classification results
簇数 标签 准确率 20 粗粒度标签 86.32% 100 细粒度标签 64.13%
Table 4. Influence of different clustering grades on classification results
View tableView in ArticleTable 4. Influence of different clustering grades on classification results
一级簇数 准确率 二级簇数 准确率 20 86.32% 5 68.29%
Tools
Get Citation
Copy Citation Text
Hongqiang LU, Junlin Wang, Ya’nan WANG, Xuezhi AN, Xinchao NING, kun QIAN. Fine-grained target classification method based on deep clustering[J]. Journal of Applied Optics, 2022, 43(4): 669
Paper Information
Category: OE INFORMATION ACQUISITION AND PROCESSING
Received: May. 20, 2022
Accepted: --
Published Online: Aug. 10, 2022
The Author Email: Junlin Wang (aogetuya@qq.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20