摘要
随着大数据和硬件的快速发展,细粒度分类任务应运而生,其目的是对粗粒度的大类别进行子类分类。为利用类间细微差异,提出基于RPN(Region Proposal Network)与B-CNN(Bilinear CNN)的细粒度图像分类算法。利用OHEM(Online Hard Example Mine)筛选出对识别结果影响大的图像,防止过拟合;将筛选后的图像输入到由soft-nms(Soft Non Maximum Suppression)改进的RPN网络中,得到对象级标注的图像,同时减少假阴性概率;将带有对象级标注信息的图像输入到改进后的B-CNN中,改进后的B-CNN可以融合不同层特征并加强空间联系。实验结果表明,在CUB200-2011和Standford Dogs数据集平均识别精度分别达到85.50%和90.10%。
With the rapid development of big data and hardware, fine-grained classification has emerged. Its purpose is to classify the coarse-granted categories into subclasses. In order to use the subtle differences between similarities, we proposed a fine-granted classification algorithm based on RPN and B-CNN. The online hard example mine(OHEM) algorithm was used to screen out the images which had a great impact on the recognition results to prevent the over-fitting. Then, the selected image was input into the RPN network improved by soft non maximum suppression(soft-nms). The false negative probability was reduced, and the image with object-level annotation was obtained. The image with object-level annotation was input the improved B-CNN. The improved B-CNN could fuse features of different layers and enhanced their spatial connection. The experimental results demonstrate that the average recognition accuracy of CUB200-2011 and Stanford Dogs datasets is 85.50% and 90.10%.
作者
赵浩如
张永
刘国柱
Zhao Haoru;Zhang Yong;Liu Guozhu(College of Information Science and Technology, Qingdao University of Science and Technology, Qingdao 266000, Shandong, China)
出处
《计算机应用与软件》
北大核心
2019年第3期210-213,264,共5页
Computer Applications and Software
基金
国家自然科学基金项目(61472196
61672305
61702295)
山东省自然科学基金项目(ZR2014FM015)
关键词
细粒度分类
类间差异
双向卷积网络
非极大值抑制
特征融合
Fine-granted classification
Interclass difference
B-CNN Non-maximum suppression
Feature fused
