摘要
针对陶俑文物的图像理解问题,陶俑分类可为其提供有价值的信息,该文提出了一种融合深度特征的多示例学习(MIL)方法用于陶俑图像分类。首先,对陶俑图像进行分割,提取出分割区域的手工特征(包括尺度不变特征变换和形态特征)和卷积神经网络特征;接着,采用联合字典学习获取多示例学习的多概念点,并使用多核将深度学习特征与传统手工特征融合到多示例学习框架;最后,利用直推式支持向量机进行分类。在陶俑图像集和MIL数据集上的实验结果表明,该文方法是有效的,且相较其他深度和非深度MIL算法具有更高的分类准确度。
Focusing on the image understanding problem of ancient pottery figurine,pottery figurine classification can provide valuable information.In this paper,a multiple instance learning(MIL)with deep feature fusion method is proposed to classify pottery figurine images.Firstly,the pottery figurine image is segmented into regions and the hand-crafted features(scale invariant feature transformation and morphological features)and convolutional neural network features is extracted.And then the joint dictionary learning is adoped to acquire the multiple target concepts of MIL and fuse the deep and hand-crafted features into multiple instance learning framework by using multiple kernel learning.Finally,transductive support vector machine is used to classify the pottery figurine images.Experimental results on pottery figurine image sets and MIL datasets show that the proposed method is effective and has higher classification accuracy comparing with the state-of-the-art algorithms including deep and other non-deep MIL algorithms.
作者
温超
屈健
李展
WEN Chao;QU Jian;LI Zhan(School of Arts,Northwest University,Xi′an 710127,China;School of Information Science and Technology,Northwest University,Xi′an 710127,China)
出处
《西北大学学报（自然科学版）》
CAS
CSCD
北大核心
2019年第6期895-902,共8页
Journal of Northwest University（Natural Science Edition）
基金
国家重点研发计划课题(2017YFB1402103)
教育部人文社会科学研究项目(17YJCZH186)
陕西省自然科学基金项目(2013JQ8022)
陕西省教育厅自然科学基金项目(2013JK1181,2014JK1724)
关键词
陶俑图像
多示例学习
卷积神经网络
多核
直推式支持向量机
pottery figurine image
multiple instance learning
convolutional neural network
multiple kernel
transductive support vector machine
