摘要
深度学习是目前图像分类的主流方法之一,其重视感受野内的局部信息,却忽略了类别的先验拓扑结构信息。本文提出了一种新的图像分类方法,即Key-D-Graph,这是基于关键点的图对比网络方法,在识别图像类别时可以显式地考虑拓扑先验结构。具体地,图像分类需要2个步骤,第一步是基于关键点构建图像的图表达,即采用深度学习方法识别图像中目标类别的可能关键点,并采用关键点坐标生成图像的拓扑图表达;第二步基于关键点的图像图表达建立图对比网络,以估计待识别图与目标类别之间的结构差异,实现类别判断,该步骤利用了物体的拓扑先验结构信息,实现了基于图像全局结构信息的物体识别。特别的,Key-D-Graph的中间输出结果为类别关键点,具有语义可解释性,便于在实际应用中对算法逐步分析调试。实验结果表明,提出的方法可在效率和精度上超过主流方法,且通过消融实验分析验证了拓扑结构在分类中的作用机制和有效性。
At present, deep learning is one of the mainstream methods for image classification. It focuses more on local features in the receptive field than the prior information of topological structure of the category. In this paper, We propose a Keypoint-based Discriminator Graph neural network(Key-D-Graph) for image binary classification method,which is a graph comparison network method based on key points. It explicitly introduces the topology prior structure when identifying image categories. The method contains two main steps. The first step is to build the graph representation of an image with the keypoints, that is, identifying possible key points of the target category in the image by a deep learning method, and then using the coordinates of the key points to generate the topological representation of the image.The second step is to build a graph contrastive network based on the image representation of key points, so as to estimate the structural difference between the graph to be identified and the object graph, realizing object discrimination. In this step, the topological prior structure information of the object is used to realize object recognition based on the global structure information of the image. Especially, the intermediate output results of Key-D-Graph are the key points of categories containing explicit semantic information, which facilitates analysis and debugging of the algorithm step by step in practical application. Contrast experiments show that the proposed method outperforms the mainstream methods both in efficiency and precision. And the mechanism and effectiveness of topological structure in classification are verified by the ablation experiments.
作者
卢毅
陈亚冉
赵冬斌
刘暴
来志超
王超楠
LU Yi;CHEN Yaran;ZHAO Dongbin;LIU Bao;LAI Zhichao;WANG Chaonan(The State Key Laboratory of Management and Control for Complex Systems,Institute of Automation,Chinese Academy of Sciences,Beijng 100190,China;School of Information Science and Engineering,Shan Dong Normal University,Ji’nan 250358,China;Chinese Academy of Medical Sciences,Peking Union Medical College Hospital,Beijing 100730,China;Department of Vascular Surgery,Peking Union Medical College Hospital,Beijing 100730,China)
出处
《智能系统学报》
CSCD
北大核心
2023年第1期36-46,共11页
CAAI Transactions on Intelligent Systems
基金
国家重点研发计划项目(2019YFB1311700)
山东省自然科学基金青年项目(ZR2021QF085)
国家自然基金青年基金项目(62006223,62006226)
中国科学院战略重点研究项目(XDA27030400)
中央级公益性科研院所基本科研业务费临床与转化医学研究基金项目(2019XK320004)
中国医学科学院医学与健康科技创新工程医学人工智能科技先导专项(2018-12MAI-004)
中央高校基本科研业务费重点项目(3332020009)。
关键词
关键点识别
图拓扑结构
图像分类
图对比学习
距离学习
图神经网络
暹罗网络
图分类
keypoint detection
graph topological structure
image classification
graph contrastive learning
metric learning
graph neural network
siamese network
graph classification
