摘要
针对刑侦图像分类问题,提出一种基于多核支持向量机的多示例学习(MIL)算法。首先,该方法采用金字塔网格划分法对刑侦图像进行分块,再将每幅图像作为一个多示例包,每个子块的底层视觉特征作为包中的示例,将刑侦图像分类问题转化为MIL问题;然后,采用K-means双重聚类方法对所有多示例包进行聚类生成聚类中心并定义为视觉字,再把视觉字的集合构造成视觉投影空间;最后,通过设计的非线性投影函数将每个包映射为视觉投影空间中的一个点,则MIL问题被转化为一个标准的有监督学习问题,并采用多核支持向量机(MKSVM)来训练刑侦图像分类器。基于真实刑侦图像库的对比实验表明,所提方法具有较好的鲁棒性,且分类精度高于其他方法。
A multiple instance learning （MIL） algorithm based on multiple-kernels support vector machine is proposed to solve the problem of criminal investigation classification.Firstly,this method uses the pyramid grid partition to block the criminal investigation image,and then each image as a multiple instance bag,each sub-block of the low-level visual features as an instance of the bag,the criminal investigation image classification problem can be converted to MIL problems;Secondly,the K-means double clustering method is used to cluster all multiple instance bags to form a cluster center and is defined as a visual word,and then the set of visual words is constructed into a visual projection space;Finally,the MIL problem is transformed into a standard supervised learning problem by using the designed nonlinear projection function to map each bag into a point in the visual projection space.The multiple-kernels support vector machine （MKSVM） is used to train the criminal investigation image classifier.Based on the real investigation of the criminal investigation library,the proposed method has good robustness and classification accuracy higher than other methods.
作者
吴倩
李大湘
刘颖
WU Qian;LI Daxiang;LIU Ying(School of Communication and Information Engineering, Xi＇ an University of Posts and Telecommunications, Xi＇an 710121, China;Ministry of Public Security Key Laboratory of Electronic Information Application Technology for Scene Investigation, Xi＇ an 710121, China)
出处
《电视技术》
北大核心
2017年第11期59-63,共5页
Video Engineering
基金
陕西省国际合作交流项目(2017KW-013)
陕西省教育厅科学研究项目(15JK1660)
公安部科技强警基础工作专项项目(2014GABJC022)
关键词
刑侦图像分类
多示例学习
多核支持向量机
criminal investigation image classification
multiple instance learning
multiple-kernels support vector machine
