图像分块及惰性多示例学习鞋印图像识别  

Shoeprint image recognition algorithm based on image block and lazy multi-instance learning

作　　者：李大湘[1,2] 吴倩[1] 李娜 刘颖[1,2] 

机构地区：[1]西安邮电大学通信与信息工程学院,陕西西安710121 [2]电子信息现场勘验应用技术公安部重点实验室,陕西西安710121

出　　处：《西安邮电大学学报》2016年第1期59-62,共4页Journal of Xi’an University of Posts and Telecommunications

基　　金：公安部科技强警基础工作专项基金资助项目(2014GABJC022);陕西省自然科学基金资助项目(2013JM8031;2015KW-014);陕西省教育厅科学研究计划资助项目(15JK1660;15JK1661);中国博士后科研基金资助项目(2013M542386)

摘　　要：结合图像分块与惰性多示例学习(MIL)给出一种鞋印识别新算法。将整个鞋印图像当作包,根据脚底生物特征比例,采用均匀网格分块的方法将鞋印图像分成15个子块,并提取每个子块的纹理与形状特征,当作包中的示例,将鞋印图像识别问题转化成MIL问题;然后,将推土机距离(EMD)应用到K最近邻(KNN)算法中,得出一种惰性MIL新方法用于鞋印识别。在包含5种不同类型花纹的鞋印库中进行实验,识别正确率可达91.28%,较之基于欧氏距离的KNN算法,识别精度平均提高4.0%。A novel shoeprint recognition algorithm is proposed based on image block and lazy multi-instance learning （MIL）. Firstly, the algorithm regards every shoeprint image as a bag. Then according to the proportion of foot biometrics, a uniform grid partitioning method is used to divide shoeprint image into 15 sub-blocks, and the texture and shape features of each block are extracted. Therefore the shoeprint image recognition problem is transform into a MIL problem. Secondly, the earth mover＇s distance （EMD） is introduced into the traditional k-nearest neighbour （KNN） method to improve the accuracy of image similarity measure, and a new lazy MIL algorithm is designed for shoe prints recognition. Experimental results on shoeprint images datasets where contains five different types of shoeprint pattern indicate that the recognition accuracy of the new method can reach 91.280%, and that compared with Euclidean distance based KNN algorithm, its average recognition accuracy is improved by 4 %.

关 键 词：多示例学习 鞋印图像识别 纹理-形状特征 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]