摘要
跨场景行人再识别方法的关键在于特征识别和度量模型的建立,而这两方面的问题都受到图像样本分布的局限,进而使得模型参数的估计出现过拟合现象。针对以上跨场景的行人再识别问题,提出了一种基于半监督的改进KISSME算法。该算法在KISSME学习算法的基础上,根据样本数据的正态分布特性进行重采样,并通过构建循环优化的学习方式弱化模型的拟合强度,增强度量模型的泛化能力,以此建立泛化后的度量模型。再通过联合KISSME度量,构建改进的半监督度量模型。最后,利用行人再识别通用公开数据集VIPe R对改进算法的有效性进行验证,并与SLDDL、RDC、ITML、PCCA、QARR-RSVM和KISSME等算法精度相比较,实验结果表明基于半监督的改进KISSME算法在不同排名下都有明显的优势,尤其在rank-1识别精度上,相较于现有的KISSME算法提升了3.14%,充分验证了该算法的有效性。
As two critically important parts of the cross-camera pedestrian re-recognition method,feature recognition and metric model establishment have been constrained by notorious overfitting of model parameter estimation that arises from improper image sample distribution.This study proposed an improved,semi-supervised KISSME learning algorithm-based method for pedestrian re-recognition.The proposed method succeeded to construct a generalized measurement model by re-sampling normally distributed data,weakening fitting strength through establishment of a circular optimization metric learning method,and improving the model’s generalization capacity.Then,it introduced KISSME metrics to further improve the semi-supervised model.Finally,it verified effectiveness of the improved algorithm by pedestrian re-recognition using the public open VIPeR dataset,results of which were compared with accuracies of SLDDL,RDC,ITML,PCCA,QARR-RSVM and KISSME.It demonstrated the improved,semi-supervised KISSME algorithm to be superior in all recognition accuracy ranks,especially in the rank-1.It achieved an accuracy that was 3.14%higher than that of the existing KISSME algorithm,thereby validating effectiveness of the algorithm proposed in this study.
作者
宋丽丽
李彬
赵俊雅
刘国峰
Song Lili;Li Bin;Zhao Junya;Liu Guofeng(College of Engineering&Technical,Chengdu University of Technology,Leshan Sichuan 614000,China;School of Mechanical Engineering,Wuhan Polytechnic University,Wuhan 430023,China;School of Science,Wuhan University of Technology,Wuhan 430070,China)
出处
《计算机应用研究》
CSCD
北大核心
2020年第7期2227-2231,共5页
Application Research of Computers
