摘要
针对基于无监督特征提取的目标检测方法效率不高的问题,提出一种在无标记数据集中准确检测前景目标的方法.其基本出发点是:正确的特征聚类结果可以指导目标特征提取,同时准确提取的目标特征可以提高特征聚类的精度.该方法首先对无标记样本图像进行局部特征提取,然后根据最小化特征距离进行无监督特征聚类.将同一个聚类内的图像两两匹配,将特征匹配的重现程度作为特征权重,最后根据更新后的特征权重指导下一次迭代的特征聚类.多次迭代后同时得到聚类结果和前景目标.实验结果表明,该方法有效地提高Caltech-256数据集和Google车辆图像的检测精度.此外,针对目前绝大部分无监督目标检测方法不具备增量学习能力这一缺点,提出了增量学习方法实现,实验结果表明,增量学习方法有效地提高了计算速度.
Aiming at the low accuracy of object detection methods based on unsupervised object detection, this paper proposes a foreground object detection method in unlabeled dataset. The basic idea is that correct feature clustering results can guide future object feature extraction, while the accurate foreground object features can improve the accuracy of feature clustering. The proposed method extracts local features from unlabeled images and then clusters features based on minimum feature distances. By matching pairwise images in the same cluster, feature weights are computed through feature correspondence. Finally, the updated feature weights are used to guide feature clustering in the next iteration. We simultaneously group similar images and detect foreground objects after iterations. The experimental results on Caltech-256 and Google car side images demonstrate the effectiveness of our method. Furthermore, due to the present unsupervised object detection methods lacking of incremental learning ability, we propose an incremental implementation of our method. The experimental results show the incremental learning method can improve the computation speed greatly.
出处
《计算机研究与发展》
EI
CSCD
北大核心
2012年第8期1721-1729,共9页
Journal of Computer Research and Development
基金
国家自然科学基金项目(60273064)
广东省工业攻关计划基金项目(2004B10101032)
关键词
前景目标检测
无监督学习
特征聚类
增量学习
特征提取
foreground object detection
unsupervised learning
feature clustering
incremental lear.ning
feature extraction
