期刊文献+

加权KNN的图文数据融合分类 被引量:10

Fusion method via KNN with weight adjustment for the classification of image-text co-occurrence data
原文传递
导出
摘要 目的图文数据在不同应用场景下的最佳分类方法各不相同,而现有语义级融合算法大多适用于图文数据分类方法相同的情况,若将其应用于不同分类方法时由于分类决策基准不统一导致分类结果不理想,大幅降低了融合分类性能。针对这一问题,提出基于加权KNN的融合分类方法。方法首先,分别利用softmax多分类器和多分类支持向量机(SVM)实现图像和文本分类,同时利用训练数据集各类别分类精确度加权后的图像和文本正确判别实例的分类决策值分别构建图像和文本KNN模型;再分别利用其对测试实例的图像和文本分类决策值进行预测,通过最邻近k个实例属于各类别的数目确定测试实例的分类概率,统一图像和文本的分类决策基准;最后利用训练数据集中图像和文本分类正确的数目确定测试实例中图像和文本分类概率的融合系数,实现统一分类决策基准下的图文数据融合。结果在Attribute Discovery数据集的图像文本对上进行实验,并与基准方法进行比较,实验结果表明,本文融合算法的分类精确度高于图像和文本各自的分类精确度,且平均分类精确度相比基准方法提高了4.45%;此外,本文算法对图文信息的平均整合能力相比基准方法提高了4.19%。结论本文算法将图像和文本不同分类方法的分类决策基准统一化,实现了图文数据的有效融合,具有较强的信息整合能力和较好的融合分类性能。 Objective Most existing fusion methods are suitable in situations where the image classification method is identical with the text classification method. However, better classification methods of image and text for image-text co-occurrence data are not identical in many application scenarios. The decision benchmarks of different classification methods are not unified, which would reduce the classification precision of the fusion method. The fusion method based on KNN with weight adjustment for the classification of image-text co-occurrence data is proposed to overcome the problem. Method First, the softmax and multiple classification SVM are used to classify the image and text. The image' s KNN model and the text' s KNN model are constructed using the weighted classification decision values of image and text instances, which are correctly discriminated on a training dataset. Then, the classification decision values of the test instance are predicted by the image' s KNN model and the text' s KNN model. The image classification probability and the text classification proba- bility of the test instance are determined by the number of each category in the k nearest neighbors, and the classification probabilities would unify the classification decision benchmarks. Finally, the fusion coefficient is calculated by the number of image instances and the number of text instances discriminated correctly on a training dataset applied to fuse the classifi- cation probabilities of the image and text for the test instance. Result We performed an experiment on the Attribute Discov- ery dataset and compared the proposed fusion method with the baseline method. Experimental results show that the proposed fusion method achieved higher classification precision than the image classification method and text classification method, and the proposed fusion method increased the average classification precision by 4. 45%. Moreover, the proposed fusion method increased the average information integration ability for image-text data by 4. 19%. Conclusion The proposed fusion method unifies the classification decision benchmarks of the different classification methods for image and text and implements the effective fusion of image-text co-occurrence data. Therefore, the proposed fusion method has a better ability to integrate information, with a better classification performance.
出处 《中国图象图形学报》 CSCD 北大核心 2016年第7期854-864,共11页 Journal of Image and Graphics
基金 国家自然科学基金项目(41301493) 高分对地观测领域学术交流基金项目(GFEX04060103)~~
关键词 图文数据 softmax多分类器 多分类支持向量机 加权KNN 融合分类方法 image-text co-occurrence data classification method softmax multiple classification SVM KNN with weight adjustment fusion
  • 相关文献

参考文献22

  • 1Wu L Z, Oviatt S L, Cohen P R. Multimodal integration-a statis- tical view[ J]. IEEE Transactions on Multimedia, 1999, 1 (4) : 334-341.
  • 2Lanckriet G R G, Christianini N, Bartlett P L, et al. Learning the kernel matrix with semidefinite programming [ J ]. Journal of Machine Learning Research, 2004, 5 : 27-72.
  • 3Scholkopf B, Smola A J. Learning with Keme|s : Support Vector Machines, Regularization, Optimization and Beyond [ M ]. Cam- bridge, MA: MIT Press, 2002.
  • 4Li T, Li L. Music Data Mining: An introduction [ M ]//Li T, Ogihara M, Tzanetakis G. Music Data Mining. London: CRC Press, 2011.
  • 5Li T, Ogihara M. Semisupervised learning from different informa- tion sources [ J ]. Knowledge and Information Systems, 2005, 7(3) : 289-309.
  • 6Tirilly P, Lu K, Mu X, et al. On modality classification and its use in text-based image retrieval in medical databases [ C ]//Pro- ceedings of the 9th International Workshop on Content-Based Multimedia Indexing. Madrid, Spain: IEEE, 2011: 109-114.
  • 7Clinchant S, Csurka G, Ah-Pine J, et al. XRCE' s participation in wikipedia retrieval, medical image modality classification and ad-hoc retrieval tasks of imageclef 2010 [ C ]//Proceedings of Cross-Language Evaluation Forum. Padua, Italy: CLEF, 2010.
  • 8Cheng B, Stanley R J, Antani S, et al. Graphical figure classifi- cation using data fusion for integrating text and image features [ C]//Proceedings of the 12th International Conference on Docu- ment Analysis and Recognition. Washington, DC: IEEE, 2013: 693-697.
  • 9Tian L, Zheng D, Zhu C. Research on image classification based on a combination of text and visual features [ C ]//Proceedings of the 8th Intemational Conference on Fuzzy Systems and Knowledge Discovery. Shanghai, China: IEEE, 2011: 1869-1873.
  • 10Wang G, Hoiem D, Forsyth D. Building text features for object image classification [ C ]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, 2009. Miami, FL, USA: IEEE, 2009 : 1367-1374.

二级参考文献40

  • 1Lowe D G. Distinctive image features from scale-invariant keypoints[J].{H}International Journal of Computer Vision,2004,(2):91-110.
  • 2Dalal N,Triggs B. Histograms of oriented gradients for human detection[A].San Diego,CA,USA:IEEE,2005.886-893.
  • 3Ojala T,Pietikainen M,Harwood D. Performance evaluation of texture measures with classification based on Kullback discrimination of distributions[A].Jerusalem,Irsael:IEEE,1994.582-585.
  • 4Matas J,Chum O,Urban M. Robust wide-baseline stereo from maximally stable extremal regions[J].{H}IMAGE AND VISION COMPUTING,2004,(10):761-767.
  • 5Hinton G E,Osindero S,Teh Y W. A fast learning algorithm for deep belief nets[J].{H}Neural Computation,2006,(7):1527-1554.
  • 6Hinton G E. Learning multiple layers of representation[J].{H}Trends in Cognitive Sciences,2007,(10):428-434.
  • 7Hinton G E,Zemel R S. Autoencoders,minimum description length,and Helmholtz free energy[A].Burlington,USA:Morgan Kaufmann,1994.3-10.
  • 8Rumelhart D E,Hinton G E,Williams R J. Learning Representations by Back-Propagating Errors[M].Cogmitive Modeling:MIT Press,2002.213.
  • 9Vincent P,Larochelle H,Bengio Y. Extracting and composing robust features with denoising autoencoders[A].New York,NY,USA:ACM,2008.1096-1103.
  • 10Lee H,Grosse R,Ranganath R. Convolutional deep belief networks for scalable unsupervised learning of hierarchical representations[A].New York,NY,USA:ACM,2009.609-616.

共引文献145

同被引文献84

引证文献10

二级引证文献81

相关作者

内容加载中请稍等...

相关机构

内容加载中请稍等...

相关主题

内容加载中请稍等...

浏览历史

内容加载中请稍等...
;
使用帮助 返回顶部