Image-to-class distance metric learning for image classification

  • Authors:

  • Zhengxiang Wang;Yiqun Hu;Liang-Tien Chia

  • Affiliations:

  • Center for Multimedia and Network Technology, School of Computer Engineering, Nanyang Technological University, Singapore;Center for Multimedia and Network Technology, School of Computer Engineering, Nanyang Technological University, Singapore;Center for Multimedia and Network Technology, School of Computer Engineering, Nanyang Technological University, Singapore

  • Venue:
  • ECCV'10 Proceedings of the 11th European conference on Computer vision: Part I
  • Year:
  • 2010

Quantified Score

Hi-index 0.00

Visualization

Abstract

Image-To-Class (I2C) distance is first used in Naive-Bayes Nearest-Neighbor (NBNN) classifier for image classification and has successfully handled datasets with large intra-class variances. However, the performance of this distance relies heavily on the large number of local features in the training set and test image, which need heavy computation cost for nearest-neighbor (NN) search in the testing phase. If using small number of local features for accelerating the NN search, the performance will be poor. In this paper, we propose a large margin framework to improve the discrimination of I2C distance especially for small number of local features by learning Per-Class Mahalanobis metrics. Our I2C distance is adaptive to different class by combining with the learned metric for each class. These multiple Per-Class metrics are learned simultaneously by forming a convex optimization problem with the constraints that the I2C distance from each training image to its belonging class should be less than the distance to other classes by a large margin. A gradient descent method is applied to efficiently solve this optimization problem. For efficiency and performance improved, we also adopt the idea of spatial pyramid restriction and learning I2C distance function to improve this I2C distance. We show in experiments that the proposed method can significantly outperform the original NBNN in several prevalent image datasets, and our best results can achieve state-of-the-art performance on most datasets.