Towards Category Based Large-Scale Image Retrieval Using Transductive Support Vector Machines
In this study, we use transductive learning and binary hierarchical trees to create compact binary hashing codes for large-scale image retrieval applications. We create multiple hierarchical trees based on the separability of the visual object classes by
- PDF / 1,263,527 Bytes
- 17 Pages / 439.37 x 666.142 pts Page_size
- 4 Downloads / 223 Views
Electrical and Electronics Engineering, Eskisehir Osmangazi University, Eski¸sehir, Turkey [email protected], [email protected] 2 TUBITAK UZAY, Ankara, Turkey [email protected]
Abstract. In this study, we use transductive learning and binary hierarchical trees to create compact binary hashing codes for large-scale image retrieval applications. We create multiple hierarchical trees based on the separability of the visual object classes by random selection, and the transductive support vector machine (TSVM) classifier is used to separate both the labeled and unlabeled data samples at each node of the binary hierarchical trees (BHTs). Then the separating hyperplanes returned by TSVM are used to create binary codes. We propose a novel TSVM method that is more robust to the noisy labels by interchanging the classical Hinge loss with the robust Ramp loss. Stochastic gradient based solver is used to learn TSVM classifier to ensure that the method scales well with large-scale data sets. The proposed method improves the Euclidean distance metric and achieves comparable results to the stateof-art on CIFAR10 and MNIST data sets and significantly outperforms the state-of-art hashing methods on NUS-WIDE dataset. Keywords: Image retrieval · Transductive support vector machines Semi-supervised learning · Ramp loss
1
·
Introduction
Large-scale image retrieval has recently attracted great attention due to the rapid growth of visual data brought by Internet. Image retrieval can be defined as follows: Given a query image, finding and representing (in an ordered manner) the images depicting the same scene or objects in large unordered image collections. Despite the great research efforts, image retrieval is still a challenging problem since large-scale image search demands highly efficient and accurate retrieval methods. For large scale image search, the most commonly used method is the hashing method that enables us to approximate the nearest neighbor search. Hashing methods convert each image feature vector in the database into a compact code c Springer International Publishing Switzerland 2016 G. Hua and H. J´ egou (Eds.): ECCV 2016 Workshops, Part I, LNCS 9913, pp. 621–637, 2016. DOI: 10.1007/978-3-319-46604-0 44
622
H. Cevikalp et al.
(typically a binary code) and provide constant or sub-linear search time. Most of the current popular hashing methods [6,7,10,17,18,22,26,27,31] are unsupervised methods and they are built on the assumption that the similar images in the Euclidean space must have similar binary codes. Among these, Locality Sensitive Hashing (LSH) [6] chooses random projections so that two closest image samples in the feature space falls into the same bucket with a high probability. However, due to the semantic gap between the low-level features and semantics, Euclidean distances in the feature space do not reflect the semantic similarities between the images. Furthermore, the state-of-art image visual features are typically high-dimensional vectors ranging from several thousands to millions. As
Data Loading...
