Article


Article Code : 13950803192273717

Article Title : Concept Detection in Images Using SVD Features and Multi-Granularity Partitioning and Classification

Keywords :

Journal Number : 19 Summer 2017

Visited : 71

Files : 946 KB


List of Authors

  Full Name Email Grade Degree Corresponding Author
1 Kamran Farajzadeh k.farajzadeh@iau-tnb.ac.ir Post Graduate Student Graduate Student
2 Esmail Zarezadeh zarezadeh@aut.ac.ir Post Graduate Student Graduate Student
3 Jafar Mansouri jafar.mansouri@gmail.com Graduate PhD

Abstract

Effective low-level visual features are crucial factors for the semantic concept detection in images. In this paper, new static features (namely, right singular feature vector, left singular feature vector, and also singular value feature vector) are proposed. These features are derived by applying singular value decomposition (SVD) "directly" to the "raw" images. The proposed SVD features are different from features which are obtained from the principal component analysis (PCA) or the latent semantic analysis (LSA). The advantage of the proposed SVD features is that in which edge, color and texture information is integrated simultaneously and this information is sorted based on their importance for concept detection. Feature extraction is performed in a multi-granularity partitioning manner. Since SVD features have the high dimensionality, classification is carried out with the K-nearest neighbor (K-NN) algorithm which utilizes a new distance function, namely, multiplicative distance. This distance is "stable" in the high-dimensional space. Classification is carried out for each grid partition of each granularity separately (in contrast to the existing systems in which classification is performed just for the whole image, not for each partition). Therefore, if some partitions do not contain the target concept, the results of classifications on these partitions do not affect the results of classifications on partitions containing that concept. This leads to the performance improvement. Experimental results on PASCAL VOC and TRECVID datasets show the effectiveness of the proposed SVD features and multi-granularity partitioning and classification method.