Scene categorization using boosted back-propagation neural networks

  • Authors:

  • Xueming Qian;Zhe Yan;Kaiyu Hang;Guizhong Liu;Huan Wang;Zhe Wang;Zhi Li

  • Affiliations:

  • Department of Information and Communication Engineering, Xi'an Jiaotong University, Xi'an, China;Department of Information and Communication Engineering, Xi'an Jiaotong University, Xi'an, China;Department of Information and Communication Engineering, Xi'an Jiaotong University, Xi'an, China;Department of Information and Communication Engineering, Xi'an Jiaotong University, Xi'an, China;Department of Information and Communication Engineering, Xi'an Jiaotong University, Xi'an, China;Department of Information and Communication Engineering, Xi'an Jiaotong University, Xi'an, China;Department of Information and Communication Engineering, Xi'an Jiaotong University, Xi'an, China

  • Venue:
  • PCM'10 Proceedings of the 11th Pacific Rim conference on Advances in multimedia information processing: Part I
  • Year:
  • 2010

Quantified Score

Hi-index 0.00

Visualization

Abstract

Scene categorization plays an important role in computer vision, image content understanding, and image retrieval. In this paper, back-propagation neural network (BPN) is served as the basic classifier for multi-class scene/image categorization. Four features, namely, SPM (spatial pyramid appearance descriptor represented by scale invariant feature transform), PHOG (pyramid histogram of oriented gradient), GIST, and HWVP (hierarchical wavelet packet transform) are selected as the basic inputs of BPNs. They are the appearance, shape and texture descriptors respectively. For an M (M2) classes scene categorization problem, we cascade M one-versus-all BPNs to determine the accurate label of an image. An offline multi-class Adaboost algorithm is proposed to fuse multiple BPN classifiers trained with complementary features to improve scene categorization performance. Experimental results on the widely used Scene-13 and Sport Event datasets show the effectiveness of the proposed boosted BPN based scene categorization approach. Scene categorization performances of BPN classifiers with input features: SPM, PHOG, GIST and HWVP, boosted BPN classifiers of each of the four features, and the boosted classifiers of all the four features are given. Relationships of boosted classifiers number and the scene categorization performance are also discussed. Comparisons with some existing scene categorization methods using the authors' datasets further show effectiveness of the proposed boosted BPN based approach.