Image pre-classification based on saliency map for image retrieval

  • Authors:

  • Zhen Liang;Hong Fu;Zheru Chi;Dagan Feng

  • Affiliations:

  • Centre for Multimedia Signal Processing, Department of Electronic and Information Engineering, The Hong Kong Polytechnic University, Hong Kong, China;Centre for Multimedia Signal Processing, Department of Electronic and Information Engineering, The Hong Kong Polytechnic University, Hong Kong, China;Centre for Multimedia Signal Processing, Department of Electronic and Information Engineering, The Hong Kong Polytechnic University, Hong Kong, China;Centre for Multimedia Signal Processing, Department of Electronic and Information Engineering, The Hong Kong Polytechnic University, Hong Kong, China and School of Information Technologies, The Un ...

  • Venue:
  • ICICS'09 Proceedings of the 7th international conference on Information, communications and signal processing
  • Year:
  • 2009

Quantified Score

Hi-index 0.00

Visualization

Abstract

In content-based image retrieval, it is helpful to add a pre-classification module to classify a query image into attentive class or non-attentive class. Based on the pre-classification result, a suitable retrieval strategy is adopted for the query image presented. In this paper, we proposed a Multi-Layer Perceptron (MLP) classifier with the features extracted from saliency map to classify both the query image and database images into attentive images and non-attentive classed. A dataset of 1,000 images was selected from the 7,346 Hemera color image database for our experiments. Various features extracted from the saliency map are investigated, including the number and the average size of saliency regions, the variance of the sizes of saliency regions, as well as the sizes of the three most conspicuous saliency regions. The number and average size of saliency regions in the saliency map are shown to be the most discriminative features with which a classification rate of better than 98% can be achieved. To facilitate multi-strategy image retrieval, we define an attentive index between 0 and 1 based on the two most discriminative features to indicate how attentive an image is. Finally, two image retrieval strategies based on the attentive index are proposed, and the corresponding image retrieval performances are evaluated on the 7,346 Hemera color image database with a comparison with the other conventional image retrieval methods.