Download Publication : Download Publication

 

 Class 1, Raw SIFT features

Class 2, Raw SIFT features

Class 1, 35 selected Codebook features

Class 2, 35 selected Codebook features

 

Abstract:

This paper introduces a new method of codebook-based image categorization by building the codebook using scored and selected local features in the image. Different from traditional clustering-based codebook generation that may lead to codeword uncertainty and plausibility, the proposed Matching and Consensus (M&C) process follows the paradigm of feature selection: Based on distance metrics, the M&C process examines salient local features recurring over training images and produces scores that quantify the levels of relevance of the features to the image categories. By selecting features with the highest scores into the codebook, the method is expected to filter out non-representative candidates and keeps the most informative codewords for the category. We evaluate on five image sets for tasks of binary object identification and multi-class biological image classification. Experiments show that our method promotes very parsimonious codebooks that contain highly representative features and deliver a robust classification performance.

Paper:

Bala S. Divakaruni and Jie Zhou: Image Categorization using Codebooks Built from Scored and Selected Local Features. In: Proceedings of The 2011 International Conference on Image Processing, Computer Vision and Pattern Recognition (IPCV), 2011, Volume 1, pages 3--9.

Implementation:

DOWNLOAD, instructions for using are in the README file. Please read the COPYRIGHT notice before using.

Document:

Click here for brief introductory slides.

Contact:

Bala S. Divakaruni
mrdivakaruni [at] gmail.com
Jie Zhou
jzhou [at] niu.edu



Example Image Source Acknowledgement: Dr. J H Simpson, Janelia Farm Research Campus, HHMI.