Previous [ 1] [ 2] [ 3] [ 4] [ 5] [ 6] [ 7] [ 8] [ 9] [ 10] [ 11] [ 12]

@

Journal of Information Science and Engineering, Vol. 20 No. 6, pp. 1197-1212 (November 2004)

Learning Visual Concepts from Image Instances

Jun-Wei Hsieh, Cheng-Chin Chiang1, Yea-Shuan Huang2 and W. E. L. Grimson3
Department of Electrical Engineering
Yuan Ze University
Chungli, 320 Taiwan
E-mail: shieh@saturn.yzu.edu.tw
1Department of Computer Science and Information Engineering
National Dong Hwa University
Hualien, 974 Taiwan
2Advanced Technology Center, Compuer and Communication Laboratories
Industiral Technology Reserach Institute
Hsinchu, 310 Taiwan
3Artificial Intelligence Laboratory
Massachusetts Institute of Technolgoy
MA 02139-4307, U.S.A.

This paper presents a novel method of retrieving images by learning the commonality of instances from a set of training examples. The proposed scheme uses a coarse- to-fine algorithm to find the desired visual concepts from a set of instances for successful image retrieval. The learner at the coarse stage attempts to partition training data into two smaller compact sets (relevant and irrelevant) to reduce the size of the training examples, thus improving the efficiency of concept learning at the refined stage. At the refined stage, a proposed verification scheme is employed to verify each instance obtained at the coarse stage by examining its indexing and filtering capabilities based on a pool of images. Due to this extra examination step, the desired visual concepts can be learned more accurately, leading to significant improvement in image retrieval. Since no time-consuming optimization process is involved, all the desired visual concepts can be learned online. Experimental results are provided to verify the superiority of the proposed method.

Keywords: multiple instances, diverse density algorithm, relevance feedback, region instances, image retrieval

Full Text () Retrieve PDF document (200411_10.pdf)

Received April 29, 2003; revised August 11, 2003 & September 19, 2003 & October 8, 2003; accepted Novemmber 21, 2003.
Communicated by Pau-Choo Chung.