Object and Scene Category Recognition using a Combination of Dense Color SIFT Descriptors

• Main Authors: Rassem, Taha H., Khoo, Bee Ee, Bayuaji, Luhur, Makbol, Nasrin M., Suryanti, Awang
• Format: Article
• Language: English
• Published: AENSI Publishing 2015
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://umpir.ump.edu.my/id/eprint/9226/1/93-103.pdf; http://umpir.ump.edu.my/id/eprint/9226/; http://www.ajbasweb.com/old/Ajbas_Special-IPN-Penang%20_2015.html
• Institution: Universiti Malaysia Pahang

Nowadays, many features have been proposed for image category recognition. Scale Invariant Feature Transform (SIFT) is one of the important descriptors, which is used in these systems. It is robust against rotation change, viewpoint change, scaling change, but it is partially robust against illumination change. Color SIFT descriptors are proposed to increase the illumination invariant. In this paper, the performances of different color SIFT descriptors densely extracted from the images were evaluated for object and scene recognition. RGB color SIFT, HSV color SIFT, Opponent color SIFT, Transformed-color SIFT and a new proposed color SIFT descriptor based on Ohta color space (Ohta Color SIFT) were used instead of the traditional gray SIFT. The performances of these descriptors and all their possible combinations were evaluated using challenging data sets. Caltech-04, Caltech-101, Caltech-256, Graz-02 are examples of object data sets used, whereas Oliva and Torralba data set (OT) and SUN-398 are examples of scene data sets. Using some combination of dense color SIFT descriptors, remarkable results of classification accuracy were achieved for some data sets such as Caltech-04 and Graz-02 and acceptable accuracy results for the remaining data sets as shown in experimental results.