Unsupervised Regions of Interest Extraction based on Visual Attention and SIFT

B. Chaudhury, O. Marques (USA), G.B. Borba, and H.R. Gamba (Brazil)


Region of interest (ROI) extraction, visual attention model, scale invariant feature transform (SIFT), keypoints, saliency.


This paper presents an unsupervised approach to ROI extraction, using a combination of salient keypoints from Scale Invariant Feature Descriptors (SIFT) and points of attention (POAs) from a computational model of human visual attention (VA). In our framework, the main idea is to locate keypoints in the image by means of SIFT and cluster those keypoints using the POAs as centroids. The obtained clusters originates polygons that are further joint together to form the ROI contours. In spite of the popularity and robustness of the SIFT algorithm, it has a major problem of generating a large number of unnecessary keypoints, belonging mostly to the background. This imposes a limitation for its straight application on the support for ROI extraction, since the immediate effect is a poor discrimination of the saliencies. In order to solve this, a two-folded ļ¬ltration method of the numerous keypoints has been applied, based on the least probable colors of the image. The algorithm has been evaluated on a wide variety of images and the results show that the proposed architecture can provide promising results in the task of ROI extraction.

Important Links:

Go Back