Automatic extraction of regions of interest from images based on visual attention models

Abstract

This thesis presents a method for the extraction of regions of interest (ROIs) from images. By ROIs we mean the most prominent semantic objects in the images, of any size and located at any position in the image. The novel method is based on computational models of visual attention (VA), operates under a completely bottom-up and unsupervised way and does not present con-straints in the category of the input images. At the core of the architecture is de model VA proposed by Itti, Koch and Niebur and the one proposed by Stentiford. The first model takes into account color, intensity, and orientation features and provides coordinates corresponding to the points of attention (POAs) in the image. The second model considers color features and provides rough areas of attention (AOAs) in the image. In the proposed architecture, the POAs and AOAs are combined to establish the contours of the ROIs. Two implementations of this architecture are presented, namely 'first version' and 'improved version'. The first version mainly on traditional morphological operations and was applied in two novel region-based image retrieval systems. In the first one, images are clustered on the basis of the ROIs, instead of the global characteristics of the image. This provides a meaningful organization of the database images, since the output clusters tend to contain objects belonging to the same category. In the second system, we present a combination of the traditional global-based with region-based image retrieval under a multiple-example query scheme. In the improved version of the architecture, the main stages are a spatial coherence analysis between both VA models and a multiscale representation of the AOAs. Comparing to the first one, the improved version presents more versatility, mainly in terms of the size of the extracted ROIs. The improved version was directly evaluated for a wide variety of images from different publicly available databases, with ground truth in the form of bounding boxes and true object contours. The performance measures used were precision, recall, F1 and area overlap. Experimental results are of very high quality, particularly if one takes into account the bottom-up and unsupervised nature of the approach.