Extraction of regions of interest (work in progress)

An algorithm for extracting the regions of interest in images of man-made outdoor scenes (buildings) is proposed. The proposed extraction mechanism exploits the fact that the most interesting parts of an image are those that contain some variation i.e. local texture. A scheme independent of scale and colour is proposed, based on the use of Gaussian kernels, Toboggan enhancement and also an intuitive image processing algorithm which can split concave regions into convex subparts. We compare our algorithm using 100 manually labelled images with the state of the art saliency map model. We show that the proposed algorithm has superior performance in detecting regions of interest in structured images (man-made scenes) compared with the saliency model.