Detection of incomplete rectangular contours with application in archaeology

We introduce a feature useful for detection of structures in images that are perceived as approximate rectangular contours. Given a set of extracted line segments, this rectangularity feature quantifies the degree of alignment of an optimal subset with a contour of rectangular shape, arbitrary size, and aspect ratio. The rectangularity feature has high values not only for perfect rectangles, but also for rectangles with distorted angles, fragmented, or with a completely missing side. However, it has zero response for configurations of lines that do not form at least three sides of a rectangular contour. This feature is based on a graph with nodes corresponding to line segments in a particular neighborhood and with edges encoding spatial relations between the lines. It is defined as the maximum of a target function of node and edge attributes of the graph cliques. The target function assigns high values to nearly convex line configurations with angles close to zero or right angles. We show that the rectangularity feature is effective for detection of ruins of rectangular livestock enclosures in high-resolution remotely sensed images. We also show how to improve the detection performance using an additional feature that quantifies the size of a structure, and by learning a linear feature combination from a small number of representative examples of the livestock enclosures and a large number of available negatives. 1