A Resegmentation Approach for Detecting Rectangular Objects in High-Resolution Imagery

General image segmentation techniques are divided into twomain classes. The first, called pixel-oriented segmentation,considers each pixel of the image as one graph node. Edgesgenerally connect every pixel to its 3 × 3 neighborhood. Thesecond class considers regions as graph nodes. Such regions arepixel groups obtained by segmentation techniques. This class,called region-oriented segmentation, uses graph structures torepresent the topology, connecting neighbor regions with graphedges. Neighboring regions apply the topological relation oftouch （Point-set topological spatial relations）.

Benediktsson et al. [12] presented a hybrid approach forsegmenting urban images with high spatial resolution, applyingmorphological operations. Afterward, they applied a neuralnetwork to classify the extracted features. When dealingwith panchromatic imagery, Gaetano et al. [13] segmentedmultispectral remote-sensing images, creating regions usingpanchromatic data. Subsequently, regions were merged usingspectral features (i.e., textural based) from the remainingmutispectral information.The nature of urban areas is inherently complex. Althoughhuman operators may extract information from urban imageseasily, computer-based automated interpretation remains a challengingtask. Cinque et al. [14] proposed the use of rectanglesin image queries to express the properties of target regions. Theauthors used resegmentation to build the index of images in thedatabase, which provided the features needed for determiningthe distance between the query and candidates [14].