Building Detection in Aerial Images Based on Watershed and Visual Attention Feature Descriptors

This paper investigates a novel solution for the recognition of objects of interest in aerial images. The solution builds on a combination of algorithms inspired from the human visual system with classical and modern algorithms. The goal is to achieve intelligent and powerful approaches that allow for fast and automatic treatment of complex images. The methodology that is proposed innovatively combines a variation of the classical watershed segmentation algorithm with a series of feature descriptors derived from a computational model of visual attention. The feature descriptors are tuned with a machine learning approach for the task of detecting buildings in aerial images. The experimental evaluation that is conducted demonstrates that objects recognition with features derived from human visual attention performs better than when only traditional features, such as statistical texture descriptors and shape descriptors, are used. As well, the proposed solution obtains better classification rates than those reported on image processing-based recognition of buildings in the remote sensing literature.