Object recognition on satellite images with biologically-inspired computational approaches

The human vision system is often significantly superior in extracting and interpreting visual information when compared to classical computer vision systems. The exploitation of existing knowledge about human perception is expected to improve the performance of computational vision systems. Computational visual attention has been reported to improve scene understanding performance. This paper discusses some of the difficulties faced by the current generation of visual attention systems when applied on satellite images. Next, a novel technique for top-down attention is devised which is based on the energy of bottom-up feature maps and overcomes some of the limitations of previous approaches. The computed top-down map is then used as a method of object localization in the object recognition phase that makes use of texture and shape information using local binary patterns, Legendre moments and Hu moment invariants. The proposed algorithm is shown to perform better than other similar systems on satellite images in many aspects.