Biologically-inspired algorithms for object recognition

Neurobiologically-inspired algorithms were investigated for their ability to recognize airplanes on satellite images. Standard segmentation algorithms often segregate objects in several segments or merge sections of objects to one segment. To correct such segmentation errors, we used information on the template object (top-down information) to correct such segmentation errors by merging several segments to one object (binding). This strategy was influenced by neurobiological findings demonstrating that top-down information modulates visual representations in primary, visual areas. Our algorithms use gray-scale, size, moment of inertia, and distance between segments to detect segment groups that resemble airplanes. For the images that contained our template, 3 out of 3 large airplanes were correctly identified without false positives. Note that this successful identification was achieved despite the fact that in the original segmentation, each of the three recognized planes was broken up in several segments. In further simulations, we used the same template for 15 more airport images and achieved 8 correct matches, 29 missed matches, and 8 false positives. These results demonstrate that template information is indeed crucial to improve segmentation errors and can lead to substantial improvements of recognition rates. Further research should be conducted to systematically investigate the optimal criteria for reliable object detection.