Aircraft detection: A case study in using human similarity measure

The problem of screening images of the skies to determine whether or not ai rcraft are present is of both theoretical and practical interest. After the most prominent signal in an infrared image of the sky is extracted, the qu estion is whether the signal corresponds to an aircraft. Common approaches calculate the degree of similarity of the shape of the 2D signal with a mod el aircraft using a similarity measure such as Euclidean distance, and make a decision based on whether the degree of similarity exceeds a (prespecifi ed) threshold. We present a new approach that avoids metric similarity meas ures and the use of thresholds, and instead attempts to learn similarity me asures like those used by humans. In the absence of sufficient real data, t he approach allows us to specifically generate an arbitrarily large number of training exemplars projecting near the classification boundary. Once tra ined on such a training set, the performance of our neural network-based sy stem was comparable to that of a human expert and far better than a network trained only on the available real data. Furthermore, the results were con siderably better than those obtained using an Euclidean discriminator.