Blind detection of targets from LADAR data

We propose a method of blind detection of targets of defense interest from LADAR images. Each pixel of the image is represented by a feature vector computed from a co-occurrence matrix defined over a neighborhood of the pixel. Usual co-occurrence based features are modified to make them independent of gray-scale quantization and a new feature based on Weber´s law of contrast perception is suggested. A set of 14 features are first computed and from there a set of 5 features are selected using a feature selection multilayer perceptron (FSMLP) network. A classifier is trained using only the selected features; we use a multilayered perceptron (MLP), but our system will work with almost any crisp or fuzzy classifier. For an image with unknown content, we then compute the five selected features for each pixel and assign a class label using the trained classifier. Then we find the dense regions in the image, that is, regions having high membership to the target class and declare each such dense region as a possible target. Our simulation results are quite encouraging