Focus-of-attention strategies for finding discrete objects in multispectral imagery

Tools that perform pixel-by-pixel classification of multispectral imagery are useful in broad area mapping applications such as terrain categorization, but are less well-suited to the detection of discrete objects. Pixel-by-pixel classifiers, however, have many advantages: they are relatively simple to design, they can readily employ formal machine learning tools, and they are widely available on a variety of platforms. We describe an approach that enables pixel-by-pixel classifiers to be more effectively used in object-detection settings. This is achieved by optimizing a metric which does not attempt to precisely delineate every pixel comprising the objects of interest, but instead focusses the attention of the analyst to these objects without the distraction of many false alarms. The approach requires only minor modification of exisiting pixel-by-pixel classifiers, and produces substantially improved performance. We will describe algorithms that employ this approach and show how they work on a varitety of object detection problems using remotely-sensed multispectral data.