Distinguishing Buried Mines from Battlefield Clutter UsingHolographic Radar

It is said that 90% of a deminer's time is spent locating and removing harmless clutter objects. In order to address whether this situation can be improved, a series of holographic radar scans have been made using the Rascan holographic radar system of a simulated mine (a cylindrical plastic sweet container ¯lled with sugar) and several common battle¯eld clutter objects: a short length of barbed wire, a small unexploded bullet, a coke tin and a large °int stone. The raw Rascan images which consist of 5 frequencies at parallel and cross polarisations have been combined together by taking the modulus of the image amplitude less the mean background and summing over all the frequencies and two polarisations. The series consists of the same objects placed on a 400mm £ 470mm area measured at10mm resolution in a sand box ¯lled with ¯ne dry sand at about 60mm depth. Some 15 similar images were taken with the same objects placed at di®erent positions over the area and at di®erent depths and orientations. The above ¯gure shows that the eye, armed with the knowledge of what the Rascan images represent, is readily able to identify them from a combination of their shape, their amplitude, and their \texture" or degree of un-evenness of amplitude. The question is, \Can the skills of the human eye be reproduced by an automated system?". The image is scanned by a receptive ¯eld of a size large enough to contain the objects to be discriminated. If the amplitude above background over the receptive ¯eld exceeds some de¯ned threshold, the position of maximum integrated amplitude is located. The part of the image within the receptive ¯eld placed optimally over the located object is saved by the system, as an example representation of that object, to be used in subsequent training. The coloured boxes in the ¯gure show such automatically generated object images. The series of \training" images of known objects is presented to a neural network which is able to adjust its complexity so that other \test" images are optimally classi¯ed. Finally quite new \validation" images which have not been previously presented to the network can be used to identify the classi¯cation accuracy in terms of the percentage of correctly classi¯ed objects and the percentage of false alarms.