An algorithm or the neural fusion of IRST & radar for airborne target detection

Abstract

This paper investigates in to the possibility of using a BAM correlating encoding based neural fusion of IRST and radar at the point of the IRST's maximum range. During training phase (in peace time or at a safe place or range), intermittent appearance of a target on IRST display can be recorded in a temporal array. Corresponding intermittent appearance on radar will also be recorded on another array. Treating IRST array as horizontal array and radar array as vertical one, these two binary arrays will be made bipolar by replacing 0s with 1s and multiplied and square or rectangular arrays obtained. A large number of sets can be obtained like this representing the entire representative situations and corresponding square matrices added to form a general weight matrix. Data corresponding to the intermittent appearances of targets and other objects on radar display will be kept in the forms of binary arrays as database. In application phase, if a target is detected through the radar at the maximum range where target appears on the IRST display, radar can be switched off. IRST display will show intermittent appearances of the target, which may be difficult to track or even to discriminate from nearby bird or far off planet/star. The data collected for a number of frames for a single target's estimated intermittent appearance will be stored in an array as binary data. This binary array will be multiplied with the general weight matrix and resulting vertical matrix after thresholding represents an estimated radar data. This approximated radar binary array can be compared with stored radar representations and nearest class can be declared the class of the object present in the scene. As a further improvement, this whole experiment can be performed in a peaceful condition and the estimated radar representation obtained can be compared with exact radar representation and error calculated. Another neural model (like multilayer perceptron) can be used to provide a feedback to correct the errors in the radar estimation. The process basically works as an adaptive filter and predicts a radar array corresponding to the IRST array. The success of the algorithm depends on the training (selecting representative situations) and the implementation methods. Optical implementation with optical associative memories can also be experimented for faster processing.