Comparison of model based, vision, statistical based and neural net based ATRs

An effort is made to establish a common ground upon which a comparison of model-based vision (MBV), statistical-based, and neural-net-based (NN) automatic target recognizer (ATR) approaches can be performed. A definition for each type of ATR as compared to a generic ATR is provided. Upon these definitions, the differences, purported risks, and benefits are described. It is found that the comparison between statistical, MBV and NN approaches to ATR can only be made at a very high system level. The differences primarily deal with how the desired target is represented within the ATR. These representation differences lead to other implementation differences, which affect the performance flexibility and technical achievability of each approach as it is faced with the realities of new target types and engagement conditions. It is noted that as attempts are made to become more specific, there are always attempts to indicate that a particular technique does not belong exclusively to one class of recognizers versus another. Indeed, a hybrid approach of using models to train a statistical-based classifier is valid, but not clearly separable into one class of recognizers