Neural networks have been proposed as solutions to complex pattern recognition problems at which humans excel but for which algorithmic approaches have not been very successful. Examples of such problems include recognizing an object regardless of its viewing angle or perspective in an image and recognizing objects in SAR imagery for which the actual object signature bears little resemblance to its optical view. We are investigating the performance of back propagation to solve these problems in a multi-layer feedtorward architecture.
We have trained a multi-layer net with back propagation to distinguish different classes of objects in SAR images. The resulting multi-layer feedtorward net operates on the raw grey level images using a small target-size receptive field for translation invariant recognition. To achieve perspective invariance, the network was trained to recognize different views of the same object as being in the same class. We examined the sensitivity of the resulting perspective invarianca to changes in the number of hidden layers, changes in the number of hidden nodes, and changes in the number of perspective views used in training. Results of these sensitivity experiments are presented.
We show how to shape the system response of the net in the vicinity of the object of interest to trade localization accuracy for throughput demands. By training the net using patterns with small shifts, we give up localization but gain invariance to small shifts within the receptive field resulting in a wider system response. Consequently, we need only generate the output image every n pixels (where n is the width of the system response) resulting in a n2 savings in computation. This can be significant in using a dedicated processing engine for real-time applications.