Comparing classical and neural network classification techniques for image feature identification

One of the main requirements of an image processing system is the ability to automatically recognise a given object within a scene. Many military systems rely on the use of imagery based upon infra-red (IR) technology. Another requirement is for robustness over a wide range of operating conditions. Another overall consideration in any system is one of processing requirements in terms of speed, cost and physical size; military systems often impose severe constraints on the physical size. A simple approach to recognising objects is to segment the image to form a set of regions, where one or more of the regions is representative of the object. This step is then followed by the computation of a set of rudimentary numerical metrics for each object region indicated by the segmentation process. For each region, the set of metrics form a feature vector whose values are, in some sense, representative of the nature of the region. Ideally the elements of the feature vector associated with a object region of one class will significantly differ from the feature vectors describing regions from other classes. So the objective is to classify the object region given the contents of the feature vector computed for that region. Traditionally, a typical feature classification process made use of well established algorithms based upon linear discriminant or nearest neighbour techniques. More recently neural network classification techniques have emerged and the objective of the present paper is to perform some initial simple experiments to compare the traditional and more contemporary classification techniques