Image classification for automatic target recognition using artificial neural networks

摘要

Automatic Target Recognition (ATR) is a challenging problem in each case the need exists to observe targets, and to automatically recognize, characterize, and identify them. Moreover, the recognition, characterization, and identification processmust occur in a challenging observation environment. ATR typically involves background effects such as ground clutter or deliberate attempts to camouflage or obscure the presence of a target, or even in the presence of effects such as electronic countermeasures (ECM) and deception. Thus, with limited historical data and in a challenging observation environment, the need exists for human-like vision and reasoning in near real time speed. Clearly, improvements to ATR is vital to military mission,affecting both targeting and survivability. In this paper, we present an image classification system for ATR applications using artificial neural networks. The real challenge in this kind of problem is the huge feature vectors representing images to beintroduced to neural network algorithms. In order to overcome this problem, increase the number of targets to be classified, and speed up the neural network training and processing, only image features necessary for the classification process need to becomputed and used as neural network inputs. In this paper we present two different methods for image feature extraction using cosine and wavelet transform. Results from both of the two developed systems are presented and compared.