A Simple Deep Learning Network for Target Classification

Abstract

In this paper, the primary objective is to develop a target classification algorithm based on deep learning network for real time application, primarily for an automatic target recognition (ATR) system in combat operation. Deep learning exploits neural networks for the learning of useful illustrations of features from the data directly for regression and classification. For the training there are two options: 1) formulate and train a new network, or 2) use pre-trained networks. This paper shows an example on developing and training a simple deep learning convolutional neural network (CNN) for the classification of three different type of targets namely: tank, truck and apc. In order for the creation of the CNN, we have created a training database consisting of 1266 sample images of those three targets. Then 30% of those images are selected randomly for the training purposes and the rest 70% of the images are used for the validation or testing purposes. This paper presents a preliminary study on the creation of a CNN for the classification of targets from forward looking infrared imagery (FLIR) supplied by Army Missile Command (AMCOM). The accuracy of the created CNN is 95.7%. The ultimate goal is to integrate this classification algorithm into our ongoing work on the design of a complete ATR system.