<title>Detection of targets varying in fine details, rotation, and translation</title>

Abstract

Automatic target recognition (ATR) has been a topic of interest for many researchers because of its applications in the fields of defense, manufacturing, health sciences etc. The ability of massive parallelism and high-speed classification of neural network (NN) makes it a good choice for ATR. In this paper, we present a novel ATR approach for targets varying in fine details, rotation and translation using a Learning Vector Quantization (LVQ) NN. The algorithm includes two phases such as the feature extraction and the NN discrimination. The feature extraction algorithm obtains the features of the original and distorted targets in the Fourier-log-polar domain and clusters them into a set of centers. These centers are then applied as inputs to an LVQ NN for training. We explore two distinct discrimination algorithms. In the first algorithm, unrotated target features are applied as training vectors and the network is tested with features of rotated targets. In the second algorithm, the LVQ NN is trained using the rotated images and tested for unknown rotated target features. The algorithm is also applied for the cases of targets varying in fine details and translation and a combination of rotation, fine details and translation.