AUTOMATIC TARGET RECOGNITION-A SOFT COMPUTING APPROACH

Abstract

Automatic target recognition (ATR) problems involve classifying and recognizing targets in highly cluttered environment. Most works in target detection using traditional image processing and computer vision techniques contains multiple tasks such as preprocessing, segmentation, feature extraction and object classification based on hard classification. The development of these detection systems usually• involves a time consuming investigation of good preprocessing and filtering methods, and especially a hand-crafting of different programs for the extraction and selection of specific, important image features for a particular problem. These hard classification approaches does not give good detection rate in noisy environment in ATR problems. A relatively emerging soft computing approach, which is combination of neural network, genetic algorithm and fuzzy classification techniques, is proposed for improvement of target detection. In this dissertation, we directly used the pixel based data (raw image pixel data) as input to the learning systems in which the features relevant to a particular target can be automatically learnt through a back propagation multi layer feed forward GA-refined neural network. A soft classifier using fuzzy C-means classifier is proposed along with GA-refined neural network to improve detection performance. The algorithms tested using aerial static images of aircraft and tank (after adding noise with varying density) on SUN SPARC machine. We found improvement in detection rate using soft classification approach. However, the training time, which is one to two hours in our experimentation, need to be reduced further.