Automated Tracking and Classification of Infrared Images Jahangheer. S. Shaik and Khan. M. Iftekharuddin, Intelligent Systems and Image Processing Lab Department of Electrical and Computer Engineering, The University Of Memphis Memphis, TN 38018. Abstract The problem of automatic target recognition (ATR) and image classification has been active research fields in image processing and neural networks. In this paper, we explore novel ATR techniques such as object pre-processing, tracking and classification for sequence of infrared (IR) targets. We attempt to track the IR objects automatically without the prior knowledge of the object. We enhance our algorithm to track multiple objects in image frames based on the knowledge of the histogram of the targets. We also propose algorithms for classification of IR targets. We perform image segmentation and extract intensity and edge features of the object. Finally, we exploit these different types of features such as statistical (based on intensity) and shape (based on edge information) of the object for classification using a self-organizing map (SOM) classifier. Key words: Automatic Target Recognition, Edge detection, IR images, Morphological operations, and Self-organizing map classifier. Introduction and Background Review The problems of ATR and image classification have been active research areas in image processing and neural networks [2-9]. The ATR is a processing and understanding of an image in order to recognize the targets [1]. Image identification and restoration may be defined as the problem of determining the kind of degradations an image has suffered, and undoing these degradations in some optimal manner such that the target may be identified [10]. Once the target(s) are identified, it may be necessary to track the same for ATR purposes. In general, the image may have one or more objects in it, some of which are targets, while the others are different external operating conditions (EOCs) such as clutter that may be perceived as the potential targets. The present state of the art machine vision systems do not even approach the performance of human vision in image understanding (IU), suggesting that there is still much to be learned from biological vision systems. With this in mind, researchers have chosen biomorphic-engineering approaches to IU intricate ATR and IU Problems. Mathematical Preliminaries Image pre-processing and classification have been the concern of ATR researchers for a long time. Targets and backgrounds reflect the illuminating light where the reflectivities are strong functions of wavelength. The light is not only reflected by the targets but also by the background and scattered by atmospheric aerosol. It is necessary that we analyze the image in order to identify the potential targets. We propose a tracking algorithm to focus on potential targets by using the correlation techniques. The correlation is accomplished between images ) , ( y x I and ) , ( y x G by formula ) , ( * * ) , ( y x G y x I F corr - - = . Where ‘**’ represents the convolution operation. Once we track the targets we may have to extract the features of the object in order to identify it correctly. Features may be low-level features such as edges of the image. It may be inappropriate to use all the edge information for classification. One of the most frequently used transformation for image compression is the discrete cosine transform (DCT). The kernel of DCT is defined as [11]       + ×       + = N v y N u x v u v u y x g 2 ) 1 2 ( cos 2 ) 1 2 ( cos ) ( ) ( ) , , , ( π π α α . (1) (2)