Robust automatic target recognition in second generation forward-looking infrared images

This dissertation presents a general methodology for automatic target detection and recognition in complex environments. Within this methodology, new algorithms have been developed to detect, segment and recognize tactical targets from second generation Forward Looking Infrared (FLIR) images for Automatic Target Recognition (ATR) applications. These images are obtained from sensors that are at distances greater than 1500m from the target; thus the target occupies less than 5% of the image. An initial detection algorithm has been developed to robustly identify all possible regions in the image that are candidate locations of targets. This detection is obtained by accurately modeling the background using Gaussian and Weibull functions. A two-stage focused analysis of each candidate target location is then performed to get an accurate representation of the target boundary. A region-growing procedure that uses a diffusion process driven by the underlying probability distribution of the background, and modulated by local shape changes of the target, is used to get an initial estimate of the target shape. The boundary of the target is then combined with salient edge information in the image to arrive at a more accurate representation of the target boundary. A computationally efficient and flexible method to incorporate the salient edge information into the region boundary has been developed by formulating it as a Bayesian classification problem. Finally, to reduce the false alarm rate, a higher level interpretation module is used to classify the detected areas as man-made or natural objects using geometric and FLIR-intensity based features extracted from the target.; For target recognition, a methodology for recognition based on target parts is proposed. The methodology is based on a hierarchical, modular structure for object recognition. In the most general form, the lowest level consists of classifiers that are trained to recognize the class of the input target, while at the next level, classifiers are trained to recognize specific targets. At each level, the targets are recognized by their parts, and thus each classifier is made up of modules, each of which is an expert on a specific part of the target. Each modular expert is trained to recognize one part under different viewing angles and transformations. A Bayesian realization of the proposed methodology is presented in this dissertation, in which the expert modules represent the probability density functions of each part, modeled as a mixture of densities to incorporate different views (aspects) of each part. Recognition is performed using two different techniques. In the first method, recognition relies on the sequential presentation of the parts to the system, without using any relational information between the parts. In the second approach, relative positional information of the parts is used by incorporating this information within a Markov Random Field framework. A new method to decompose a target into its parts is also presented here. Results obtained for target detection, segmentation and recognition in numerous second generation Forward Looking Infrared (FLIR) images are also presented here.