Target segmentation and automatic target recognition using global and local invariant features in forward-looking infrared images

Abstract

In this dissertation, a new automatic target recognition (ATR) algorithm for forward-looking infrared (FLIR) images is presented to identify targets such as battle tanks and armored personal carriers in ground-to-ground scenario. It consists of sequential steps of segmentation, feature extraction and classification. The proposed ATR algorithm satisfies the following requirements: Segmentation algorithm has to be simple and robust to clutters in the real battlefield environment. Features have to be more robust to noise and more invariant to similarity transform than conventional features. Moreover, nonoccluded and partially occluded targets can be identified in the classification phase. The proposed segmentation algorithm of FLIR image is based on fuzzy thresholding and edge detection. At first, a region of interest (ROI) selection method is applied to the FLIR image to remove complex background and to automatically extract a small rectangular ROI including a target. Fuzzy thresholding, which uses image intensity and spatial information, is then performed on the ROI. The Canny edge detection method is utilized to supplement the deficiency of the threshold method. As a preprocessing of edge detection, a new edge sharpening method using a gray-level mathematical morphology operation is performed to sharpen edges. After segmenting a target, the target contour is partitioned into four local boundaries. Radial and distance functions are defined from the target contour and local boundaries, and are used to define global and local shape features, respectively. To identify nonoccluded and partially occluded targets, a new classification method using multiple feature vectors and multilayer perceptrons (MLPs) is proposed. Four feature vectors are composed of the global and local shape features, and are used as inputs of MLPs. The outputs of MLPs are combined to identify targets. The proposed segmentation, feature extraction and classification methods are applied to natura...