Toward Advanced Baggage Screening: Reconstruction and Automatic Target Recognition (ATR)
Computed Tomography (CT) has been widely used in applications of medical diagnosis and security inspection. In the transportation security area, one of the most important CT applications is checked luggage screening. The CT scanner will generate the cross-sectional images of the bag. Traditionally, images are reconstructed using direct methods, such as filtered back projection (FBP). MBIR algorithms have recently been shown effective in improving the image quality by reducing metal artifacts and increasing spatial resolution. Instead of relying on a single pass through the measurement data alone, the model-based methods are based on the mathematical optimization of a statistical model of both the acquired data and the unknown image. (By incorporating better models of the system and increased knowledge of the image, the model-based approach has the potential to further reduce metal artifacts, improve resolution and suppress structured noises due to clutter.) Given CT scans, automatic image analysis for recognizing objects of interest can reduce the cost of human labor, help to extract important information, and support human judgments. A typical Automatic Target Recognition (ATR) system for security applications mainly consists of three steps: image segmentation, feature extraction, and target classification. Our goal is to advance each of these three steps in the ATR system, in order to improve the current state-of-the-art ATR performance. To achieve this goal, we will investigate the dataset in depth to understand the challenges in the detection process, design new algorithms for each unit separately, and integrate them to build up the whole ATR system. In particular, we will study the use of advanced segmentation, feature extraction and classification algorithms for improving performance in challenging cases, such as images with cluttered objects, artifacts, and in-accuracy CT numbers.