Toward Advanced Baggage Screening: Reconstruction and Automatic Target Recognition (ATR)
R4-B1

Download Project Report (Phase 2, Year 7)

Project Description

Overview and Significance

A.1. Research on reconstruction

X-ray Computed Tomography (CT) for checked baggage scanning is among the most important elements of transportation security. The performance of image analysis algorithms is highly dependent on the quality of reconstruction imagery that is used as input to this analysis. When components of a reconstructed slice of a bag are poorly resolved or corrupted by artifacts resulting from highly attenuating materials such as metal objects, poor segmentation of materials may result in sufficient ambiguity in the bag’s content to require human intervention due to a “false alarm”. Any improvement in image quality is expected to reduce the number of such cases and reduce the cost of operation of the overall system. The great majority of deployed CT systems utilize image reconstruction methods based on deterministic descriptions of the mapping from the data (sinogram) domain and the image domain. Variants of filtered back-projection are most common and can be implemented at high frame rates appropriate for continuous-flow baggage scanning. Inversion methods, based on more accurate descriptions of the instrument and modeling of reliability of data, may demand more computation in their iterative solution but show promise in related CT applications, which may transfer to the security arena. We call this class of methods “model-based image reconstruction” (MBIR) because they rely on the relatively precise modeling of pixel/Xray interactions, detector behavior, photon counting and electronic noise. The objective of this project is the improvement of MBIR in its specific application to security scanning. Because a major contributor to costly false alarms is poor image quality in the presence of the many metal objects that may be part of baggage or packed within it, our primary focus is a technique to automatically compensate for the beam hardening of metal. The technique separates metal from other image content and models the total attenuation as a polynomial function of both the total attenuation in metal and the total in other materials. The coefficients of the polynomial, which will vary with the X-ray’s spectral shape, are estimated along with the image to allow the best it to the sinogram data, eliminating some of the large inconsistencies due to beam hardening and other metal effects that force artifacts in images when attempting to match data.

A second thrust is variation in the weighting of measurements according to their approximate variances. The most direct model, taken from the Poisson log-likelihood function, dictates weighting proportional to received photon counts. However, the dynamic range of these counts may produce estimated images with disadvantageous properties such as poor noise texture or unnecessary emphasis of artifacts in cases,where the modeling of high-attenuation rays are inadequately accurate. Both the advantages of generic MBIR and the enhancements of our beam-hardening correction approach are evaluated using the Imatron datasets shared among participants in the Task Order 3 effort. The iterative approaches show advantages in subjective image quality. The results in the project’s metrics for segmentation performance are mixed relative to standard, one-pass filtered backprojection (FBP).

A.2. Research on Automatic Target Recognition (ATR)

Automatic target detection and recognition from the scanned images can help to extract important information and support human judgments. However, developing an ATR system is challenged due to metal presence and tight packing. For example, metal introduces strong streaking artifacts with which ATR detects divided objects. In addition, ATR may not be able to separate cluttered objects because of tight packing. The objective of this research is to investigate and develop a new ATR system that can handle the above-mentioned challenging cases. To do this, we will incorporate advanced computer vision algorithms upon the baseline software provided by ALERT through its ATR Task Order. While recent research in computer vision has shown a lot of promising results in each of the ATR components such as image denoising, image segmentation and object detection, most of them are for the application of natural images and very few have been applied to CT images, in particular, for security application. So the question remains as to the potential advantages of the advanced techniques in ATR applications. During the last project period, we successfully developed a new ATR system that incorporates advanced computer vision algorithms such as shape filter and multi-label segmentation. We evaluated the performance using the standard specified metrics (i.e. probability of detection and probability of false alarm) and it already gives improvements over the standard ATR. We will continue to develop new techniques, particularly in advanced feature extraction, in order to further improve the detection accuracy. With these results, we can propose potential directions for the improvement of ATR in aviation security.

Automatic target detection and recognition from the scanned images can help to extract important information and support human judgments.
Phase 2 Year 2 Annual Report
Project Leader
  • Charles A. Bouman
    Professor
    Purdue University
    Email

Students Currently Involved in Project
  • Venkatesh Sridhar
    Purdue University