Image Analysis of Radiographic Scans for Detection of Threats in Cargo Containers
Motivation
The current standard methods for examining containers that pose a potential terrorist threat involve Department of Homeland Security (DHS) officers generally conducting either non-intrusive or physical inspections. The non-intrusive inspection (NII) involves use of X-ray or gamma ray scanners to generate an image of the contents, which DHS officers review for anomalies. DHS officers also scan cargo using radiation detection devices. When an irregularity is identified, officers may physically examine all or a portion of the container’s contents (see Figure 1 below). [pic]
Figure 1: Cargo container being examined by portable VACIS system.
Problem Importance
The methods implemented will enable the identification and evaluation of cargo radiographic images having an extremely widespread and powerful impact for Homeland Security.
Project Description
Radiographic imaging has become an important tool for screening cargo containers for potential nuclear or radiological threats. We are investigating methods to extract features from these images that effectively characterize the contents and when combined with other measurements and information could indicate whether or not a threat is present. Analysis of single-energy radiographs is made particularly challenging by the large variety of cargo contents and the overall volume and mass of standard intermodel shipping containers. Once these features are extracted, we will leverage machine learning methodologies to perform threat detection utilizing these features along with other signature measurements and contextual information. The other features may include spatial profiles, gamma-ray spectra and neutron emissions from radioactive materials, weight, volume, location, condition, origin, shipper, destination, stated contents, etc. The data used to support this analysis consists of 669 radiographic scans measured by a SAIC VACIS (Vehicle and Cargo Imaging System) at the Port of Oakland. The machine learning methods that will be investigated include: particle filters (PF), support vector machines (SVM’s) and Variational Bayesian factor analysis (VBFA). Features that will be studied include those generated via Sobel Edge operators and those based upon both local and global statistical characteristics of the images. We will apply each of these techniques to the entire image dataset, as well as to a duplicate set of images that have a simple surrogate threat overlaid on each image. This allows us to characterize the separability and variability of both individual and combinations of features due to the presence of a threat and provide threat and non-threat training data for machine learning algorithms that could be used to analyze these features in combination with the features from other sources.
This effort supports both directorate technology and infrastructure by providing an environment facilitating the rapid prototyping and development of feature extraction and machine learning algorithms.
Technical Approach and Results
Variational Bayesian Factor Analysis (VBFA) methods have been utilized to extract features from VACIS-type radiographic scans that could in the future be used with the RIID, RPM, PRD and context features to improve the detection, classification and identification of potential nuclear threats. An overview of feature extraction and VBFA training is shown in Figures 2 and 3 below.
Figure 2: Edge detection process used for VBFA training.
In this report, all cargo images were divided into three categories: vehicles, merchandise, and bundles. Opaque circular objects were added to each image to simulate a threat image & non-threat image pair.
Vehicle Merchandise Bundles
Each image paired with added threat objects.
Figure 3: Radiographic Cargo images with simulated opaque threats and non-threat pairs.
In each category VBFA training factors were computed from a selected set of training images. The testing performance of the merchandise images is shown first (Figure 4) followed by the bundles images (Figure 5).
In Figure 4 the left plot shows the performance testing of non-threat merchandise images. The Y axis of the plot is the percent difference between two likelihood values A and B. A is the likelihood that the non-threat test image is a non-threat, and B is the likelihood that the same image is a threat. If A is greater than B then the plot will show a positive difference meaning the image has greater likelihood of being a non-threat (which is the correct case).
The right-hand side of Figure 4 plots the performance of the same sequence of merchandise images except this time they are drawn with a dark disc marking a threat. In this case A is the likelihood of the image is a threat, and B is the likelihood that it is a non-threat. A positive difference (A greater than B) is what is desired. This difference (shown as the amount of percent difference of A with respect to B) is plotted on the Y axis.
Figure 4: Non-threat and threat classification performance for 53 merchandise images.
Figures 5a-b plot the performance of 264 bundles images. Figure 5a shows the non-threat images, and 5b shows the images with a threat. Just as in Figure 4, the plot shows the difference between the likelihoods of each image being a non-threat and threat. The sign (not the magnitude) of the difference is the most relevant. A positive sign means correct classification.
Figure 5a: classification performance for 264 non-threat bundles images.
(Value above zero means image was classified correctly as a non-threat.)
Figure 5b: classification performance for 264 threat bundles images.
(Value above zero means image was classified correctly as a threat.)
In Figure 5b we see a majority of the 264 bundles threat images were classified incorrectly because most of the points have negative value. We reason this is due the the dense opacity present in the bundles images. The opacity often prevents the threat to show (or hinders its contrast with its surroundings), thus making the edge-detected threat less distinguishable with the regular non-threat training images. However from Figure 5a we can also see that a majority of the non-threat bundle images were distinguished correctly with most points having positive value (i.e., had a larger likelihood of being a non-threat vs. a threat).
Figure 6 shows histograms of likelihood values for 382 images taken across all image categories. The histograms in Figures 6-a,b,c show likelihoods of the images compared against the bundles, merchandise, and vehicle training, respectively. The left and right plots in each figure are for the non-threat and threat versions of the images. The likelihoods of non-threat images are computed using non-threat training. Threat likelihoods are computed using threat training.
Figure 6a: Bundle likelihood histogram for 382 non-threat (left) and 382 threat (right) images.
(X axis is likelihood range)
(Y axis is # of images)
Figure 6b: Merchandise likelihood histogram for 382 non-threat (left) and 382 threat (right) images.
Figure 6c: Vehicle likelihood histogram for 382 non-threat (left) and 382 threat (right) images.
The bundles likelihoods (Figure 6a) have the largest maximum value of 130,731 meaning that the bundles training set produces a stronger likelihood classification when used against all the images. Indeed, all the images in Figure 6a with likelihoods above 100,000 are bundles. In Figure 6b, the merchandise training set is not a robust as the bundles. Hence we found a lot of bundle images distributed in the upper likelihoods of Figure 6b. However, no merchandise images were in the negative to low likelihood range. For the vehicle histogram in Figure 6c, we found that some merchandise images were located near the peak of likelihood range. This is likely due to similarity between the structural elements of merchandise contents with that of vehicles. Finally no vehicle images were present in the negative to low likelihood range in Figure 6c.
Other features were extracted from images: average pixel intensity, opacity (i.e., amount of image darkness) with intensity below 10% of white intensity, and amount of transparency with intensities greater than 90% of white. Examples, of opaque images are illustrated in Figure 7a-b. These feature points can used used in later work to compliment the VBFA process and give weight in classifying images that possess similar pixel intensity statistics.
Figure 7a: Examples of a highly opaque image at 68% (upper right) and a low opaque image at approximately 0.067 % (lower left).
[pic]
Figure 7b: Examples of a mid-level transparent radiographic image at 0.388 (upper right) and highly transparent image at approximately 0.15 (lower left).
Technical Challenges:
Technical challenges posed in analyzing radiographic cargo images are noisy or non-existent signals, Special Nuclear Materials (SNM) threats surrounded by a variety of material and shapes, rendering identification of the SNM threat very difficult (Aufderheide et al., 2007).
Construction of Region Of Interest (ROI) images showing just the cargo container (ie, free of areas showing the truck or outside the container) will be necessary as a pre-processing step before use in a production quality VBFA classification system. Snakes or active contour models will need to be applied to images to bound just the cargo container before being given to the edge detection process.
Highlights
We have a Bayesian Factor Analysis system that can be trained to distinguish threat versus non-threating features present in radiographic images of cargo containers. Because VBFA is sensitive to subtle object features which might otherwise go unnoticed, the classification system acts as a force multiplier for increasing the throughput and security of cargo screening. The system is coupled with a simple graphical user-interface letting the user browse and select images both for training and for classification. So far, this VBFA study has been done with edge detected image data. The training and testing can also be applied to 1-D histogram data of the images rather than the 2-D data of the images themselves.
Major Accomplishments
Currently, we have generated the base implementation standalone algorithms utilizing Bayesian network models and radiograph image visualization as intended in the original project plan. The testbed components have been designed and reconciled to work on single-energy radiographs. More extensive machine learning algorithms and I/O modifications to support programs in side of Engineering Systems for Knowledge and Inference are in progress at the present time such as the National Ignition Facility Automatic Alignment (NIFAA) program for beam commissioning (Candy, McClay, Awwal, 2005). The need for this algorithm testbed is substantial, as LLNL programs continue to invest time and money in machine learning and Bayesian methodologies (computational inference models) to support knowledge discovery. Renewal of this project will ensure that tools are available to help Homeland Security analysts use computational inference models more efficiently and effectively on single-energy radiographic cargo images.
References
1. H Attias. ICA, graphical models, and variational methods. In Independent Component Analysis: Principles and Practice (eds: S Roberts, R Everson), 95-112, 2001. Cambridge UP.
2. H Attias. Learning in high dimensions: modular mixture models. Proceedings of the 8th International Conference on Artificial Intelligence and Statistics, 144-148, 2001.
3. H Attias. A variational Bayesian framework for graphical models. Advances in Neural Information Processing Systems 12, 209-215, 2000.
4. S. Ogorodnikov and V. Petrunin. Processing of interlaced images in 4-10 MeV dual energy customs system for material recognition. Physical Review Special Topics – Accelerators and Beams, Volume 5.
5. M. Aufderheide, H. Martz, Alan Ross, D. Slaughter, P. Sokkappa, D. Schneberk, R. Wheeler, Workshop Summary on Defining SNM and RDD Benchmarks and Cargo Scenes for System Performance of X-ray Radiography Non-Intrusive Inspection Systems, February 2007.
-----------------------
A sequence of 53 non-threat merchandise images along the X axis. Two likelihoods are computed for each image: one if the image is a non-threat, and the other if it is a threat. The difference value between the two likelihoods is on the Y axis. Each point above the black center line means the image is more likely to be a non-threat.
The same sequence of 53 merchandise images with a threat added. Two likelihoods are computed for each image: one if the image is a threat, and the other if it is a non-threat. The percent difference between the two is show on the Y axis. A value above zero mean the image is more likely to be a threat.
