Search

This thesis reports an investigation into dual-energy X-ray discrimination techniques. These techniques are designed to provide colour-coded materials discrimination information in a sequence of perspective images exhibiting sequential parallax. The methods developed are combined with a novel 3D imaging technique employing depth from motion or kinetic depth effect (KDE). This technique when applied to X-ray images is termed KDEX imaging and was developed previously by the university team for luggage screening applications at security checkpoints. A primary motivation for this research is that the dual-energy X-ray techniques, which are routinely incorporated into ‘standard’ 2D luggage scanners, provide relatively crude materials discrimination information. In this work it was critical that robust materials discrimination and colour encoding process was implemented as the sequential parallax exhibited by the KDEX imagery may introduce colour changes, due to the different X-ray beam paths associated with each perspective image. Any introduction of ‘colour noise’ into the resultant image sequences could affect the perception of depth and hinder the ongoing assessment of the potential utility of the dual-energy KDEX technique. Two dual-energy discrimination methods have been developed, termed K-II and W-E respectively. Employing the total amount of attenuation measured at each energy level and the weight fraction of layered structures, a combination of the K-II and the W-E techniques enables the computation and extraction of a target objects’ effective atomic number (Zeff) and its surface density (ρS) in the presence of masking layers. These material parameters (Zeff and ρS) together with laminographic layer thickness estimation enable mass density extraction. A series of experiments investigated the computation of Zeff and ρS as a function of system noise and repeatability. The estimation of thickness depends on the depth increment provided by the image capture geometry and the laminographic processing. Within the atomic number range of 6 to 30 and with up to 4 masking layers, the investigated techniques produce an accuracy of Zeff and ρS up to 97% and 95% respectively. The thickness estimation technique provided a relatively high accuracy for object thickness’ greater than 4cm. Although, the measurement accuracy for relatively thin layers is inherently limited by the minimum resolvable depth increment of the image collection geometry. The mass density extracted had an overall accuracy of greater than 90% for well-estimated thicknesses. Implementing a new four colours scheme highlights the presence of potential threat materials in the resultant KDEX imagery. This thesis forms part of a larger programme of activity in collaboration with and funded by the UK Home Office Scientific Development Branch, and the US Department of Homeland Security and the EPSRC.

Object recognition in infrared imagery has important applications, for example, in security and defence, and surveillance, due to the passive night-time and bad weather capabilities of infrared sensors. The objective of this thesis is to find a preferred method for the identification of static targets in single infrared images, concentrating on appearance-based methods. This has included thermal modelling of infrared signatures and the identification of images of different objects with variation in pose and thermal state. This thesis reviews several popular approaches in object recognition In visible and infrared imagery, concentrating on the appearance based approach. Using principal component analysis, the variances among the images are extracted and represented in a low-dimensional feature Eigenspace. Any new image can be projected into the Eigenspace by taking an inner product with the basis. The object of interest can be recognized by a nearest-neighbour classification rule, made more accurate by application of over-sampling to the surface manifold by B-spline surface fitting, and made more efficient by a k-d tree search algorithm. To address the problems of recognizing targets in noisy and cluttered images, we have also employed a random sampling approach that is based on the principle of high-breakdown point estimation. As a final step, a probabilistic framework in employed to improve the recognition rate and give a confidence measure for the result. The probability is determined . by two facts: distance from the Eigenspace and distance in the Eigenspace. Using this probabilistic framework, we set an 'image window' on the test image and adjust the position of the window according to the recognition result in the form of probability. The 'image window' method makes the system able to bear small in-plane transformation of the object in the test image and to recognize poorly segmented test images. We also discuss the possibility of using a non-linear dimensionality reduction method, Isomap, to replace peA as the basis of data decomposition in the appearance based method. Results show that although Isomap has some advantages in separating poses, it does not improve the recognition result using sufficient basis vectors compared to PCA. Therefore, we still use PCA as the basis for dimensionality reduction. A new way of modelling thermal change has been proposed under the framework of an appearance based method. Possible thermal state changes of an object are modelled by several single component changes and the combinations of these changes. Hence, we build an Eigenspace model inwhich each object is represented by several lines (or vectors) in the Eigenspace and each line represents one pose and one thermal state change. Using this model, it is possible to predict subspace projection of changes in thermal state and to recognize new unseen thermal images. Using a recognition algorithm that measures scene to model similarity by the distance between the unknown point and the learnt linear object representation, we are able to show an improvement in the recognition accuracy over the conventional appearance based approach. We have made extensive use of simulated data for both learning and recognising targets by appearance. As we have two degrees of freedom in viewpoint, azimuth and elevation, and several further degrees of freedom in allowing thermal state changes on different parts of the object, we have used as many as 33700 thermal images for a single object in the most extreme case. Hence, it is not feasible to both control the thermal state and acquire infrared data for the complete set of objects and viewpoints in the learning phase. In the recognition phase, we have used simulated data to test the algorithms, but have also embedded simulated vehicles within real infrared image data, a practice which is common in the literature on IR target recognition which is reviewed in Chapter 2. Although the simulation package, CameoSim, has been evaluated in comparison with real data, this is less than ideal, but necessary in the circumstances to evaluate and test the approach.