| Summary: | This thesis reports the development and analysis of feature based synthesis of transmission X-ray images. The synthetic imagery is formed through matching and morphing or warping line-scan format images produced by a novel multi-view X-ray machine. In this way video type sequences, which periodically alternate between synthetic and detector based views, may be formed. The purpose of these sequences is to provide depth from motion or kinetic depth effect (KDE) in a visual display; while the role of the synthesis is to reduce the total number of detector arrays, associated collimators and X-ray flux per inspection. A specific challenge is to explore the bounds for producing synthetic imagery that can be seamlessly introduced into the resultant sequences. This work is distinct from the image collection and display technique, termed KDEX, previously undertaken by the Imaging Science Group at NTU. The ultimate aim of the research programme in collaboration with The UK Home Office and The US Dept. of Homeland Security is to enhance the detection and identification of threats in X-ray scans of luggage. A multi-view „KDEX scanner‟ was employed to collect greyscale and colour coded image sequences of 30 different bags; each sequence comprised of 7 perspective views separated from one another by 10. This imagery was organised and stored in a database to enable a coherent series of experiments to be conducted. Corresponding features in sequential pairs of images, at various different angular separations, were identified by applying a scale invariant feature transform (SIFT). The SIFT was additionally bounded by epipolar-line and disparity window criteria. Experiments revealed that around 37% of unfaithful matches were removed. This approach was applied to greyscale imagery and colour coded materials discriminated views, produced by a dual-energy X-ray technique. Two competing approaches namely, direct matching versus material segmentation were developed and tested. The „direct method‟ provided relatively enhanced matching performance. The SIFT matching results are employed by a novel material based morphing to synthesise colour coded X-ray imagery. The technique was designed to conserve the colour coding scheme employed to represent the three different material classes. Error analysis revealed the presence of blurring and ghosting artefacts in the resultant synthetic views. In response a competing translation based image synthesis algorithm was proposed. This rival method reduced the blurring and ghosting artefacts to zero but at the expense of some shape distortion. However, it did provide marginally improved performance when the angular separation between generating views exceeded 50. The fidelity of the synthetic imagery degraded as the angular separation between perspective views was increased. This finding is consistent with the commensurate increase in disparity and therefore dissimilarity between perspective views. The best fidelity intermediary views recorded in this research occurred at 20, which is the minimum angular separation of interest. This result is highly significant as for a given 3D or KDE imaging performance it can reduce the total number of X-ray sensor arrays, collimators and X-ray beams required by a factor of two.
|
|---|
