In general, the 3D data contained in a stereoscopic image is difficult to extract and sparse in comparison to the number of pixels contained in the image pair. The lack of reliable data makes computational reconstruction for solid image modelling highly problematic and error prone. To address this problem a novel object recognition process that is able to Identify primitive shapes by their geometric, colour and intensity signature has been developed. This information is used to 'look up' the recognised shape in a database to enable a solid image representation of the primitive form to be available for redisplay from different viewpoints. An experimental binocular stereoscopic dual-energy x-ray system previously developed by the university team for aviation security screening applications provided the image data for the empirical analysis. The dual-energy x-ray sensors enable materials discrimination information to be colour encoded in the resultant images: organic materials are displayed in orange, metallic in blue and a mixture of organic and metallic in green. Several new feature extraction methods have been developed to overcome shortcomings found in existing methods. The algorithms and databases are compiled Into a complex program suite that performs the automatic detection and recognition of defined objects. The program has built in flexibility than enables testing of a wide range of feature extraction methods. Detected objects are subsequently 'reconstructed' and displayed as solid image models that facilitate viewing from different angles. Information stored in a database is used in hierarchical associative comparison of image structures as the basis for shape recognition. Initially the empirical study of this approach utilised a range of test objects composed of simple primitive forms: cube, sphere and cylinder. Later experiments concentrated on the analysis of composite objects that form a humanoid shape or 'ski man'. Image processing techniques utilising neural networks, morphology, and wavelets have been developed to tackle the inherent uncertainty in the image data. The final algorithmic approach was influenced by a preliminary study of the human visual system that enabled the development of a well-structured information extraction method and a robust image feature classification. This work is an initial step towards producing 3D image models for a range of 'threat items' of interest in security x-ray screening. The techniques developed are applicable to the latest generation of multiple view dual-energy x-ray systems currently under development by the University team. This research program was undertaken in collaboration with the Police Scientific Development Branch (PSDB), Sandridge, UK, part of the Home Office Science and Technology Group.