In order to recognise objects from images of scenes that typically involve overlapping and partial occlusion, traditional computer vision systems have relied on domain knowledge to achieve acceptable performance. However there is much useful structural information about the scene, for example the resolution of figure-ground ambiguity, which can be recovered or at least plausibly postulated in advance of applying domain knowledge. This thesis proposes a generic information theoretic approach to the recognition and attribution of such structure within an image. It reinterprets the grouping process as a model selection process with MDL (minimum description length) as its information criterion. Building on the Gestalt notion of whole-part relations, a computational theory for grouping is proposed with the central idea that the description length of a suitably structured whole entity is shorter than that of its individual parts. The theory is applied in particular to form structural interpretations of images under partial occlusion, prior to the application of domain knowledge. An MDL approach is used to show that increasingly economical structural models (groups) are selected to describe the image data while combining lower level primitives to form higher level structures. From initially fitted segments, progressive groups are formed leading to closed structures that are eventually classified as foreground or background. Results are observed which conform well with human interpretations of the same scenes.