Graph-theoretical approach to colour picture segmentation and contour classification zyx T. Vlachos A.G. Constantinides zyxwvutsrqpo Indexing terms: Colour picture segmentation, Contour class@cution zyxwvutsrqpo Abstract: The segmentation of colour pictures in a graph-theoretical context is considered. The procedure aims at identifying, extracting and classifying visually important features on the image plane, such as regions of homogeneous colour and chromatic transitions. Well established principles of colour theory and graph theory are combined to obtain a unified representation of a colour picture. The picture is represented by means of a weighted graph, constructed so as to reflect the specification of the colour space employed as well as important relationships between picture elements. A spanning tree of the graph is obtained by iteratively minimising a spe- cific picture distortion measure. zyxwvutsrq This tree structure describes a hierarchy of partitions on the image plane. Each partition comprises disjoint regions containing elements with similar attribute. Due to the fact that region identification and edge detec- tion form dual problems from the graph- theoretical viewpoint, region contours defined by such partitions form a hierarchy. To avoid artifi- cial contouring, a specific type of artefact intro- duced by the segmentation algorithm, the use of higher level information, is considered. It is shown that, when texture (which is taken into account at an intermediate stage of picture segmentation) is combined with colour as joint similarity attributes of regions, an improved hierarchical description of contours is possible. This facilitates the progres- sive elimination of the undesirable contours and leads to the visual enhancement of the segmen- tation obtained. 1 Introduction Segmentation has been regarded as a useful tool in image analysis and low level processing and has been employed in a wide range of machine vision applications [l]. More importantly, it has been identified as the critical com- ponent of a hierarchical picture description that takes into account the principles which underlie human percep- tion, and in that capacity it has been considered for object recognition, scene analysis and bandwidth com- pression tasks 123. Paper 93421 zyxwvutsrqpo (M), first received 29th June and in revised form 27th November 1992 The authors are with the Signal Processing Section, Department of Electrical and Electronic Engineering, Imperial College of Science, Technology and Medicine, Exhibition Road, London, SW7 2BT. United Kingdom 36 Most segmentation techniques described in the liter- ature [3, 41 are not capable of capturing the actual level of significance of the features contained in a scene, either because they do not take into account universal picture information or because they ignore important spatial relationships among picture elements. Moreover, many popular techniques impose external geometric constraints on the structure of the regions obtained, such as in the form of regular boundaries, to obtain a solution to an otherwise almost intractable computational problem. Additionally, segmentation techniques based on edge detection suffer from multiple definition of borders and open contouring and hence need to be supplemented by operations such as thinning and linking. This paper describes a segmentation technique that avoids the problems commonly associated with existing segmentation algorithms. It takes into account global image properties and local spatial relationships among image elements without imposing any geometric con- straints in the structure of the segmented image obtained. The basis of this technique is obtained by extending the work described [SI to take into account higher level information contained in an image such as colour and texture. Chromatic features obtained from the original image are used, to partition the image into disjoint regions of uniform colour. The enhancement of the semantic information conveyed by the segmented image is pursued by the identification of appropriate textural features. The use of the latter facilitates the elimination of a certain class of irrelevant features contained in the par- tition. 2 We consider colour images which have been sampled both spatially and in amplitude. Such an image zyx f can be specified as a two-dimensional array of zyxwv M x N discrete amplitude elements. A more compact representation is obtained by raster scanning the array so that Pcan be Graph representation of colour pictures This work was supported by the Commission of the European Communities in the framework of the ESPRIT programme under sectoral grant B/890376. The authors are grateful to Philips Research Laboratories, Redhill, UK, for providing us with the source material. It was a pleasure to collaborate on parts of this work with Dr. O.J. Morris of PRL and D. Moran formerly with PRL, and at present with Image Recognition Systems, Warrington, UK. IEE PROCEEDINGS-I, Vol. zyxwvu 140, No. I, FEBRUARY 1993