INTEGRATION OF LINGUISTIC KNOWLEDGE FOR COLOUR IMAGE SEGMENTATION T. CARRON, P. LAMBERT Laboratoire d’Automatique et de MicroInformatique Industrielle LAMII/CESALP - Université de Savoie - B.P 806 - F.74016 Annecy Cedex (France) (CNRS G1047 - Information-Signal-Image) e-mail: carron@univ-savoie.fr - lambert@univ-savoie.fr ABSTRACT The Hue, Chroma, Intensity (HCI) space is well suited to colour images segmentation processing. In this paper, we used fuzzy logic for integrating specific knowledge of the Hue component. Based upon several linguistic rules which built a symbolic cooperation between Hue and Intensity according to Chroma, a region growing segmentation with fuzzy aggregation is proposed. This fuzzy segmentation is compared with a technique using a Fuzzy C-Means algorithm in different colour spaces. Key words: Colour segmentation - Fuzzy sets - Knowledge integration. 1 INTRODUCTION For a few years, there is a growing interest in using a linguistic approach associated with fuzzy subsets theory for image processing. This is true in high level processing (inter- pretation) but also in low-level processing (filtering [1], segmentation [2][3] or edge detection- in [8] a fuzzy colour edge extractor by If-Then rules operating in Hue, Chroma, Intensity (HCI) space has been presented). In the case of colour image segmentation, the Fuzzy C-Means (FCM) algorithm [4] is widely used for clustering [5][6][7]. However, it is also widely recognized that the clustering technique based on FCM suffers from problems related to adjacent clusters frequently overlapped in colour space, inducing incorrect pixel classification. Furthermore, clustering is more difficult when the number of clusters is a priori unknown, which is typical in segmentation appli- cation. An other inconvenient of these methods is that they don’t take into account the specificity of the colour image. There are different methods to get HCI representation from RGB space. In this paper, the colour used features are calcu- lated by the following formula: Y C 1 C 2 0,33 0,33 0,33 1 0,50 – 0,50 – 0 0,87 – 0,87 – R G B × = I Y = ❏ C C 1 2 C 2 2 + = ❏ C 2 0 > H C 2 C ∕ ( 29 acos = H 2 Π C 2 C ∕ ( 29 acos – = ❏ if then else Of course the first interest of this space is that it is more suited to color perception than the RGB space. The second interest is that the noise sensitivity presents interesting properties [9]. So, Hue noise sensitivity, and consequently Hue relevance, is depending on Chroma level. This can be traduced in linguistic rules by: ❍ If the Chroma is low then the Hue is hence irrelevant, ❍ If the Chroma is medium then the Hue is weakly relevant ❍ If the Chroma is high then the Hue is very relevant and its sensitivity to noise is lower than that of the Intensity. Thus, a segmentation algorithm in HCI space can used this specific characteristic of Hue by realizing a cooperation between Hue and Intensity according to the Chroma: ❍ If the Chroma is low then Intensity is used, ❍ If the Chroma is high then Hue is used, ❍ If the Chroma is medium then Hue and Intensity are jointly used. So, in a region-growing segmentation, the part of the Hue component in the decision of aggregation will be nil, lower, identical, or more important than the one of the Intensity. The basis of this work is the definition of fuzzy sets charac- terizing the three numerical magnitudes used in the proposed colour segmentation method, i.e. the Chroma levels of two neighbouring pixels and their Hue and Intensity difference. Then several linguistic rules are defined to integrate the specific characteristic of the Hue. These rules create a fuzzy homogeneity criterion in order to realize the fuzzy aggre- gation of these pixels. In a first part a general description of the fuzzy aggregation is done. Then the fuzzy partitions of the domains are presented. The linguistic rules for the segmentation Hue- Intensity / Chroma are described in the section 4. Finally, a comparative application on biomedical images is performed. 2 GENERAL DESCRIPTION OF THE ALGORITHM The segmentation technique used in this paper is a local iterative region-growing adapted from gray level images segmentation (Blob Colouring with a L-shaped template of three pixels) [10]. The basic idea is the definition of homoge- neity criterion between two neighbouring pixels. The