This paper presents the robust colour image segmentation algorithm which can be used extensively in automated optical in- spection system on the printed circuits boards (PCB) assembly line. The robustness of the presented algorithm is increased by low computational effort what dedicates this approach to on-line vision systems explored in high-speed production lines. The second advantage of the proposed algorithm is the non-parametric solution which is invariant to changes of light and colour of the inspected components. This condition should be fulfilled with the vectoral imaging approach where the most important task is the segmentation process which divides image space into disjoint regions with similar colour or texture. gle components R, G, and B or their transformations (lin- ear/non-linear) are explored independently of the others. The results obtained from these single components are often dif- ficult for unambiguous interpretation. To overcome this shortcoming, the analysis of the multidimensional feature space is required. The analysis of this space is very often time-consuming particularly in case of computer vision problems where a large imass of the datai is connected to the image. The approach proposed in this paper is based on one of the widely used techniques for feature space analysis called clustering. The points in the feature space correspond to feature vectors describing points (pixels) in the image space. Colour images can be characterized by three-dimen- sional histograms, and thus, we have three-dimensional fea- ture space in which clusters or modes correspond to pixels with similar colours called image regions. One of the drawbacks of the feature space clustering is that the cluster analysis does not utilise any spatial infor- mation from the image space that considers the spatial rela- tion among pixels. The second and the most serious prob- lem is difficult determination of the number of clusters (im- age regions) in the unsupervised clustering scheme which is known as cluster validity. The next, crucial problem for the colour image segmentation is the selection of the appro- priate colour space (feature space). RGB colour space is not appropriate for colour space clustering because of the high correlation of among R, G, and B, therefore an object with a uniform colour but different intensities could be seg- mented into different objects. In other words, a colour im- age with shadows or some shading cannot be segmented properly in RGB colour space. The proposed algorithm is based on the modified HSI colour space. The modification of HSI colour space has been already presented in Ref. 3