Robust segmentation of complex images is a challenging problem. Performance with traditional use of statistical information, such as intensity, first and second derivatives, and local intensity histograms, is often degraded severely by noise. Good results with model-based segmentation approaches are generally sensitive to the precise initialization of the model within the image to be segmented.
The basic idea of this paper is to integrate a set of image attributes into a single, unified framework, such that their complementary information facilitates reliable and robust decisions in image segmentation. This paper proposes a new vector clustering method, called Attributed Vector Quantization (Attributed VQ), for this purpose. The new method can be considered as a variation of the existing weighted vector quantization method, but uses a different weighting scheme from the traditional method, which is clearly motivated by the complex image segmentation problem. Furthermore, the vector quantization process intrinsically produces hierarchically organized structural information that can characterize the pattern of the object to be segmented. We demonstrate the new method with industrial images as the example. We demonstrate, preliminarily, that this approach shows promise in its application to training and recognition in industrial vision systems by requiring minimal user interaction during training, and by leveraging its basis in vector quantization to reduce sensitivity to noise and other anomalies during recognition.