A connectionist approach for thresholding

Thresholding is a necessary and useful step in many applications of image processing. The general process of thresholding is first to select several gray levels, or thresholds, then use these values to classify the pixels into several subranges. Previous methods for selecting thresholds are usually designed based on assumed distributions of pixels or some sort of heuristics. It is difficult to apply any of these methods when the domain of images is changed. There is a need for seeking a more flexible and robust technique in such situation. The paper presents a connectionist approach for learning and selecting thresholds by using the Kohonen algorithm which is an unsupervised neural network. The approach is able to find thresholds for classifying images without a teacher. Experimental results show that the approach is promising