Image segmentation is one of the most important steps in computerized systems for analyzing geographic map images. We present a segmentation technique, based on fuzzy rules derived from the K-means clusters, that is aimed at achieving human-like performance. In this technique, the K-means clustering algorithm is first used to obtain mixed-class clusters of training examples, whose centers and variances are then used to determine membership functions. Based on the derived membership functions, fuzzy rules are learned from the K-means cluster centers. In the map image segmentation, we make use of three features, difference intensity, standard deviation, and a measure of the local contrast, to classify each pixel to the foreground, which consists of character and fine patterns, and to the background. A centroid defuzzification algorithm is adopted in the classification step. Experimental results on a database of 22 gray-scale map images show that the technique achieves good and reliable results, and is compared favorably with an adaptive thresholding method. By using K-means clustering, we can build a segmentation system of fewer rules that achieves a segmentation quality similar to that of using the uniformly distributed triangular membership functions with the fuzzy rules learned from all the training examples.