Colour map image segmentation based on supervised and unsupervised learning techniques

Image segmentation is a very important stage in any image analysis or computer vision system. Map images are considered to be among the most complex of images. The segmentation of colour map images is a difficult problem. In this thesis, four segmentation techniques are presented to extract characters and lines from colour geographic map images. There are: conventional adaptive thresholding, the supervised-learning neural network, the unsupervised fuzzy c—means clustering and nearest-prototype rule, and the combined supervised and unsupervised techniques. In the conventional adaptive thresholding technique, images are divided into subimages. For each bimodal histogram subimage, a threshold is located at the valley of the histogram using an automated histogram analysis technique. A threshold value is obtained for each pixel of the image by interpolation of the thresholds. The image is then segmented by the different thresholds at each pixel. In the supervised-learning neural network based technique, a neural network is first trained with feature values using known character and line pixels and background pixels, and is then used for classification. The image segmentation problem is treated as a pattern classification process and the neural network classifier is used to generate non—linear decision regions to separate the foreground and background of an image that containing a number of nonuniform regions with different colours. In the unsupervised fuzzy clustering and nearest-prototype rule based technique, segmentation is also considered as a process of pixel classification. A set of prototypes is generated using the fuzzy c—means clustering algorithm on the training areas selected from different colour map images, and each pixel of the image is classified into character and line class or background class according to the nearest—prototype rule. In the combined supervised and unsupervised technique, training samples are generated by the unsupervised fuzzy clustering technique applied to subimages and by randomly choosing pixels in the low contrast areas. A supervised learning based multi-layer neural network is trained for classifying character and line pixels and background pixels. These four techniques are applied to many colour geographic map images containing English, Japanese and Chinese characters with different printing styles. The conventional adaptive threshold technique does not work well. The proposed supervised and unsupervised techniques have achieved satisfactory segmentation results although some very low contrast areas require improvement in the unsupervised technique. The combined technique is a way of enchancing the performance of the supervised technique, and it has yielded good segmentation results