Image compression aims to reduce the amount of data that is stored or transmitted for images. One technique that may be used to this end is vector quantization. Vectors may be used to represent images. Vector quantization reduces the number of vectors required for an image by representing a cluster of similar vectors by one typical vector that is part of a set of vectors referred to as the code book. For compression, for each image vector, only the closest codebook vector is stored or transmitted. For reconstruction, the image vectors are again replaced by the the closest codebook vectors. Hence vector quantization is a lossy compression technique and the quality of the reconstructed image depends strongly on the quality of the codebook. The design of the codebook is therefore an important part of the process. In this thesis we examine three clustering algorithms which can be used for codebook design in image compression: c-means (CM), fuzzy c-means (FCM) and learning vector quantization (LVQ). We give a description of these algorithms and their application to codebook design. Edges are an important part of the visual information contained in an image. It is essential therefore to use codebooks which allow an accurate representation of the edges. One of the shortcomings of using vector quantization is poor edge representation. We therefore carry out experiments using these algorithms to compare their edge preserving qualities. We also investigate the combination of these algorithms with classified vector quantization (CVQ) and the replication method (RM). Both these methods have been suggested as methods for improving edge representation. We use a cross validation approach to estimate the mean squared error to measure the performance of each of the algorithms and the edge preserving methods. The results reflect that the edges are less accurately represented than the non - edge areas when using CM, FCM and LVQ. The advantage of using CVQ is that the time taken for code book design is reduced particularly for CM and FCM. RM is found to be effective where the codebook is trained using a set that has larger proportions of edges than the test set.