This paper introduces a new classification method, termed vector-scaler classification, where each DCT (ac) vector is indexed using vector quantization (VQ) on the energy vector of the corresponding DCT subvectors, and the subvectors each are then classified according to the class assignment for the VQ. Vector-scalar classification achieves an average classification gain about 6.5 dB higher than block classification proposed by Chen and Smith (1977), since VQ has high clustering performance and classification of subvectors can exploit compactness of DCT coefficients more efficiently. The complexity of this method is relatively low. For coding small images, a fixed-rate DCT coding system developed using vector-scaler classification is competitive with some best variable-rate coding systems reported in the literature, and the classification can capture most of the performance gain obtained by entropy coding.