A predictive vector quantization scheme exploiting the intervector correlations of adjacent blocks (vectors) of pixels is developed. The model presented utilizes the statistical dependencies of the previously encoded pairs of adjacent blocks to predict future
blocks of picture elements. The state of the vector predictor is represented by a subcodebook composed of a finite number of code vectors. These patterns constitute the most probable candidates for encoding purposes. The entries of the subcodebook are replenished at each state employing interblock dependencies. To further increase the performance of the quantizer, the difference between predicted pixels and original image samples was vector quantized in the second stage. Excellent subjective performance and SNRs were achieved for monochrome still images, while the range of the bit rates was lower than those of memoryless vector quantization schemes.