Image compression is an essential task for image storage and transmission. We propose a compression technique in which an MLP predictor takes advantage of the topological properties of the Kohonen algorithm. The Kohonen algorithm creates a code-book which is used for Vector Quantization of the source image. Then,an MLP is trained to predict references to code-book, allowing further compression. Even with difficult images, the result is a reduction of 15% to 20% of the bit rate compared with classical Vector Quantization techniques, for the same quality of decoded images.