An extended multi-valued exponential bi-directional associative memory (EMV-eBAM) model is presented in this paper based on Wang’s MV-eBAM model, which is a special case of EMV-eBAM (extended MV-eBAM). EMV-eBAM has higher storage capacity and stronger error-correcting capability. Using these performances in image compression, a novel image compression algorithm based on EMV-eBAM is proposed. In noise-free situations, this algorithm can acquire similar performances compared with vector quantization algorithm (VQ). However, in noisy context, this algorithm possesses strong noise-restraining capability. The experimental results show that while VQ amplified 5% random noises appended in the image, this algorithm can hold back nearly all noises and acquire similar performances as in noise-free context. Furthermore, in transmitting there may be some errors in the channel, in this situation, this algorithm has much better error-correcting capability than the result by using the cyclic encoding method, so this algorithm is a robust image compression algorithm.