IMAGE COMPRESSION USING HYBRID NEURAL NETWORKS COMBINING THE AUTOASSOCIATIVE MULTILAYER PERCEPTRON AND THE SELF-ORGANIZING FEATURE MAP

A new image compression technique is presented using hybrid neural networks that combine two different learning networks, the auto-associative multi-layer perceptron (AMLP) and the self-organizing feature map (SOFM). The neural networks simultaneously perform dimensionality reduction with the AMLP and categorization with the SOFM to compress image date. Two hybrid neural networks forming parallel and serial architectures are examined through theoretical analysis and computer simulation. The parallel structure network reduces the dimensionality of input pattern vectors by mapping them to different hidden layers of the AMLP selected by winner-take-all units of the SOFM. The serial structure network categorizes the input pattern vectors into several classes representing prototype vectors. Both the serial and parallel structures are combinations of the AMLP and SOFM networks. These hybrid neural networks achieve clear performance improvement with respect to decoded picture guality and compression ratios, compared to existing image compression techniques.