Image data compression using a neural network model

Data compression and generalization capabilities are important for neural network models as learning machines. From this point of view, the image data compression characteristics of a neural network model are examined. The applied network model is a feedforward-type, three-layered network with the backpropagation learning algorithm. The implementation of this model on a hypercube parallel computer and its computation performance are described. Image data compression, generalization, and quantization characteristics are examined experimentally. Effects of learning using the discrete cosine transformation coefficients as initial connection weights are shown experimentally.<>