Using feed forward multilayer neural network and vector quantization as an image data compression technique

摘要

Single hidden layer feed forward neural networks with different number of hidden neurons are used for image data compression. A subimage of size 4/spl times/4 pixels forms the input vector of size 16 pixels. The hidden vector, which is the output of the hidden layer whose size is smaller than that of the input vector represents the compressed form of the image data. The hidden vector is transmitted by a vector quantizer with codebook of 256 codevectors which corresponds to a bit rate of 0.5 bit/pixel. The reconstructed subimage, at the receiver, is obtained from the output layer which consists of 16 neurons. Good reconstructed images are obtained with a PSNR of about 30 dB for the in-training set image (Lena) and 27 dB for the outside-training set image (Boats).