Artificial neural network for discrete cosine transform and image compression
K.S. Ng, L.M. Cheng · 2002
Efficient adaptive image compression using a structured artificial neural network (ANN) is described. An image is first divided into a series of sub blocks with size 8/spl times/8 pixels. Then each of them is transformed by a discrete cosine transform (DCT) using a structured ANN. Then, all the sub blocks are sorted into 4 classes using another layer of structured ANN, according to their level of activity within each sub block. Adaptivity is provided by assigning bits between classes. The neural network used is a structured one instead of a fully connected one, so that convergency and speed of learning are dramatically improved. Each subnetwork is trained and tested independently. Excellent performance is achieved, in comparison to traditional fully connected neural network image compression methods.