Image compression using auto-associative neural network and embedded zero-tree coding

This paper presents an image compression method using auto-associative neural network and embedded zero-tree coding. The role of the neural network (NN) is to decompose the image stage by stage, which enables analysis similar to wavelet decomposition. This works on the principle of principal component extraction (PCE). Network training is achieved through a recursive least squares (RLS) algorithm. The coefficients are arranged in a four-quadrant sub-band structure. The zero-tree coding algorithm is employed to quantize the coefficients. The system outperforms the embedded zero-tree wavelet scheme in a rate-distortion sense, with best perceptual quality for a given compression ratio.