Improved compression can be attained by selecting the transform adaptively to the signal. Some fixed signal bases are supported by statistical image models. The justifications are as follows: (1) images generally have high pixel correlation - DCT decorrelates image data well, (2) images have most energy at low frequencies - the wavelet transform provides increasing frequency resolution at lower frequencies, (3) images are non-stationary - block-based transforms compensate for long-term non-stationarity. However, in many cases real images differ significantly from the models. As such, the non-adaptive transforms provide only suboptimal organization of the image data. The incorporation of a basis selection mechanism for the first stage improves the potential for compression in the system.