An image compression technique in which patterns identified, the means of separating the image components, the parameterization of the patterns, and the lower level numerical encodings are all designed around a narrow class of images, such as two-dimensional projections of three-dimensional visualizations of data generated by numerical weather simulations. The process analyzes an image in terms of perceptual constructs of the human visual system and searches for patterns among analyzed abstractions of the image. The image is then described in terms of the perceptual constructs and the patterns found among them. The image is re-represented by describing the image as a collection of parameterized versions of the patterns prevalent in that class of image. A resulting description is taken outside of the context of abstract patterns. Redundancies in the description are looked for and the data is re-represented so as to eliminate the redundancies and compress the description.