Classification preserving image transform compression

A data compression device comprising:

• an input configured to receive image data corresponding to an input image comprising a plurality of blocks;
• a statistical mode selector coupled to the input and configured to generate a histogram corresponding to a color or an intensity of pixels of the input image, to determine a statistical mode of pixels based on the histogram, and to generate a plurality of classifications of pixels based on the determined statistical mode, wherein the plurality of classifications of pixels comprises at least two color classifications of pixels representing color values selected to preserve at least one of a mean and a variance of each block of the input image;
• a bitmask generator coupled to the statistical mode selector and configured to receive the plurality of classifications of pixels, and to generate a bitmask based on a plurality of classifications of pixels;
• a processor coupled to the bitmask generator and configured to run a first sparse transform in a spatial domain on a first color classification of pixels of the at least two color classifications of pixels to generate a first transform coefficients and a second sparse transform on a second color classification of pixels of the at least two color classifications of pixels to generate a second transform coefficients, the second sparse transform being different from the first sparse transform;
• and
• a scalar quantizer coupled to the processor and configured to generate quantized transform coefficients.

A method of color image compression, the method comprising:

• receiving image data corresponding to a plurality of pixels and having a plurality of blocks;
• classifying the pixels based on the image data to generate a plurality of classifications of the pixels, wherein classifying the pixels comprises generating a histogram corresponding to features of the pixels of an input image, determining a statistical mode of the pixels based on the histogram, and classifying the pixels representing the statistical mode as one of the classifications of the pixels, wherein at least two pixel classifications representing color values may be selected to preserve at least one of a mean and a variance of each block of the input image;
• running a first sparse transform for a first color classification of pixels of the at least two pixel classifications to generate a first transform coefficients and a second sparse transform for a second color classification of pixels of the at least two pixel classifications to generate a second transform coefficients, the second sparse transform being different from the first sparse transform;
• and
• running a scalar quantizer for each transform coefficient,
• wherein sparseness of the sparse transform corresponds to a spatial domain of the pixels corresponding to the image data.