A method for producing a compressed bit-stream from a digital image includes
the
steps of a) processing the digital image to produce a main subject belief map containing
a continuum of belief values relating to the importance of subject and background
regions in the image, b) performing a spatio-frequency transformation on the digital
image to produce an array of transform coefficients, c) deriving a distortion-weighting
factor for each transform coefficient from the belief map, and d) producing a compressed
bit stream using an image compression system that is responsive to the distortion-weighting
factors. The specific image compression system may be selected from a variety of
image compression systems, including JPEG compression, JPEG2000 compression or
vector quantization. The production of the main subject belief map involves the
extraction of at least one structural saliency feature and at least one semantic
saliency feature from the image, and the integration of the structural saliency
feature and the semantic feature using a probabilistic reasoning engine into an
estimate of a belief that a region of the digital image is the main subject.
