Rounded reduced-difference pyramid for semilossy image coding

An extension of the reduced-difference pyramid (RDP) algorithm is derived and proposed for lossy image compression with upper-bounded maximum error (semilossy). Quantization has been introduced into the basic RDP in two different ways: approximations are made either before or after taking the differences among pairs of adjacent nodes. The resulting peak-error (PE) propagation along the pyramid is theoretically investigated, and quantization strategies are suggested. Tests on the standard grey-scale Lena image show significant and steady rate-distortion improvements over the original RDP. As few fixed-point operations are required per pixel, on-line compression/decompression is feasible on general-purpose computers.