JPEG 2000 scalar quantization using an optimally frequency localized modulated lapped transform

JPEG 2000 is a lossy integer-to-integer transform-based compression method that first quantizes the separable 2-D wavelet transform coefficients, then entropy codes them. The image is restored by performing an inverse wavelet transform on the dequantized and decoded coefficients. In this paper we consider closely related coding strategies using modulated lapped transforms rather than wavelet transforms, where performance is studied as a function of time (space) localization, frequency localization, and joint localization. In terms of reconstruction error at a given coding gain (quantization step size), we find that the modulated lapped transform admitting the best frequency localization offers superior performance relative to both other lapped transforms and orthogonal wavelet transforms.