Theory and applications of estimated spectrum adaptive postfilter

This thesis investigates a number of image-adaptive, JPEG-compatible postfiltering and pre-post filtering methods designed to minimize the DCT blocking distortion. A pre-post filtering system uses inverse pair filters for high frequency preemphasis before encoding and high frequency deemphasis after decoding. The inverse pair pre-post filters are related by the inverse relationship P(o) = 1/D(o), where P is the preemphasis filter and D is the deemphasis postfilter. A postfiltering system does not preemphasize the image before encoding. These techniques minimize the mean square error (MSE), improve the objective and subjective quality of low bit rate JPEG gray-scale images, and simultaneously enhance their perceptual visual quality. All the variants of the algorithms presented minimize the MSE below the level of baseline JPEG image compression, which is used as our comparison basis for similar bit rates. Convergence to a unique MMSE is possible for fixed quantization matrices, however, it cannot be guaranteed when image-adaptive quantization is jointly optimized under pre-post filtering.; We develop the theoretical basis of the Estimated Spectrum Adaptive Postfilter (ESAP) algorithm. ESAP is the main postfiltering algorithm used to minimize DCT blocking. ESAP utilizes either the default JPEG quantization table or image-adaptive DCT quantization matrices created in a preprocessing stage prior to image compression. At the decoder, the algorithm estimates 2-D pixel-adaptive bandwidths directly from the dequantized DCT coefficients to control a 2-D spatially-adaptive non-linear postfilter. Consistent with the human visual system tolerance to quantization errors in the high frequency regions, the algorithm performs directional filtering parallel to the edges and no filtering across the edges, subject to filter design constraints. Postfiltered images show minimal blurring of their true edges while blocking is significantly removed. ESAP relies on a DFT analysis of the DCT and is compliant with the coded stream syntax of the Independent JPEG Group (IJG) Version 5b Software.; Additionally, this thesis explores several other variants of the Estimated Spectrum Adaptive Postfilter applied to non-DCT coders such as vector quantization (VQ), subband coders (SBC), and Projection Onto Convex Sets (POCS). We compare the performance of these basic image coding methods against the same extended coders used with image preprocessing, ESAP postprocessing, or both, in a coder-compliant manner. These methods may or may not incorporate image-adaptive quantization and pre-post filtering. We also extend the concepts of the dbx audio noise reduction systems to model and demonstrate an Iterative Pre-post Filter (IPF). The IPF is applied to JPEG and to Set Partitioning In Hierarchical Trees (SPIHT) octave-band subband coders. We also process JPEG color images, synthetic aperture radar (SAR) images, and image sequences.; Typical PSNR improvement depends on the image, the encoding method, and the bit rate, and can range between 0.5--3.2 dB over baseline JPEG for 512 x 512 8-BPP gray-scale images. A comparison including all the treated techniques is presented at the conclusion of the thesis.