Robust image coding with perceptual-based scalability

Abstract

Summary form only given. We present a multiresolution-based image coding technique that achieves high visual quality through perceptual-based scalability and robustness to transmission errors. To achieve perceptual coding, the image is first segmented at a block level (16/spl times/16) into smooth, edge, and highly-detailed regions, using the Holder regularity property of the wavelet coefficients as well as their distributions. The activity classifications are used when coding the high-frequency wavelet coefficients. The image is compressed by first performing a 3-level hierarchical decomposition, yielding 10 subbands which are coded independently. The LL band is coded using reconstruction-optimized lapped orthogonal transforms, followed by quantization, runlength encoding, and Huffman coding. The high-frequency coefficients corresponding to the smooth regions are quantized to zero. The high-frequency coefficients corresponding to the edge regions are uniformly quantized, to maintain Holder regularity and sharpness of the edges, while those corresponding to the highly-detailed regions are quantized with a modified uniform quantizer with a dead zone. Bits are allocated based on the scale and orientation selectivity of each high-frequency subband as well as the activity regions inside each band corresponding to the edge and highly-detailed regions of the image. The quantized high-frequency bands are then run-length encoded.