Low-Complexity Predictive Coding of Color Video Using Optimum Switched Two-Level Quantizers

Two-Level Quantizers Frank Hartung and Bernd Girod Telecommunications Institute University of Erlangen-Nuremberg Cauerstrasse 7, 91058 Erlangen, Germany fhartung, [email protected] ABSTRACT Di erential pulse code modulation (DPCM) and its signal-adaptive derivative, the adaptive di erential pulse code modulation (ADPCM), are well-known schemes for coding of images and video [1, 2]. We present a novel low-complexity scheme for adaptive di erential pulse code modulation of color video using a set of two-level quantizers which are switched adaptively on a block-by-block basis. The quantizers are designed by statistical optimization with a new design criterion that borrows from DPCM, block truncation coding (BTC) and vector quantization (VQ). Coding results show that our scheme has a better performance than other lowcomplexity schemes. At 2 bit/pixel, it performs more than 2.0 dB better in peak signal-to-noise ratio than DPCM and up to 3.5 dB better than BTC. Possible applications include VLSI implementations and real-time compression for video transmission over local area networks. 1. ADAPTIVE DIFFERENTIAL PULSE CODE MODULATION (ADPCM) The principle of DPCM (Di erential Pulse Code Modulation) of images is that pixels to be coded are predicted from previously coded neighbouring pixels, and only the quantized prediction error is transmitted for each pixel. Coder and decoder use the same reconstructed pixel amplitudes for prediction and are thus conform to each other. Usually predictor and quantizer are xed. Adaptive DPCM has the additional feature of several predictors and/or several quantizers that can be chosen adaptively. If the choice of the predictor and/or quantizer is based on previously transmitted pixels, no side information has to be signaled. If the choice of the predictor and/or quantizer is based on future pixels to be coded and transmitted, side information has to be transmitted in order to notify the decoder which predictor/quantizer to use. 2. ADPCM WITH SWITCHED TWO-LEVEL QUANTIZERS We employ an ADPCM scheme with adaptive one-bit (i.e., two-level) quantizers for the prediction error. The choice of a quantizer is based on pixels to be coded, thus side information has to be transmitted (Figure 1). Each pixel to be coded is predicted from already trans-