Pub. online:1 Jan 2017Type:Research ArticleOpen Access
Volume 28, Issue 4 (2017), pp. 703–724
Scalar quantizer selection for processing a signal with a unit variance is a difficult problem, while both selection and quantizer design for the range of variances is even tougher and to the authors’ best knowledge, it is not theoretically solved. Furthermore, performance estimation of various image processing algorithms is unjustifiably neglected and there are only a few analytical models that follow experimental analysis. In this paper, we analyse application of piecewise uniform quantizer with Golomb-Rice coding in modified block truncation coding algorithm for grayscale image compression, propose design improvements and provide a novel analytical model for performance analysis. Besides the nature of input signal, required compression rate and processing delay of the observed system have a strong influence on quantizer design. Consequently, the impact of quantizer range choice is analysed using a discrete designing variance and it was exploited to improve overall quantizer performance, whereas variable-length coding is applied in order to reduce quantizer’s fixed bit-rate. The analytical model for performance analysis is proposed by introducing Inverse Gaussian distribution and it is obtained by discussing a number of images, providing general closed-form solutions for peak-signal-to-noise ratio and the total average bit-rate estimation. The proposed quantizer design ensures better performance in comparison to the other similar methods for grayscale image compression, including linear prediction of pixel intensity and edge-based adaptation, whereas analytical model for performance analysis provides matching with the experimental results within the range of 1 dB for PSQNR and 0.2 bpp for the total average bit-rate.
Pub. online:1 Jan 2012Type:Research ArticleOpen Access
Volume 23, Issue 1 (2012), pp. 125–140
In this paper, a piecewise uniform quantizer for input samples with discrete amplitudes for Laplacian source is designed and analyzed, and its forward adaptation is done. This type of quantizers is very often used in practice for the purpose of compression and coding of already quantized signals. It is shown that the design and the adaptation of quantizers for discrete input samples are different from the design and the adaptation of quantizers for continual input samples. A weighting function for PSQNR (peak signal-to-quantization noise ratio), which is obtained based on probability density function of variance of standard test images is introduced. Experiments are done, applying these quantizers for compression of grayscale images. Experimental results are very well matched to the theoretical results, proving the theory. Adaptive piecewise uniform quantizer designed for discrete input samples gives for 9 to 20 dB higher PSQNR compared to the fixed piecewise uniform quantizer designed for discrete input samples. Also it is shown that the adaptive piecewise uniform quantizer designed for discrete input samples gives higher PSQNR for 1.46 to 3.45 dB compared the adaptive piecewise uniform quantizer designed for continual input samples, which proves that the discrete model is more appropriate for image quantization than continual model.
Pub. online:1 Jan 2010Type:Research ArticleOpen Access
Volume 21, Issue 3 (2010), pp. 375–391
In this paper new semilogarithmic quantizer for Laplacian distribution is presented. It is simpler than classic A-law semilogarithmic quantizer since it has unit gain around zero. Also, it gives for 2.97 dB higher signal-to-quantization noise-ratio (SQNR) for referent variance in relation to A-law, and therefore it is more suitable for adaptation. Forward adaptation of this quantizer is done on frame-by-frame basis. In this way G.712 standard is satisfied with 7 bits/sample, which is not possible with classic A-law. Inside each frame subframes are formed and lossless encoder is applied on subframes. In that way, double adaptation is done: adaptation on variance within frames and adaptation on amplitude within subframes. Joined design of quantizer and lossless encoder is done, which gives better performances. As a result, standard G.712 is satisfied with only 6.43 bits/sample. Experimental results, obtained by applying this model on speech signal, are presented. It is shown that experimental and theoretical results are matched very well (difference is less than 1.5%). Models presented in this paper can be applied for speech signal and any other signal with Laplacian distribution.