PUMA publications for /tag/quantizationhttps://puma.ub.uni-stuttgart.de/tag/quantizationPUMA RSS feed for /tag/quantization2024-03-29T12:14:06+01:00Charge Based Mixed-Signal Multiply-Accumulate Circuit for Energy Efficient In-Memory Computinghttps://puma.ub.uni-stuttgart.de/bibtex/20f60983a15955478092676950ebc085f/felixwiewelfelixwiewel2021-11-15T10:37:38+01:00deep efficient energy hardware implementation myown network neural quantization <span data-person-type="author" class="authorEditorList "><span><span itemtype="http://schema.org/Person" itemscope="itemscope" itemprop="author"><a title="Raphael Nägele" itemprop="url" href="/person/159b6e153a9af501a7f73b8203e663b0d/author/0"><span itemprop="name">R. Nägele</span></a></span>, </span><span><span itemtype="http://schema.org/Person" itemscope="itemscope" itemprop="author"><a title="Felix Wiewel" itemprop="url" href="/person/159b6e153a9af501a7f73b8203e663b0d/author/1"><span itemprop="name">F. Wiewel</span></a></span>, </span><span><span itemtype="http://schema.org/Person" itemscope="itemscope" itemprop="author"><a title="Sebastian Kelz" itemprop="url" href="/person/159b6e153a9af501a7f73b8203e663b0d/author/2"><span itemprop="name">S. Kelz</span></a></span>, </span><span><span itemtype="http://schema.org/Person" itemscope="itemscope" itemprop="author"><a title="Manuel Wittlinger" itemprop="url" href="/person/159b6e153a9af501a7f73b8203e663b0d/author/3"><span itemprop="name">M. Wittlinger</span></a></span>, </span><span><span itemtype="http://schema.org/Person" itemscope="itemscope" itemprop="author"><a title="Manfred Berroth" itemprop="url" href="/person/159b6e153a9af501a7f73b8203e663b0d/author/4"><span itemprop="name">M. Berroth</span></a></span>, </span><span><span itemtype="http://schema.org/Person" itemscope="itemscope" itemprop="author"><a title="Bin Yang" itemprop="url" href="/person/159b6e153a9af501a7f73b8203e663b0d/author/5"><span itemprop="name">B. Yang</span></a></span>, </span> and <span><span itemtype="http://schema.org/Person" itemscope="itemscope" itemprop="author"><a title="Markus Grözing" itemprop="url" href="/person/159b6e153a9af501a7f73b8203e663b0d/author/6"><span itemprop="name">M. Grözing</span></a></span></span>. </span><span class="additional-entrytype-information"><span itemtype="http://schema.org/Book" itemscope="itemscope" itemprop="isPartOf"><em><span itemprop="name">2021 Kleinheubach Conference</span>, </em></span><em>page <span itemprop="pagination">1-4</span>. </em>(<em><span>2021<meta content="2021" itemprop="datePublished"/></span></em>)</span>Mon Nov 15 10:37:38 CET 20212021 Kleinheubach Conference1-4Charge Based Mixed-Signal Multiply-Accumulate Circuit for Energy Efficient In-Memory Computing2021deep efficient energy hardware implementation myown network neural quantization EEM quantization revisited: asymptotic optimality for variable rate codinghttps://puma.ub.uni-stuttgart.de/bibtex/2062e3a09a2b460e0fcc4b8ad3c8a9875/thomasrichterthomasrichter2016-03-10T09:18:49+01:00myown quantization <span data-person-type="author" class="authorEditorList "><span><span itemtype="http://schema.org/Person" itemscope="itemscope" itemprop="author"><a title="Thomas Richter" itemprop="url" href="/person/15749c603f43f68dfbc74022088ffb6a1/author/0"><span itemprop="name">T. Richter</span></a></span></span>. </span><span class="additional-entrytype-information"><span itemtype="http://schema.org/Book" itemscope="itemscope" itemprop="isPartOf"><em><span itemprop="name">Applications of Digital Image Processing XXXIV</span>, </em></span><em> 8135, </em><em>SPIE, </em><em><span itemprop="publisher">SPIE</span>, </em>(<em><span>September 2011<meta content="September 2011" itemprop="datePublished"/></span></em>)</span>Thu Mar 10 09:18:49 CET 2016Applications of Digital Image Processing XXXIVsep{EEM} quantization revisited: asymptotic optimality for variable rate coding81352011myown quantization Equal-Expectation Magnitude Quantization (EEM) aims at minimizing the distortion of a quantizer with defined reconstruction points by shifting the deadzone parameter such that the expectation value of the signal equals the reconstructed value. While intuitively clear, this argument is not sufficient to prove rate-distortion optimality. In this work, it is show that the EEM quantizer is rate-distortion optimal up to third order in an expansion in powers of the quantization bucket size in the high-bitrate appoximation, and the approximating series for the optimal quantizer is computed. This result is compared to an even simpler quantization strategy based on the LLoyd-Max quantizer which selectively sets coefficients to zero. It is shown that both strategies lead to the same asymptotic expansion for the threshold parameter, but zeroing coefficients provides optimality in one additional order in the quantization bucket size.Deadzone Based Rate Allocation for JPEG XRhttps://puma.ub.uni-stuttgart.de/bibtex/2aff6f09c8d42df4861ce6c90540d2ba8/thomasrichterthomasrichter2016-03-10T09:18:49+01:00JPEG Optimal Quantization Quantization, Rate-Distortion XR!!, <span data-person-type="author" class="authorEditorList "><span><span itemtype="http://schema.org/Person" itemscope="itemscope" itemprop="author"><a title="Thomas Richter" itemprop="url" href="/person/195c1969b1a1b6bdc1d7a31ca7c6eecc1/author/0"><span itemprop="name">T. Richter</span></a></span></span>. </span><span class="additional-entrytype-information"><span itemtype="http://schema.org/Book" itemscope="itemscope" itemprop="isPartOf"><em><span itemprop="name">Data Compression Conference (DCC), 2011</span>, </em></span><em>page <span itemprop="pagination">474</span>. </em><em>Snowbird, UT, </em><em>IEEE, </em><em><span itemprop="publisher">IEEE</span>, </em>(<em><span>March 2011<meta content="March 2011" itemprop="datePublished"/></span></em>)</span>Thu Mar 10 09:18:49 CET 2016Snowbird, UTData Compression Conference (DCC), 2011mar474{D}eadzone {B}ased {R}ate {A}llocation for {JPEG} {XR}2011JPEG Optimal Quantization Quantization, Rate-Distortion XR!!, Similar to the JPEG image compression standard, the JPEG XR image
compression solely controls the image quality loss and hence the output rate
by means of the quantizer bucket sizes; a precise rate control mechanism
like the EBCOT rate allocation algorithm in JPEG 2000 is not specified, and
hence rate-distortion optimality of the quantizer is, in general, not given.
In this work, a simple rate-control mechanism for JPEG XR is introduced that
allows an efficient control of the quantizer towards rate-distortion
optimality. One possibility to implement this quantizer control would be to
use the spatial variable quantization feature of JPEG XR, but it was seen in
an earlier work that the additional side information required to transmit
the quantization setting almost compensates the PSNR gain of variable
quantization and complicates the rate allocation process by requiring an
additional quantizer allocation step.
However, while JPEG XR defines the image reconstruction process completely,
an encoder still has the freedom to select the deadzone size of the
quantizer; this mechanism has the additional advantage that no additional
side information needs to be transmitted and that the deadzone size is not,
unlike the quantizer bucket size, constrained to a set of pre-defined
values specified in the standard.
It is found that the image quality of JPEG XR can be improved by about 0.2
to 0.4 dB by performing a rate-distortion optimal selection of the deadzone;
this gain is seen to be comparable to the PSNR loss of a JPEG 2000 codec
where, for experimental reasons, EBCOT rate control has been turned off.Universal Refinable Trellis Coded Quantizationhttps://puma.ub.uni-stuttgart.de/bibtex/2129ef12bf4cd76de2771395baf389fb4/thomasrichterthomasrichter2016-03-10T09:18:49+01:00(signal);trellis coded codes;bitplane coding;Image coding;codebook coding;embedded coding;progressive coding;residual compression;Decoding;Distortion distortion measurement;Image performance;scalar quantisation quantization quantization;Bit quantization;universal quantizer;trellis rate;Data reconstruction;Quantization;Rate refinable theory;Rate-distortion;Transform training;rate trellis <span data-person-type="author" class="authorEditorList "><span><span itemtype="http://schema.org/Person" itemscope="itemscope" itemprop="author"><a title="S. Steger" itemprop="url" href="/person/135306931e0653b50d572d8534660dde9/author/0"><span itemprop="name">S. Steger</span></a></span>, </span> and <span><span itemtype="http://schema.org/Person" itemscope="itemscope" itemprop="author"><a title="T. Richter" itemprop="url" href="/person/135306931e0653b50d572d8534660dde9/author/1"><span itemprop="name">T. Richter</span></a></span></span>. </span><span class="additional-entrytype-information"><span itemtype="http://schema.org/Book" itemscope="itemscope" itemprop="isPartOf"><em><span itemprop="name">Data Compression Conference, 2009. DCC '09.</span>, </em></span><em>page <span itemprop="pagination">312-321</span>. </em>(<em><span>March 2009<meta content="March 2009" itemprop="datePublished"/></span></em>)</span>Thu Mar 10 09:18:49 CET 2016Data Compression Conference, 2009. DCC '09.mar312-321{U}niversal {R}efinable {T}rellis {C}oded {Q}uantization2009(signal);trellis coded codes;bitplane coding;Image coding;codebook coding;embedded coding;progressive coding;residual compression;Decoding;Distortion distortion measurement;Image performance;scalar quantisation quantization quantization;Bit quantization;universal quantizer;trellis rate;Data reconstruction;Quantization;Rate refinable theory;Rate-distortion;Transform training;rate trellis We introduce a novel universal refinable trellis quantization scheme (URTCQ) that is suitable for bitplane coding with many reconstruction stages. Existing refinable trellis quantizers either require excessive codebook training and are outperformed by scalar quantization for more than two stages (MS-TCQ, E-TCQ), require a huge computational burden (SR-TCQ) or achieve a good rate distortion performance in the last stage only (UTCQ). The presented quantization technique is a mixture of a scalar quantizer and an improved version of the E-TCQ. For all supported sources only one time training to an i.i.d. uniform source is required and its incremental bitrate is not more than 1 bps for each stage. The complexity is proportional to the number of stages and the number of trellis states. We compare the rate distortion performance of our work on generalized Gaussian i.i.d. sources with the quantizers deployed in JPEG2000 (USDZQ, UTCQ). It turns out that it is in no stage worse than the scalar quantizer and usually outperforms the UTCQ except for the last stage.Universal Refinable Trellis Coded Quantizationhttps://puma.ub.uni-stuttgart.de/bibtex/2129ef12bf4cd76de2771395baf389fb4/rainerreichelrainerreichel2016-03-03T17:45:04+01:00(signal);trellis coded codes;bitplane coding;Image coding;codebook coding;embedded coding;progressive coding;residual compression;Decoding;Distortion distortion measurement;Image performance;scalar quantisation quantization quantization;Bit quantization;universal quantizer;trellis rate;Data reconstruction;Quantization;Rate refinable theory;Rate-distortion;Transform training;rate trellis <span data-person-type="author" class="authorEditorList "><span><span itemtype="http://schema.org/Person" itemscope="itemscope" itemprop="author"><a title="S. Steger" itemprop="url" href="/person/135306931e0653b50d572d8534660dde9/author/0"><span itemprop="name">S. Steger</span></a></span>, </span> and <span><span itemtype="http://schema.org/Person" itemscope="itemscope" itemprop="author"><a title="T. Richter" itemprop="url" href="/person/135306931e0653b50d572d8534660dde9/author/1"><span itemprop="name">T. Richter</span></a></span></span>. </span><span class="additional-entrytype-information"><span itemtype="http://schema.org/Book" itemscope="itemscope" itemprop="isPartOf"><em><span itemprop="name">Data Compression Conference, 2009. DCC '09.</span>, </em></span><em>page <span itemprop="pagination">312-321</span>. </em>(<em><span>March 2009<meta content="March 2009" itemprop="datePublished"/></span></em>)</span>Thu Mar 03 17:45:04 CET 2016Data Compression Conference, 2009. DCC '09.mar312-321{U}niversal {R}efinable {T}rellis {C}oded {Q}uantization2009(signal);trellis coded codes;bitplane coding;Image coding;codebook coding;embedded coding;progressive coding;residual compression;Decoding;Distortion distortion measurement;Image performance;scalar quantisation quantization quantization;Bit quantization;universal quantizer;trellis rate;Data reconstruction;Quantization;Rate refinable theory;Rate-distortion;Transform training;rate trellis We introduce a novel universal refinable trellis quantization scheme (URTCQ) that is suitable for bitplane coding with many reconstruction stages. Existing refinable trellis quantizers either require excessive codebook training and are outperformed by scalar quantization for more than two stages (MS-TCQ, E-TCQ), require a huge computational burden (SR-TCQ) or achieve a good rate distortion performance in the last stage only (UTCQ). The presented quantization technique is a mixture of a scalar quantizer and an improved version of the E-TCQ. For all supported sources only one time training to an i.i.d. uniform source is required and its incremental bitrate is not more than 1 bps for each stage. The complexity is proportional to the number of stages and the number of trellis states. We compare the rate distortion performance of our work on generalized Gaussian i.i.d. sources with the quantizers deployed in JPEG2000 (USDZQ, UTCQ). It turns out that it is in no stage worse than the scalar quantizer and usually outperforms the UTCQ except for the last stage.Deadzone Based Rate Allocation for JPEG XRhttps://puma.ub.uni-stuttgart.de/bibtex/2aff6f09c8d42df4861ce6c90540d2ba8/rainerreichelrainerreichel2016-03-03T17:45:04+01:00JPEG Optimal Quantization Quantization, Rate-Distortion XR!!, <span data-person-type="author" class="authorEditorList "><span><span itemtype="http://schema.org/Person" itemscope="itemscope" itemprop="author"><a title="Thomas Richter" itemprop="url" href="/person/195c1969b1a1b6bdc1d7a31ca7c6eecc1/author/0"><span itemprop="name">T. Richter</span></a></span></span>. </span><span class="additional-entrytype-information"><span itemtype="http://schema.org/Book" itemscope="itemscope" itemprop="isPartOf"><em><span itemprop="name">Data Compression Conference (DCC), 2011</span>, </em></span><em>page <span itemprop="pagination">474</span>. </em><em>Snowbird, UT, </em><em>IEEE, </em><em><span itemprop="publisher">IEEE</span>, </em>(<em><span>March 2011<meta content="March 2011" itemprop="datePublished"/></span></em>)</span>Thu Mar 03 17:45:04 CET 2016Snowbird, UTData Compression Conference (DCC), 2011mar474{D}eadzone {B}ased {R}ate {A}llocation for {JPEG} {XR}2011JPEG Optimal Quantization Quantization, Rate-Distortion XR!!, Similar to the JPEG image compression standard, the JPEG XR image
compression solely controls the image quality loss and hence the output rate
by means of the quantizer bucket sizes; a precise rate control mechanism
like the EBCOT rate allocation algorithm in JPEG 2000 is not specified, and
hence rate-distortion optimality of the quantizer is, in general, not given.
In this work, a simple rate-control mechanism for JPEG XR is introduced that
allows an efficient control of the quantizer towards rate-distortion
optimality. One possibility to implement this quantizer control would be to
use the spatial variable quantization feature of JPEG XR, but it was seen in
an earlier work that the additional side information required to transmit
the quantization setting almost compensates the PSNR gain of variable
quantization and complicates the rate allocation process by requiring an
additional quantizer allocation step.
However, while JPEG XR defines the image reconstruction process completely,
an encoder still has the freedom to select the deadzone size of the
quantizer; this mechanism has the additional advantage that no additional
side information needs to be transmitted and that the deadzone size is not,
unlike the quantizer bucket size, constrained to a set of pre-defined
values specified in the standard.
It is found that the image quality of JPEG XR can be improved by about 0.2
to 0.4 dB by performing a rate-distortion optimal selection of the deadzone;
this gain is seen to be comparable to the PSNR loss of a JPEG 2000 codec
where, for experimental reasons, EBCOT rate control has been turned off.