CN117812263B - Multi-stage vector product quantization system and method for memory lightweight - Google Patents
Multi-stage vector product quantization system and method for memory lightweight Download PDFInfo
- Publication number
- CN117812263B CN117812263B CN202410232424.4A CN202410232424A CN117812263B CN 117812263 B CN117812263 B CN 117812263B CN 202410232424 A CN202410232424 A CN 202410232424A CN 117812263 B CN117812263 B CN 117812263B
- Authority
- CN
- China
- Prior art keywords
- order
- sub
- vectors
- vector
- mode
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 239000013598 vector Substances 0.000 title claims abstract description 143
- 238000013139 quantization Methods 0.000 title claims abstract description 36
- 238000000034 method Methods 0.000 title claims abstract description 12
- 238000004364 calculation method Methods 0.000 claims description 9
- 108700041286 delta Proteins 0.000 claims description 4
- 238000012804 iterative process Methods 0.000 abstract description 5
- 230000009467 reduction Effects 0.000 abstract description 4
- 230000000694 effects Effects 0.000 abstract description 3
- 230000008569 process Effects 0.000 description 3
- 238000004891 communication Methods 0.000 description 1
- 238000013144 data compression Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
Landscapes
- Compression, Expansion, Code Conversion, And Decoders (AREA)
Abstract
The invention provides a memory lightweight multi-stage vector product quantization system and a method, which are used for segmenting an original vector to form initial m sub-vectors, performing iterative switching by utilizing three modes, enabling an iterative process to enter a termination mode along with continuous reduction of an offset distance, wherein before the termination mode, the iterative process possibly goes through a single codebook mode or a normal iterative mode, and the normal iterative mode possibly is converted into the single codebook mode under the condition of meeting an intermediate state, and is not converted back to the normal iterative mode in iteration after entering the single codebook mode, so that multi-stage vector product quantization is realized. The invention utilizes a plurality of code book dividing modes to carry out vector quantization in the early stage, effectively reduces the occupation of the system memory in a mode of 'becoming zero' in the whole, and then converts the code book into a single code book at proper time along with the continuous reduction of the vector offset distance in the later stage, so that the code book and the memory occupied by the code book operation are further reduced, thereby effectively realizing the technical effect of 'memory lightweight' in the invention.
Description
Technical Field
The invention relates to a multi-stage vector product quantization system and method for memory lightweight.
Background
Vector quantization is a mechanism for mapping a sequence of continuous or discrete vectors into a sequence of numbers suitable for communication or storage over a digital channel, with the primary consideration being to map a high-dimensional vector onto its nearest prototype vector, thereby achieving efficient data compression.
Currently, the most critical challenge posed to vector quantization is that the amount of data processed tends to be too large, even massive. For example, taking an image as an example, the image may be divided into 65536 (=2 64) small areas according to the position of the image, and 65536 vector components may be generated when the image is described by vectors. And multiple frames of images form a video. In this case, if a single vector is used to quantize each image in the segment of video, a codebook needs to be set to traverse through quantization of all 65536 vector components. Thus, the complexity of the codebook design and the amount of data processing that is incurred by the quantization of the traversal will likely be daunting.
Disclosure of Invention
The invention provides a multi-stage vector product quantization system with a lightweight memory, which effectively solves the technical problems existing in the prior art.
Specifically, the invention provides a multi-stage vector product quantization system with lightweight memory, in the system, an initial description vector A0 of a target object, which contains n vector components, is returned to m dimensions according to the characteristics of the target object by using a dimension dividing module, so as to form m original sub-vectors, which are respectively A 1、A2…Am, wherein 2 is less than or equal to m < n, each sub-vector in the m sub-vectors is allocated with a codebook, each codebook is correspondingly counted as U 1、U2…Um, the m original sub-vectors A 1、A2…Am are quantized according to a corresponding codebook U 1、U2…Um, so as to form m quantized sub-vectors A' 1、A'2…A'm, A0 th offset distance L 0 between the m quantized sub-vectors and the m original sub-vectors is calculated, the system further comprises a comparison module, a first distance threshold L1 and a second distance threshold L2 are set in the comparison module, wherein the comparison judgment of the comparison module is divided into three cases: first case: if the comparison module judges that L 0 is less than or equal to L1, a termination mode is started, and the system terminates operation; second case: if the comparison module judges that L1< L 0 is less than or equal to L2, a set sub-codebook is canceled, a single codebook U is started, a single codebook iteration mode is started, based on m original sub-vectors A 1、A2…Am and m quantized sub-vectors A '1、A'2…A'm, m 1 st order residual sub-vectors delta 1 1、Δ12…Δ1m are calculated, m 1 st order quantized residual sub-vectors delta' 1 1、Δ'12…Δ'1m are obtained through the m 1 st order residual sub-vectors according to the single codebook U quantization, a1 st order offset distance L 1 is calculated according to the m 1 st order residual sub-vectors and the m 1 st order quantized residual sub-vectors, if L 1 is less than or equal to L1, a termination mode is started, the system is stopped, if L1< L 1 is less than or equal to L2, the single codebook iteration mode is continuously started until a k th order residual sub-vector and a k quantized residual sub-vector are calculated, the k order offset distance L k is less than or equal to L1, and finally the system is stopped under the third condition: if the comparison module judges that L 0 is larger than L2, a normal iteration mode is started, m 1 st order residual sub-vectors delta 1 1、Δ12…Δ1m respectively obtain m 1 st order quantized residual sub-vectors delta' 1 1、Δ'12…Δ'1m according to m code division books U 1、U2…Um, then 1 st order offset distance L 1 is calculated according to m 1 st order residual sub-vectors and m 1 st order quantized residual sub-vectors, if L 1 is smaller than or equal to L1, a termination mode is started, if L1< L 1 is smaller than or equal to L2, a single codebook iteration mode is started until a k order residual sub-vector and a k order quantized residual sub-vector are calculated, the k order offset distance L k is smaller than or equal to L1, a termination mode is finally started, if L 1 is larger than L2, a2 nd order offset distance L 2 is calculated, the system stops running until a k order residual sub-vector and a k order quantized residual sub-vector are calculated, the k order residual sub-vector is calculated, the system stops running, and the system stops running until the k order residual sub-vector L k is calculated, and the system stops running.
Optionally, the target object is a video image.
Preferably, the calculation formula of the 0 th order offset distance l 0 is:
Or is:
。
More preferably, the calculation formula of the 1 st order offset distance l 1 is:
Or is:
。
More preferably, the k-th order residual sub-vector is denoted as Δk 1、Δk2…Δkm, the k-th order quantized residual sub-vector is denoted as Δ' k 1、Δ'k2…Δ'km, and the calculation formula of the k-th order offset distance l k is:
Or is:
。
In summary, the present invention provides a multi-stage vector product quantization system, firstly, the original vector is split to form initial m sub-vectors, then, three modes are utilized to perform iterative switching, the iterative process finally enters a termination mode along with the continuous decrease of the offset distance, before entering the termination mode, the iterative process may go through a single codebook mode or a normal iteration mode, and the normal iteration mode may be converted into the single codebook mode under the condition of meeting the intermediate state, and after entering the single codebook mode, the normal iteration mode is not converted back to the normal iteration mode in iteration, thereby finally realizing multi-stage vector product quantization. In the whole flow of the invention, the vector quantization is carried out by utilizing a plurality of code-dividing books in the early stage, the occupation of the system memory is effectively reduced in a mode of 'becoming zero' in the early stage, and the single code-book is converted into the single code-book at proper time along with the continuous reduction of the offset distance of the vector in the later stage, so that the memory occupied by the code-book and the code-book operation is further reduced, thereby effectively realizing the technical effect of 'memory lightweight'.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following discussion will discuss the embodiments or the drawings required in the description of the prior art, and it is obvious that the technical solutions described in connection with the drawings are only some embodiments of the present invention, and that other embodiments and drawings thereof can be obtained according to the embodiments shown in the drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flow chart of a multi-stage vector product quantization system and method for memory lightweight according to the present invention.
Detailed Description
The following description of the embodiments of the present invention will be made in detail with reference to the accompanying drawings, wherein it is apparent that the embodiments described are only some, but not all embodiments of the present invention. All other embodiments, which can be made by a person of ordinary skill in the art without the need for inventive faculty, are within the scope of the invention, based on the embodiments described in the present invention.
The invention provides a multi-stage vector product quantization system with lightweight memory, aiming at carrying out multi-stage iteration on a description vector aiming at a target object on the basis of dimension quantization and changing codebook configuration according to iteration conditions. FIG. 1 is a flow chart of a multi-stage vector product quantization system and method for memory lightweight according to the present invention.
The target object, most commonly a video image, for example. As mentioned in the background, an image may be described as an initial description vector A0 including n vector components according to different regions into which the image is segmented. Of course, the object of the present invention is not limited to video images, but may be a voice signal or other types of signals.
And then, a dimension dividing module in the multi-stage vector product quantization system divides the initial description vector into m dimensions according to the characteristics of the target object to form m original sub-vectors, wherein the m original sub-vectors are respectively A 1、A2…Am, and m is more than or equal to 2 and less than n. Each sub-vector of the m sub-vectors is allocated a sub-codebook, and each sub-codebook is correspondingly counted as U 1、U2…Um.
Each of the m original sub-vectors a 1、A2…Am is quantized according to its corresponding codebook U 1、U2…Um, thereby forming m quantized sub-vectors a' 1、A'2…A'm.
The 0 th order offset distance l 0 is calculated on the basis of this. The calculation of the offset distance may construct different formulas, such as:
Or is:
Further, a comparison module is provided in the multi-stage vector product quantization system of the present invention, i.e., a first distance threshold L1 and a second distance threshold L2 are provided in the comparison module, where L1< L2.
Next, comparing the 0 th offset distance l 0 with the two thresholds, respectively, three cases may occur:
first case: if the 0 th offset distance L 0 meets L 0 and is less than or equal to L1, a termination mode is started, and the multi-stage vector product quantization system is stopped, wherein the condition shows that the initial description vector is divided into m dimensions and the codebook is allocated respectively, so that the vector quantization requirement can be met, and the system does not need to continue to run.
Second case: if the 0 th order offset distance L 0 satisfies L1< L 0 +.L2, then an intermediate state is provided that indicates that the multi-stage vector product quantization system still needs to operate, but since the offset distance is already within the threshold L2, there is no need to set m sub-codebooks U 1、U2…Um, and instead, the single codebook U is enabled alone and the single codebook iteration mode is enabled, and the specific operation of the model will be described below.
Third case: if the 0 th order offset distance L 0 satisfies L 0 > L2, then this indicates that the "opportunity is not reached" and that the residual iteration needs to be continued, thereby turning on the normal iteration mode, which is further operational, as will be explained below.
When the second or third situation occurs, the system needs to continue to operate.
Based on the m original sub-vectors a 1、A2…Am and the m quantized sub-vectors a' 1、A'2…A'm, m 1 st order residual sub-vectors Δ1 1、Δ12…Δ1m are calculated. There are various calculation methods for the residual, such as direct vector subtraction, that is, each order residual after ,Δ11= A'1- A1;Δ12= A'2- A2…Δ1m= A'm- Am. can be obtained by referring to the above method.
In the second case described above, since the single codebook U is enabled, the single codebook iteration mode is turned on. In this mode, the m1 st order residual sub-vectors Δ1 1、Δ12…Δ1m obtain m1 st order quantized residual sub-vectors Δ1' 1、Δ'12…Δ'1m according to the unique codebook U, and then calculate the 1 st order offset distance l 1 according to the m1 st order residual sub-vectors and the m1 st order quantized residual sub-vectors.
The calculation process of the 1 st offset distance l 1 is similar to that of the 0 th offset distance l 0, for example, as follows:
Or is:
The 0 th order offset distance l 0 is the offset distance under the original vector, while the 1 st order offset distance l 1 is the offset distance under the residual vector after the iteration, and therefore, l 1<l0 is necessary. Thus, in the single codebook mode, since L1< L 0 is L2, after the comparison module is started, the 1 st offset distance L 1 can only exist in two cases, namely, L1< L 1 is L2 or L 1 is L1.
If L 1 is less than or equal to L1, the termination mode is started, and the multi-stage vector product quantization system is not operated any more.
If L1< L 1 is less than or equal to L2, calculating m 2 nd order residual sub-vectors Δ2 1、Δ22…Δ2m based on m 1 st order residual sub-vectors Δ1 1、Δ12…Δ1m and m 1 st order quantized residual sub-vectors Δ1 '1、Δ'12…Δ'1m, further obtaining m 2 nd order quantized residual sub-vectors Δ2' 1、Δ'22…Δ'2m according to a single codebook U, and then calculating a 2 nd order offset distance L 2 according to the m 2 nd order residual sub-vectors and the m 2 nd order quantized residual sub-vectors. Since there must be L 2<l1, the 2 nd offset distance L 2 can only exist in two cases, i.e., L1< L 2.ltoreq.L2 or L 2.ltoreq.L1. If L 2 is less than or equal to L1, the termination mode is started, and the multi-stage vector product quantization system is not operated any more. If L1< L 2.ltoreq.L2, then the single codebook mode is started again.
And the rest is analogiced until a k-order residual sub-vector and a k-order quantized residual sub-vector are calculated, so that the k-order offset distance L k is less than or equal to L1, a termination mode is finally started, and the multi-stage vector product quantization system does not run any more.
As can be seen above, once the system enters the single codebook mode, as the iteration progresses, it is impossible to re-enter the normal iteration mode, which can only continue to iterate back to the single codebook mode or the termination mode, and eventually will iterate to the termination mode.
In the third case, that is, the 0 th order offset distance L 0 satisfies L 0 > L2, the normal iteration mode is started, the m1 st order residual sub-vectors Δ1 1、Δ12…Δ1m each obtain m1 st order quantized residual sub-vectors Δ1' 1、Δ'12…Δ'1m according to the m code division codebooks U 1、U2…Um, and then the 1 st order offset distance L 1 is calculated according to the m1 st order residual sub-vectors and the m1 st order quantized residual sub-vectors.
In this case, the calculation process of the 1 st order offset distance l 1 is as follows, for example:
Or is:
There must be L 1<l0, and L 0 > L2, so there are still three possibilities for the offset distance of order 1, L 1: l 1≤L1、L1<l1≤L2、l1 > L2.
If L 1 is less than or equal to L1, starting a termination mode, and terminating the operation of the multi-stage vector product quantization system.
If L1< L 1 is less than or equal to L2, starting a single codebook iteration mode until a k-order residual sub-vector and a k-order quantized residual sub-vector are calculated, so that the obtained k-order offset distance L k is less than or equal to L1, and finally starting a termination mode, wherein the multi-stage vector product quantization system is not operated any more.
If L 1 is greater than L2, starting a normal iteration mode, calculating a2 nd order offset distance L 2, and the like until a k order residual sub-vector and a k order quantized residual sub-vector are calculated, wherein the k order offset distance L k is less than or equal to L1, and finally starting a termination mode, wherein the multi-stage vector product quantization system is not operated any more. It should be noted that in this case, in the iteration process from the 2 nd offset distance L 2 to the k-th offset distance L k, it is possible to transition from the normal iteration mode to the single codebook mode and then enter the termination mode, thereby determining the k-th offset distance L k, or it is also possible to directly determine that L k is less than or equal to L1 directly after the normal iteration mode is iterated without going through the single codebook mode, thereby directly entering the termination mode.
As can be seen from the foregoing, the multi-stage vector product quantization system according to the present invention firstly segments the original vector to form the initial m sub-vectors, then performs iterative switching by using three modes, and finally enters the termination mode with the continuous decrease of the offset distance, and before entering the termination mode, the iterative process may undergo a single codebook mode or a normal iteration mode, and the normal iteration mode may be converted into the single codebook mode under the condition of meeting the intermediate state, and after entering the single codebook mode, the normal iteration mode is not converted back into the normal iteration mode in the iteration, thereby finally realizing multi-stage vector product quantization. In the whole flow of the invention, the vector quantization is carried out by utilizing a plurality of code-dividing books in the early stage, the occupation of the system memory is effectively reduced in a mode of 'becoming zero' in the early stage, and the single code-book is converted into the single code-book at proper time along with the continuous reduction of the offset distance of the vector in the later stage, so that the memory occupied by the code-book and the code-book operation is further reduced, thereby effectively realizing the technical effect of 'memory lightweight'.
The foregoing description of the exemplary embodiments of the invention is not intended to limit the invention to the precise form disclosed, and any modifications, equivalents, and variations which fall within the spirit and scope of the invention are intended to be included in the scope of the invention.
Claims (6)
1. A multi-stage vector product quantization system with lightweight memory is characterized in that the system comprises a dimension dividing module and a comparison module, wherein
The dimension dividing module divides an initial description vector A0 of the target object containing n vector components into m dimensions according to the characteristics of the target object, thereby forming m original sub-vectors which are respectively A 1、A2…Am, wherein m < n is more than or equal to 2, each sub-vector in the m sub-vectors is allocated with a sub-codebook, and each sub-codebook is correspondingly counted as U 1、U2…Um,
Each of the m original sub-vectors A 1、A2…Am is quantized according to its corresponding codebook U 1、U2…Um, thereby forming m quantized sub-vectors A' 1、A'2…A'm, calculating the 0 th order offset distance l 0 between the m quantized sub-vectors and the m original sub-vectors,
A first distance threshold L1 and a second distance threshold L2 are set in the comparison module, wherein L1< L2:
First case: if the comparison module judges that L 0 is less than or equal to L1, starting a termination mode, and terminating the operation of the system;
second case: if the comparison module determines that L1< L 0 is less than or equal to L2, the setting of the sub-codebook is canceled, a single codebook U is started, a single codebook iteration mode is started, m 1 st order residual sub-vectors delta 1 1、Δ12…Δ1m are calculated based on m original sub-vectors A 1、A2…Am and m quantized sub-vectors A '1、A'2…A'm, m 1 st order residual sub-vectors are quantized according to the single codebook U to obtain m 1 st order quantized residual sub-vectors delta' 1 1、Δ'12…Δ'1m, then 1 st order offset distance L 1 is calculated according to m 1 st order residual sub-vectors and m 1 st order quantized residual sub-vectors,
If L 1 is less than or equal to L1, then a termination mode is initiated, the system terminates operation,
If L1< L 1 is less than or equal to L2, continuing to start a single codebook iteration mode until a k-order residual sub-vector and a k-order quantized residual sub-vector are calculated, the k-order offset distance L k is less than or equal to L1, finally starting a termination mode, terminating the system,
Third case: if the comparison module determines that L 0 > L2, then the normal iteration mode is started, the m 1 st order residual sub-vectors delta 1 1、Δ12…Δ1m respectively obtain m 1 st order quantized residual sub-vectors delta' 1 1、Δ'12…Δ'1m according to m code division books U 1、U2…Um, then the 1 st order offset distance L 1 is calculated according to the m 1 st order residual sub-vectors and the m 1 st order quantized residual sub-vectors,
If L 1 is less than or equal to L1, then a termination mode is initiated, the system terminates operation,
If L1< L 1 is less than or equal to L2, starting a single codebook iteration mode until a k-order residual sub-vector and a k-order quantized residual sub-vector are calculated, the k-order offset distance L k is less than or equal to L1, finally starting a termination mode, terminating the system,
If L 1 is greater than L2, starting a normal iteration mode, calculating a2 nd order offset distance L 2, and the like until a k order residual sub-vector and a k order quantized residual sub-vector are calculated, wherein the k order offset distance L k is less than or equal to L1, and finally starting a termination mode, wherein the system terminates operation.
2. The system of claim 1, wherein the target object is a video image.
3. The system of claim 1, wherein the 0 th order offset distance l 0 is calculated as:
Or is: 。
4. A system according to claim 3, wherein the calculation formula of the 1 st order offset distance l 1 is:
Or is: 。
5. The system of claim 4, wherein the k-th order residual sub-vector is denoted as Δk 1、Δk2…Δkm, the k-th order quantized residual sub-vector is denoted as Δ' k 1、Δ'k2…Δ'km, and the calculation formula of the k-th order offset distance l k is:
Or is: 。
6. A multi-stage vector product quantization method of memory lightweight, the method being performed by the system of any of claims 1-5.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410232424.4A CN117812263B (en) | 2024-03-01 | 2024-03-01 | Multi-stage vector product quantization system and method for memory lightweight |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410232424.4A CN117812263B (en) | 2024-03-01 | 2024-03-01 | Multi-stage vector product quantization system and method for memory lightweight |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117812263A CN117812263A (en) | 2024-04-02 |
| CN117812263B true CN117812263B (en) | 2024-05-07 |
Family
ID=90433817
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410232424.4A Active CN117812263B (en) | 2024-03-01 | 2024-03-01 | Multi-stage vector product quantization system and method for memory lightweight |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117812263B (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111368133A (en) * | 2020-04-16 | 2020-07-03 | 腾讯科技（深圳）有限公司 | Method and device for establishing index table of video library, server and storage medium |
| CN114600459A (en) * | 2019-11-26 | 2022-06-07 | 谷歌有限责任公司 | Vector quantization for prediction residual coding |
| CN114638905A (en) * | 2022-01-30 | 2022-06-17 | 中国科学院自动化研究所 | Image generation method, device, equipment, storage medium and computer program product |
| CN114945982A (en) * | 2020-01-13 | 2022-08-26 | 诺基亚技术有限公司 | Spatial audio parametric coding and associated decoding |
| CN114979659A (en) * | 2021-02-19 | 2022-08-30 | 三星显示有限公司 | Method for encoding, encoder, and image compression and storage system |
| CN116939226A (en) * | 2023-06-14 | 2023-10-24 | 南京大学 | Low-code-rate image compression-oriented generated residual error repairing method and device |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP4801160B2 (en) * | 2005-09-23 | 2011-10-26 | テレフオンアクチーボラゲット エル エム エリクソン（パブル） | Sequential improvement of lattice vector quantization |
| EP2048787A4 (en) * | 2006-12-05 | 2009-07-01 | Huawei Tech Co Ltd | Method and device for quantizing vector |
-
2024
- 2024-03-01 CN CN202410232424.4A patent/CN117812263B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114600459A (en) * | 2019-11-26 | 2022-06-07 | 谷歌有限责任公司 | Vector quantization for prediction residual coding |
| CN114945982A (en) * | 2020-01-13 | 2022-08-26 | 诺基亚技术有限公司 | Spatial audio parametric coding and associated decoding |
| CN111368133A (en) * | 2020-04-16 | 2020-07-03 | 腾讯科技（深圳）有限公司 | Method and device for establishing index table of video library, server and storage medium |
| CN114979659A (en) * | 2021-02-19 | 2022-08-30 | 三星显示有限公司 | Method for encoding, encoder, and image compression and storage system |
| CN114638905A (en) * | 2022-01-30 | 2022-06-17 | 中国科学院自动化研究所 | Image generation method, device, equipment, storage medium and computer program product |
| CN116939226A (en) * | 2023-06-14 | 2023-10-24 | 南京大学 | Low-code-rate image compression-oriented generated residual error repairing method and device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117812263A (en) | 2024-04-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US6836225B2 (en) | Fast search method for nearest neighbor vector quantization | |
| Shiraki et al. | LPC speech coding based on variable-length segment quantization | |
| US5398069A (en) | Adaptive multi-stage vector quantization | |
| CN116250236A (en) | Example adaptive image and video compression using machine learning system | |
| EP0709826B1 (en) | Pattern recognition method | |
| US7961787B2 (en) | Method of video content complexity estimation, scene change detection and video encoding | |
| JP2911682B2 (en) | Motion compensation using the minimum number of bits per motion block as a criterion for block matching | |
| CN117812263B (en) | Multi-stage vector product quantization system and method for memory lightweight | |
| US9451265B2 (en) | Method, apparatus and computer program product for providing improved data compression | |
| US20040176951A1 (en) | LSF coefficient vector quantizer for wideband speech coding | |
| CN111401546A (en) | Training method of neural network model, medium thereof, and electronic device | |
| EP0680032B1 (en) | Quantization of input vectors with and without rearrangement of vector elements of a candidate vector | |
| CN111344719A (en) | Data processing method and device based on deep neural network and mobile device | |
| JP2968666B2 (en) | Image coding method and apparatus | |
| CN110570877B (en) | Sign language video generation method, electronic device and computer readable storage medium | |
| JPH10124686A (en) | Method and device for encoding contour | |
| JPH0361956B2 (en) | ||
| JP2001519631A (en) | How to control the compression ratio of digital images | |
| JP4603429B2 (en) | Client / server speech recognition method, speech recognition method in server computer, speech feature extraction / transmission method, system, apparatus, program, and recording medium using these methods | |
| Shiraki et al. | Speaker Adaptation Algorithms Based on Piecewise‐Moving Adaptive Segment Quantization Method | |
| CN110780838B (en) | Self-adaptive DeMURA data compression method for OLED screen | |
| EP1152620A2 (en) | Image coding embedded in matrix operation | |
| Bage | Interframe predictive coding of images using hybrid vector quantization | |
| CN110942750A (en) | Fourier transform-based OLED screen DeMURA method | |
| AU2021106652A4 (en) | Statistical features based time-efficient high-performance fractal image encoder |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |