WO1998018262A1 - Vector quantisation codebook generation method - Google Patents
Vector quantisation codebook generation methodInfo
- Publication number
- WO1998018262A1 WO1998018262A1 PCT/GB1997/002841 GB9702841W WO9818262A1 WO 1998018262 A1 WO1998018262 A1 WO 1998018262A1 GB 9702841 W GB9702841 W GB 9702841W WO 9818262 A1 WO9818262 A1 WO 9818262A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- vectors
- space
- vector
- index
- data array
- Prior art date
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/90—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using coding techniques not provided for in groups H04N19/10-H04N19/85, e.g. fractals
- H04N19/94—Vector quantisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T9/00—Image coding
- G06T9/008—Vector quantisation
-
- H—ELECTRICITY
- H03—ELECTRONIC CIRCUITRY
- H03M—CODING; DECODING; CODE CONVERSION IN GENERAL
- H03M7/00—Conversion of a code where information is represented by a given sequence or number of digits to a code where the same, similar or subset of information is represented by a different sequence or number of digits
- H03M7/30—Compression; Expansion; Suppression of unnecessary data, e.g. redundancy reduction
- H03M7/3082—Vector coding
Definitions
- the present invention relates to vector quantisation data compression and in particular to a method of constructing a codebook for use in vector quantisation data compression and to codebooks so constructed.
- 'vector quantisation' where a codebook of reference patches (e.g. relatively small portions taken from one or more 'library' or 'archetypal' images) is utilised.
- each image to be compressed is partitioned into a number of image patches and a matching (i.e.
- the codebook index for each chosen reference patch is stored, together with the corresponding position vectors of the image patches (i.e. their position in the original image), to provide a compressed representation of the image.
- an approximation to the original image can be constructed by using the stored codebook indices to recover the required set of reference patches and inserting these into an image frame using the respective stored image patch position vectors.
- the achievable degree of compression is a function of the size of the image patches into which the image is partitioned, larger patches allowing higher compression.
- Simple vector quantisation techniques require an exhaustive search of the codebook for each image patch, a process which is extremely computationally expensive where large codebooks are used to maximise the quality of the compressed images and where large patches, necessitating a large number of pixel comparisons, are generated to maximise the compression ratio.
- One technique used to reduce this search overhead is known as 'hierarchical vector quantisation' (HVQ) . This technique is described in 'Hierarchical Vector Quantisation of Perceptually Weighted Block Transforms' ; N Chadda, M Vishwanath, and P A Chou; Proceedings of the Data Compression Conference; IEEE, 1995, pp3-12.
- An index table T 0 is provided which maps every possible value of the 2-D vectors to associated codebook addresses, such that T 0 [i,j] addresses the closest entry in B 0 (i.e. B 0 [T 0 [i, ] ] ) to the vector
- a second codebook B x contains a number of 4-D vectors [p, q, r, s] which correspond to a set of four pixel patches (where again the number of patches is considerably less than the total number of possible four bit patches) .
- a second level index table T 1 is constructed such that
- B ⁇ [T x [T 0 [i, ] ,T 0 [k, l] ] ] is the closest entry in B to [B 0 [T 0 [i, ] ] , B 0 [T 0 [k, l] ] ] ] .
- Finding the closest entry in the codebook B ⁇ to a given input vector can then be determined by looking up the table T 0 for pixel pairs [i, j] and [k, 1] , and then applying the resultant 2-D vectors T 0 [i, j] , T 0 [k, 1] to the table T x . This process is illustrated in Figure 1.
- n X m is much greater than n-1 for reasonable values of m and that the HVQ look-up process will be approximately m times faster than the exhaustive search.
- Hierarchical vector quantisation may be used to compress colour images by separating out the three components of colour images (e.g. red, blue, green) and compressing them separately, and recombining the three sets of codebook reference patches to produce a decompressed image .
- colour images e.g. red, blue, green
- codebooks B 0 1 m where m is log 2 of the number of pixels in the largest patch used for compression.
- the conventional HVQ approach is to develop codebooks first, e.g. by extracting reference patches from a number of archetypal image frames, and then to derive the tables from the codebooks .
- the index table T m is constructed by taking all possible pairs of patches at level m-1 and by conducting an exhaustive search of patches in the codebook at level m to identify the patch at level m which is most similar. Thus, if the level m-1 patches 7 and 13 when placed together most closely resemble level m patch 100, T m [7,13] would be set to 100. This process is propagated back thorough the levels to create a codebook and an index table at each level .
- a method of constructing a vector quantisation codebook from at least one archetypal data array composed of a set of data values comprising:
- each sample vector consisting of a set of data values which are contiguous in the array (s) and each sample vector defining a point in a finite n-dimensional space;
- step 2) partitioning said space into a predetermined number of regions, each region containing substantially the same number of sample vectors; 3) assigning to each said region a unique index, where the indices are selected to codify the partitioning process carried out in step 2) ;
- n m dimension reference vector which is representative of the associated set of n dimension sample vectors in the archetypal data array
- step a) can be made easier if it is performed on spaces of low dimensions
- step 1) the selected vectors are distributed substantially evenly about said data array (s) . More preferably, every data value is contained within at least one vector.
- step 5) comprises sub-dividing the data array (s) into said second multiplicity of sample vectors, so that every entry is contained within at least one vector.
- the selected vectors may encompass only a fraction of the data array, where that fraction coincides with those areas from which the first multiplicity of sample vectors are subsequently chosen.
- step 2) comprises: (a) determining the mean and variance of the distribution of vectors for each dimension of said space; (b) identifying that dimension for which the vectors have the greatest variance; and (c) dividing the space into two regions, a first region containing those vectors which have a value for the selected dimension which exceeds the mean value and a second region containing those vectors which do not. For each region, steps (a) to (c) are recursively repeated until the required number of regions is obtained. This number is determined by the desired size of the lookup index tables.
- step 2) comprises determining the Principal Component (or regression line) of the vector distribution and then determining the mean of the Principal Component .
- the space is then split into two by a line perpendicular to the Principal Component and passing through the mean thereof to create two regions or sub-spaces. Each of these regions is in turn split, with the process continuing recursively until a desired number of regions is achieved.
- the number of regions created by the partitioning operation is typically less than 1000 to 4000, e.g. 128 or 256. For higher numbers of regions, the computational complexity may be unmanageable.
- the sequence in which the regions are created is used to determine the assignment of indices.
- the first split determines the most significant bit of the index
- the second split determines the next most significant bit etc.
- the appropriate bit is set to 1 for the region containing vectors exceeding the mean and to 0 for the region containing vectors below the mean. It will be appreciated that a similar result can be achieved by replacing l's with 0's and 0's with l's.
- the result of assigning the indices is that the indices preserve the order of the regions having regard to the average value of data values making up the vectors of each region.
- the present invention is particularly applicable to creating codebooks for use in compressing digital images which comprise an array of image pixels, each defined by an intensity value.
- the vectors at the first level represent patches of contiguous pixels in an archetypal image frame whilst the vectors at the higher levels correspond to patches in a shrunk image.
- the regions are assigned indices which preserve the subjective order of the regions. For example, regions containing relatively dark patches are assigned relatively low indices whilst regions containing relatively bright patches are assigned relatively high indices.
- n 2. More preferably, in the first iteration, every pair of horizontally adjacent data values in the archetypal data array (s) is selected to provide the vectors for use in step 2) so that substantially all data values appear in two data value pairs. In step 5) , the or each data array is sub- divided into horizontally adjacent pairs of data values so that each pair abuts the neighbouring pair(s). In the second iteration, in step 1) every pair of vertically adjacent data values are extracted whilst in step 5) [if the process comprises three or more iterations] the or each further data array is sub-divided into vertically adjacent pairs of data values. For each subsequent iteration, horizontally and vertically adjacent pairs of data values are chosen alternately.
- the sample vectors of the first and second multiplicities preferably all have the same orientation, this is not essential. For example, there may be a mixture of horizontal and vertical patches in either of the multiplicities .
- the above method is capable of producing optimised codebooks for use in any vector quantisation process. However, it is particularly suited to producing codebooks for use in hierarchical vector quantisation (HVQ) processes given that index look-up tables for each level, and which are necessary for HVQ, are an inherent product of the codebook creation method.
- HVQ hierarchical vector quantisation
- steps 7) and 8) it may be desired to additionally carry out steps 7) and 8) at the end of each iteration so as to create a codebook at each level. This may be necessary, for example, where an image is to be compressed using variable sized reference vectors.
- a method of constructing a vector quantisation codebook from at least one archetypal data array composed of a set of data values comprising:
- each sample vector consisting of a set of data values which are contiguous in the array (s) and each sample vector defining a point within a finite n-dimensional space;
- a method of constructing a set of index tables for use in a hierarchical vector quantisation process the tables being constructed from at least one archetypal data array composed of a set of data values, the method comprising:
- each sample vector consisting of a set of data values which are contiguous in the array (s) and each sample vector defining a point in a finite n-dimensional space;
- each index generated in the final iteration is derived from a set of n dimension sample vectors in the archetypal data array (s), where m is the number of iterations carried out.
- the method of the above third aspect of the present invention is able to generate a set of index tables for use in applications where it is desired to encode data but where it is not necessary to be able to reconstruct the original data from the encoded data, i.e. where a codebook is unnecessary.
- Figure 1 illustrates schematically a two layer hierarchical vector quantisation process.
- Figure 2 illustrates a 2-D digital image frame for compression
- Figure 3 shows a first level frequency map obtained from the image frame of Figure 2 ;
- Figure 4 shows a first level shrunk image frame obtained from the image frame of Figure 2 using the frequency map of Figure 3 ;
- Figure 5 shows a second level frequency map obtained from the image of Figure 4;
- Figure 6 shows a second level shrunk image frame obtained from the image frame of Figure 2 using the frequency map of Figure 5.
- Figure 2 a two dimensional black and white digital image which is composed of 320 x 200 pixels.
- the brightness level or intensity of each pixel is defined within the range 0 to 256 (i.e. by an 8-bit binary integer) .
- This process is illustrated in Figure 1 and has been described above.
- the image is firstly sub-divided into regular 2 x 2 pixel image patches. Each of these patches is then divided into upper and lower pixel pairs. For each patch in turn, the upper and lower pairs are fed in parallel to identical first level look-up tables T 0 to derive the first level indices for each pair.
- the resulting pair of indices are then fed to a second level look-up table T 1 to derive a second level index which addresses a reference patch in a codebook.
- a compressed representation of the image is created by storing only the positions in the image of each of the image patches and the respective indices. In order to decompress the data, it is only necessary to access the codebook, or a copy of the codebook, using the indices and to correctly order the resulting reference patches.
- codebooks are generally created using a set of test or archetypal images which are expected to be typical of images to be compressed. Therefore, if the patches in the codebook occur with substantially equal frequency in the test images, it is likely that these patches will also occur with equal frequency in images to be compressed.
- the first step in constructing a codebook of 2 x 2 pixel patches suitable for compressing the image frame of Figure 2 is to extract every pair of horizontally adjacent pixels from a digital archetypal image (in this case the image frame of Figure 2 may be considered an archetypal image) such that every pixel, other than pixels at the edges of the frame, appear in two different pairs.
- the archetypal image is selected to have similar properties to images to be compressed.
- the intensities of the right and left pixels of each pair are then plotted against each other in a two dimensional space to obtain a first level frequency map (or occurrence map) which shows the intensity distribution of the pixel pairs.
- FIG. 3 The map for the image of Figure 2 is shown in Figure 3 where, for the purpose of illustration, bright areas correspond to a high density of similar pixel pairs whilst dark areas correspond to a low density of similar pixel pairs.
- adjacent pairs of pixels will tend to have similar intensities (except where a pixel pair straddles a sharp edge) resulting in a high intensity band running across a diagonal of the frequency map from the bottom left of the map to the top right.
- the intensity distribution is then determined for both axes of the map.
- the distribution for the x-axis is illustrated alongside the x-axis. From the distributions, it is possible to determine the mean, and hence the variance, of the distribution with respect to each axis.
- the frequency map is then partitioned into two regions by a partitioning line (in this case it is the line denoted A) parallel to the axis having the least variance and passing through the mean intensity of the other axis to create two new sub-maps. It will be apparent that both of these sub- maps will contain approximately the same number of pixel pairs.
- the first split in the process determines the most significant bit of the index, the second split the next most significant bit etc. Also, where a horizontal split occurs, e.g., by partitioning line A, the relevant bit for the upper sub-map is set to 1 and that for the lower sub-map is set to 0. Where a vertical split occurs, e.g., by partitioning line D, the relevant bit for the right sub-map is set to 1 whilst that for the left sub-map is set to 0.
- the result is a set of indices which tend to preserve the subjective intensity of the maps. For example, the sub-map at the top right of Figure 3 will have the highest index, corresponding to a high intensity, whilst that at the bottom left will have the lowest index, corresponding to a low intensity.
- Each of the sub-maps or groups encompasses a region of the frequency map and each point within the total space falls within a sub-map.
- a first level look-up index table T 0 is created which contains every possible point in the space, where each point is mapped to the index for the sub- map in which that point is located. For example, where the pixel intensities range from 0 to 255, the index table T 0 will contain 256 X 256 entries.
- a 2 X 1 pixel reference patch is then created by obtaining the mean intensity values for the right and left pixels of all of the pixel pairs which comprise that sub-map.
- These reference patches are then stored as a first level codebook B 0 , where the location in the codebook of a reference patch is pointed to by the associated index.
- the next step in the codebook creation process is to sub-divide the test image into pairs of horizontally adjacent pixels and to replace each of these pairs with the associated first level index. This results in an image which is shrunk from left to right as shown in Figure 4.
- An important point to note is that, because the indices preserve the intensity order of the sub-maps, the first level shrunk image remains recognisable.
- a second level frequency map is next created from the first level shrunk image, only this time every pair of vertically adjacent pixels are extracted and the intensities of the upper and lower pixels plotted against each other. This is illustrated in Figure 5, which also shows the result of partitioning the map in the way set out above. Indices are again dynamically assigned.
- each pixel in the first level shrunk image is an index which is mapped to a two pixel reference patch in the first level codebook B 0 and that each entry in the second level frequency map therefore corresponds to a 2 X 2 patch of pixels.
- a 2 X 2 pixel reference patch can be created by determining the mean values for the four pixels from each of the patches in that map.
- a second level index look-up table T 0 is constructed which contains every possible combination of vertically adjacent first level indices together with the associated second level indices.
- a second level codebook is then constructed to contain the reference patches, where the location of a reference patch is pointed to by the associated second level index.
- the process can be iteratively repeated to create a third level codebook (4 X 2 pixel patches) , a fourth level codebook (4 X 4 pixel patches) etc.
- a surprising advantage of the present invention is that it may provide a compressed representation of a data array which is tolerant to the loss of least significant bit (lsb) data from index byte-integers .
- lsb least significant bit
- these indices preserve the subjective order of the codebook patches which they represent.
- the patches pointed to by indices 00110010 and 00110011 will tend to be similar so that an error in the lsb of an index, whilst resulting in the selection of the incorrect patch for the decompressed image, will not substantially distort that image.
- the method of the present invention is ideally suited to transmission by such transmission systems.
- the tolerance of the compressed representations resulting from the present invention enables increased compression ratios to be obtained merely by removing one or more lsb's from the indices .
- the method set out above involves repeatedly partitioning the frequency maps at each level into two using either a horizontal or a vertical axis. By performing the partitioning along the axis of greatest variance it is ensured that the expected distances between the map entries and the means of the new sub-maps is lower than would be the case if the sub-maps are partitioned along the axis of least variance. This principle can be exploited further by determining the variance along more than two axes. One practical method of achieving this is to calculate the Principal Component of the frequency map. This is described in 'Negative feedback as an organising principle for Artificial Neural Nets' , Ph. D. Thesis, Strathclyde University, UK. The map is partitioned by a line extending at right angles to the Principal Component and passing through the mean of the Component. The process is repeated for each sub-map.
- colour image frames may be compressed by considering each primary colour separately.
- the same codebook may be used for each colour.
- compression is performed on image patches whose mean and variance values have been pre-normalised.
- the codebook indices are stored/transmitted along with the mean and variance values of the original patches. Codebooks for use with these methods comprise reference patches which also have their mean and variance values normalised.
- codebooks of this type When preparing codebooks of this type it is appropriate to avoid biasing the frequency map with a lot of entries drawn from patches having a low variance since, when a patch of low variance is sent, most of the transmitted information is encoded in the value for the mean of the patch.
- the construction of codebooks can therefore be improved by pre-processing the training set to remove patches having a low variance. This can readily be done by making the probability that a patch is retained in the training set to be some increasing function of its variance .
Abstract
Description
Claims
Priority Applications (6)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP51907598A JP2001505730A (en) | 1996-10-23 | 1997-10-15 | How to generate a vector quantization codebook |
EP97909426A EP0934662B1 (en) | 1996-10-23 | 1997-10-15 | Vector quantisation codebook generation method |
AT97909426T ATE223636T1 (en) | 1996-10-23 | 1997-10-15 | METHOD FOR GENERATING A VECTOR QUANTIZATION CODEBOOK |
DE69715233T DE69715233D1 (en) | 1996-10-23 | 1997-10-15 | METHOD FOR PRODUCING A VECTOR QUANTIZATION CODEBOOK |
CA002269820A CA2269820A1 (en) | 1996-10-23 | 1997-10-15 | Vector quantisation codebook generation method |
US09/284,826 US6438268B1 (en) | 1996-10-23 | 1997-10-15 | Vector quantization codebook generation method |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
GB9622055.3 | 1996-10-23 | ||
GBGB9622055.3A GB9622055D0 (en) | 1996-10-23 | 1996-10-23 | Vector quantisation |
Publications (1)
Publication Number | Publication Date |
---|---|
WO1998018262A1 true WO1998018262A1 (en) | 1998-04-30 |
Family
ID=10801837
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/GB1997/002841 WO1998018262A1 (en) | 1996-10-23 | 1997-10-15 | Vector quantisation codebook generation method |
Country Status (10)
Country | Link |
---|---|
US (1) | US6438268B1 (en) |
EP (1) | EP0934662B1 (en) |
JP (1) | JP2001505730A (en) |
KR (1) | KR20000052771A (en) |
CN (1) | CN1238100A (en) |
AT (1) | ATE223636T1 (en) |
CA (1) | CA2269820A1 (en) |
DE (1) | DE69715233D1 (en) |
GB (1) | GB9622055D0 (en) |
WO (1) | WO1998018262A1 (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR100613106B1 (en) | 1999-09-17 | 2006-08-17 | 삼성전자주식회사 | An index assignment method for coding based on Tree-Structured Vector Quantization |
WO2009127097A1 (en) * | 2008-04-16 | 2009-10-22 | Huawei Technologies Co., Ltd. | Method and apparatus of communication |
JP2012053880A (en) * | 1999-04-30 | 2012-03-15 | E I Du Pont De Nemours & Co | Method for distributed hierarchical evolutionary modeling and visualization of empirical data |
CN103905149A (en) * | 2012-12-27 | 2014-07-02 | 联发科技(新加坡)私人有限公司 | Methods for designing codebook, communication device and codebook training method |
Families Citing this family (29)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
GB0002623D0 (en) * | 2000-02-05 | 2000-03-29 | Res Councils | Improvements in or relating to data compression |
US7518652B2 (en) * | 2000-05-03 | 2009-04-14 | Aperio Technologies, Inc. | Method and apparatus for pre-focus in a linear array based slide scanner |
US6711283B1 (en) | 2000-05-03 | 2004-03-23 | Aperio Technologies, Inc. | Fully automatic rapid microscope slide scanner |
US7668362B2 (en) * | 2000-05-03 | 2010-02-23 | Aperio Technologies, Inc. | System and method for assessing virtual slide image quality |
US7738688B2 (en) * | 2000-05-03 | 2010-06-15 | Aperio Technologies, Inc. | System and method for viewing virtual slides |
WO2002017538A2 (en) * | 2000-08-18 | 2002-02-28 | The Regents Of The University Of California | Fixed, variable and adaptive bit rate data source encoding (compression) method |
JP4393720B2 (en) * | 2001-01-31 | 2010-01-06 | 富士通株式会社 | Pattern recognition apparatus and method |
US8023563B2 (en) * | 2002-10-04 | 2011-09-20 | Stmicroelectronics S.R.L. | Method and system for processing signals via perceptive vectorial quantization, computer program product therefor |
US7257268B2 (en) * | 2003-02-28 | 2007-08-14 | Aperio Technologies, Inc. | Systems and methods for image pattern recognition |
US7116440B2 (en) | 2003-02-28 | 2006-10-03 | Aperio Technologies, Inc. | Image processing and analysis framework |
US20050201470A1 (en) * | 2004-03-12 | 2005-09-15 | John Sievers | Intra block walk around refresh for H.264 |
WO2005119575A2 (en) | 2004-05-27 | 2005-12-15 | Aperio Technologies, Inc | Systems and methods for creating and viewing three dimensional virtual slides |
DE602004026645D1 (en) * | 2004-07-23 | 2010-05-27 | Telecom Italia Spa | METHOD FOR PRODUCING AND USING A VECTOR CODE BOOK, METHOD AND DEVICE FOR COMPRESSING DATA AND DISTRIBUTING LANGUAGE IDENTIFICATION SYSTEM |
US7787674B2 (en) * | 2005-01-27 | 2010-08-31 | Aperio Technologies, Incorporated | Systems and methods for viewing three dimensional virtual slides |
US8164622B2 (en) * | 2005-07-01 | 2012-04-24 | Aperio Technologies, Inc. | System and method for single optical axis multi-detector microscope slide scanner |
FR2893432B1 (en) * | 2005-11-16 | 2008-01-04 | Atmel Corp | VECTOR QUANTIFIER BASED ON SPACE DICHOTOMY WITH N DIMENSIONS |
US7602745B2 (en) * | 2005-12-05 | 2009-10-13 | Intel Corporation | Multiple input, multiple output wireless communication system, associated methods and data structures |
TWI323129B (en) * | 2006-05-17 | 2010-04-01 | Novatek Microelectronics Corp | Block truncation coding (btc) method and apparatus |
US8010555B2 (en) | 2006-06-30 | 2011-08-30 | Aperio Technologies, Inc. | System and method for managing images over a network |
US8086077B2 (en) * | 2006-06-30 | 2011-12-27 | Aperio Technologies, Inc. | Method for storing and retrieving large images via DICOM |
US8165363B2 (en) | 2007-05-04 | 2012-04-24 | Aperio Technologies, Inc. | System and method for quality assurance in pathology |
CN101163239B (en) * | 2007-11-28 | 2010-12-08 | 天津大学 | Novel vector quantization inceptive code book generating method |
CN101281750B (en) * | 2008-05-29 | 2010-12-22 | 上海交通大学 | Expanding encoding and decoding system based on vector quantization high-order code book of variable splitting table |
EP3764085A3 (en) | 2008-10-24 | 2021-03-24 | Leica Biosystems Imaging Inc. | Whole slide fluorescence scanner |
TWI376960B (en) * | 2009-07-31 | 2012-11-11 | Univ Nat Pingtung Sci & Tech | Codebook generating method for image compression |
US8098247B2 (en) * | 2009-09-24 | 2012-01-17 | Crucs Holdings, Llc | Systems and methods for geometric data compression and encryption |
EP2510494B1 (en) | 2009-12-11 | 2021-12-22 | Leica Biosystems Imaging, Inc. | Improved signal to noise ratio in digital pathology image analysis |
US9710695B2 (en) | 2013-03-15 | 2017-07-18 | Sony Corporation | Characterizing pathology images with statistical analysis of local neural network responses |
JP6434162B2 (en) * | 2015-10-28 | 2018-12-05 | 株式会社東芝 | Data management system, data management method and program |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP0444717A2 (en) * | 1982-06-11 | 1991-09-04 | Mitsubishi Denki Kabushiki Kaisha | Vector quantizer |
US5300931A (en) * | 1987-09-04 | 1994-04-05 | Unisys Corporation | Memory based vector quantization |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5371544A (en) * | 1992-02-07 | 1994-12-06 | At&T Corp. | Geometric vector quantization |
-
1996
- 1996-10-23 GB GBGB9622055.3A patent/GB9622055D0/en active Pending
-
1997
- 1997-10-15 CN CN97199861A patent/CN1238100A/en active Pending
- 1997-10-15 JP JP51907598A patent/JP2001505730A/en active Pending
- 1997-10-15 EP EP97909426A patent/EP0934662B1/en not_active Expired - Lifetime
- 1997-10-15 KR KR1019990703579A patent/KR20000052771A/en not_active Application Discontinuation
- 1997-10-15 WO PCT/GB1997/002841 patent/WO1998018262A1/en not_active Application Discontinuation
- 1997-10-15 AT AT97909426T patent/ATE223636T1/en not_active IP Right Cessation
- 1997-10-15 DE DE69715233T patent/DE69715233D1/en not_active Expired - Lifetime
- 1997-10-15 CA CA002269820A patent/CA2269820A1/en not_active Abandoned
- 1997-10-15 US US09/284,826 patent/US6438268B1/en not_active Expired - Lifetime
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP0444717A2 (en) * | 1982-06-11 | 1991-09-04 | Mitsubishi Denki Kabushiki Kaisha | Vector quantizer |
US5300931A (en) * | 1987-09-04 | 1994-04-05 | Unisys Corporation | Memory based vector quantization |
Non-Patent Citations (2)
Title |
---|
NASRABADI N M ET AL: "A MULTILAYER ADDRESS VECTOR QUANTIZATION TECHNIQUE", IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, vol. 37, no. 7, 1 July 1990 (1990-07-01), NEW YORK, pages 912 - 921, XP000160478 * |
RISKIN E A ET AL: "INDEX ASSIGNMENT FOR PROGRESSIVE TRANSMISSION OF FULL-SEARCH VECTOR QUANTIZATION", IEEE TRANSACTIONS ON IMAGE PROCESSING, vol. 3, no. 3, 1 May 1994 (1994-05-01), NEW YORK, pages 307 - 312, XP000453368 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2012053880A (en) * | 1999-04-30 | 2012-03-15 | E I Du Pont De Nemours & Co | Method for distributed hierarchical evolutionary modeling and visualization of empirical data |
KR100613106B1 (en) | 1999-09-17 | 2006-08-17 | 삼성전자주식회사 | An index assignment method for coding based on Tree-Structured Vector Quantization |
WO2009127097A1 (en) * | 2008-04-16 | 2009-10-22 | Huawei Technologies Co., Ltd. | Method and apparatus of communication |
US8364476B2 (en) | 2008-04-16 | 2013-01-29 | Huawei Technologies Co., Ltd. | Method and apparatus of communication |
CN103905149A (en) * | 2012-12-27 | 2014-07-02 | 联发科技(新加坡)私人有限公司 | Methods for designing codebook, communication device and codebook training method |
CN103905149B (en) * | 2012-12-27 | 2017-12-29 | 联发科技(新加坡)私人有限公司 | Codebook design method, communication equipment and codebook training method |
Also Published As
Publication number | Publication date |
---|---|
EP0934662A1 (en) | 1999-08-11 |
US6438268B1 (en) | 2002-08-20 |
CA2269820A1 (en) | 1998-04-30 |
DE69715233D1 (en) | 2002-10-10 |
JP2001505730A (en) | 2001-04-24 |
CN1238100A (en) | 1999-12-08 |
GB9622055D0 (en) | 1996-12-18 |
KR20000052771A (en) | 2000-08-25 |
ATE223636T1 (en) | 2002-09-15 |
EP0934662B1 (en) | 2002-09-04 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
EP0934662B1 (en) | Vector quantisation codebook generation method | |
Jacquin | Fractal image coding based on a theory of iterated contractive image transformations | |
Nasrabadi et al. | A multilayer address vector quantization technique | |
Caso et al. | Fast methods for fractal image encoding | |
CN111954000B (en) | Lossless compression method for high-speed toll collection picture set | |
Memon et al. | Lossless image compression with a codebook of block scans | |
Li et al. | Fractal color image compression | |
Kubrick et al. | Classified vector quantisation of images: codebook design algorithm | |
Riccio et al. | Deferring range/domain comparisons in fractal image compression | |
Agahian et al. | Spectral compression using subspace clustering | |
Hooda et al. | Tree based search algorithm for binary image compression | |
Markas et al. | Image compression methods with distortion controlled capabilities | |
Chang et al. | Image coding by a neural net classification process | |
Krishnamoorthi et al. | Codebook generation for vector quantization on orthogonal polynomials based transform coding | |
JPH10173538A (en) | Method for speeding up vector quantization compression | |
Jaseela et al. | A new approach to fractal image compression using DBSCAN | |
Agahian et al. | Subspace-clustering-based multispectral image compression | |
JP2693557B2 (en) | Encoding device | |
Tsay et al. | Image compression using VQ and fuzzy classified algorithm | |
Wang et al. | Adaptively hybrid fractal image coding | |
Kim et al. | Vector quantizer of medical image using wavelet transform and enhanced neural network | |
Quweider | New classification, codebook design and encoding algorithms for vector quantization of images | |
Kim et al. | Medical image vector quantizer using wavelet transform and enhanced SOM algorithm | |
El-Harby | Two Improved Color Images Compression Systems | |
Saranya et al. | An efficient image compression algorithm to improve the compression-ratio of dual-color image |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
WWE | Wipo information: entry into national phase |
Ref document number: 97199861.2 Country of ref document: CN |
|
AK | Designated states |
Kind code of ref document: A1 Designated state(s): CA CN JP KR US |
|
AL | Designated countries for regional patents |
Kind code of ref document: A1 Designated state(s): AT BE CH DE DK ES FI FR GB GR IE IT LU MC NL PT SE |
|
DFPE | Request for preliminary examination filed prior to expiration of 19th month from priority date (pct application filed before 20040101) | ||
121 | Ep: the epo has been informed by wipo that ep was designated in this application | ||
ENP | Entry into the national phase |
Ref document number: 2269820 Country of ref document: CA Ref document number: 2269820 Country of ref document: CA Kind code of ref document: A |
|
ENP | Entry into the national phase |
Ref document number: 1998 519075 Country of ref document: JP Kind code of ref document: A |
|
WWE | Wipo information: entry into national phase |
Ref document number: 1019997003579 Country of ref document: KR |
|
WWE | Wipo information: entry into national phase |
Ref document number: 1997909426 Country of ref document: EP |
|
WWE | Wipo information: entry into national phase |
Ref document number: 09284826 Country of ref document: US |
|
WWP | Wipo information: published in national office |
Ref document number: 1997909426 Country of ref document: EP |
|
WWP | Wipo information: published in national office |
Ref document number: 1019997003579 Country of ref document: KR |
|
WWG | Wipo information: grant in national office |
Ref document number: 1997909426 Country of ref document: EP |
|
WWW | Wipo information: withdrawn in national office |
Ref document number: 1019997003579 Country of ref document: KR |