WO2003071687A1 - Data processing device - Google Patents
Data processing device Download PDFInfo
- Publication number
- WO2003071687A1 WO2003071687A1 PCT/JP2003/001629 JP0301629W WO03071687A1 WO 2003071687 A1 WO2003071687 A1 WO 2003071687A1 JP 0301629 W JP0301629 W JP 0301629W WO 03071687 A1 WO03071687 A1 WO 03071687A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- template
- vector
- pattern
- input
- difference
- Prior art date
Links
Classifications
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
- G06V10/75—Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
- G06V10/751—Comparing pixel values or logical combinations thereof, or feature values having positional relevance, e.g. template matching
-
- 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
-
- 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
- H03M7/3084—Compression; Expansion; Suppression of unnecessary data, e.g. redundancy reduction using adaptive string matching, e.g. the Lempel-Ziv method
-
- 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
Definitions
- the present invention relates to a data processing device for compressing data by vector quantization.
- Vector quantization is a technique used for data compression processing of image data and audio data, for example.
- image data corresponding to each block of a predetermined size for example, 4 ⁇ 4 pixels
- each image data is converted into 4 pixels. If it is a block of X 4 pixels, it is treated as a 16-dimensional input vector.
- several pixel patterns (template vectors) having the same size as the block from which the image data is to be extracted are prepared in advance.
- the input vector is compared with the template vector, and a template vector most similar to the input vector is found from among the plurality of template vectors. Specifically, the distance between the input vector data and the template vector data is calculated for each template vector data, and as a result, the template vector with the smallest distance has the highest similarity. And thus, in the vector quantization, image data compression is performed by quantizing the input vector using the template vector.
- betattles One way to solve this problem is to use the features of betattles and reduce the amount of computation in betattle quantization.
- This method uses the characteristics of the input vector. The collected amount is compared with the feature amount of the template vector obtained in advance, and the number of template vectors to be searched (to be subjected to distance calculation) is reduced using the comparison result. As a result, the amount of computation such as distance computation in beta quantization can be reduced, and beta quantization can be accelerated.
- the template vector data is searched for all template vector data, so that the input vector data and the template vector data are Even if the pattern is clearly different, the distance calculation is performed (started), and there is a problem that a lot of useless calculation is performed.
- the present invention has been made to solve such a problem, and an object of the present invention is to enable data compression processing using vector quantization to be executed at high speed by hardware. Disclosure of the invention
- the data processing device of the present invention is a data processing device that outputs information indicating the template pattern most similar to an input pattern from a plurality of template patterns prepared in advance, and searches for a similar template pattern.
- a first search means for selecting a template pattern for calculating a similarity between the input pattern and the template pattern using the feature amounts of the input pattern and the template pattern, and a first search means
- a second search means for executing the calculation in a bit serial manner is provided.
- the second search means executes the operation on the data input by bit serial in the operation unit time, and outputs the operation result for each operation unit time. It is characterized in that it is determined whether or not the similarity is maximized based on this.
- the first search means reduces the number of template patterns for calculating similarity using the feature amounts of the input pattern and the template pattern
- the searching means is characterized in that a similarity calculation between the input pattern and the template pattern selected by the first searching means is executed in a bift serial manner, and the number of steps required for the calculation is reduced.
- the input pattern and the template pattern are one vector data composed of at least one or more elements.
- the input pattern and the template pattern are composed of at least one or more elements
- the second search means is configured to determine the similarity between the input pattern and the template pattern.
- a similarity calculation judging means for judging not to execute.
- Another feature of the data processing device of the present invention is that the similarity calculating means is
- Both the template pattern selected by the search means (1) and the input pattern are bit-serial input from the upper digit side of the data for each calculation time unit, and the input data is calculated within the calculation time unit and calculated. It is characterized in that the resulting difference absolute value distance is output in bit serial from the upper digit side.
- the feature amount is a sum of individual elements constituting the input pattern and the template pattern, and a value of a value that can take a part of the individual elements. It is characterized by being at least one of the sum of the elements after the inversion operation based on the intermediate value.
- the similarity calculating means is constituted by a semiconductor circuit, the input data is input from the upper digit side by a binary number system, and the output data is redundant. It is characterized by output from the upper digit side using a simple numerical system.
- the similarity calculation means calculates difference absolute values of elements of the input pattern and the template pattern for each element, and a difference absolute value calculation means. Adding means for adding the difference absolute values for each element calculated by the means, wherein the difference absolute value calculating means is provided by the number of elements of the pattern.
- the absolute difference calculating means receives data from the upper digit side in a binary number system, calculates the difference value of the input data in a bit serial manner, It is characterized in that it is output from the high-order send side using a number system.
- the difference absolute value calculating means further stores, as a state, a value other than "0" that first appears in the difference value output by bit serial from the upper digit. And a function of inverting or non-inverting the difference value according to the stored state to calculate the absolute value of the difference.
- the addition operation unit includes an operation unit. It has a delay means for delaying by the order time, data is input from the upper digit side by a redundant number system, and the operation result is output from the upper digit side by a redundant number system.
- the second search means includes: holding means for holding a minimum difference absolute value distance between the input pattern and the template pattern; and a minimum value held by the holding means. The difference absolute value distance is input by bit serial from the upper digit side, and the difference absolute value distance obtained by the similarity calculation means for calculating the difference absolute value distance between the input pattern and a template pattern different from the above template pattern.
- FIG. 1 is a block diagram illustrating a configuration example of a data processing device according to the first embodiment.
- FIG. 2 is a diagram for explaining a process performed by the first search unit.
- FIG. 3 is a diagram illustrating another example of the feature amount extraction process performed by the first search unit.
- FIG. 4 is a diagram illustrating an example of a feature amount extraction process when a plurality of inversion patterns are used.
- FIG. 5 is a diagram showing the ratio of the number of times of calculating the absolute difference distance when five types of inversion patterns are used.
- FIG. 6 is a diagram illustrating an example of a circuit block for calculating a difference absolute value distance.
- FIG. 7 is a diagram illustrating an example of a circuit block of the difference absolute value calculation unit.
- FIG. 8 is a diagram illustrating an example of a circuit block of the addition operation unit.
- FIG. 9 is a diagram illustrating an example of a circuit block of an addition operation unit in which the number of input data is expanded to four.
- FIG. 10 is a diagram illustrating a truth table of each addition unit of the addition operation unit.
- FIG. 11 is a diagram showing a state transition diagram for finding the minimum value.
- FIG. 12 is a block diagram illustrating a configuration example of a minimum value search operation unit.
- FIG. 13 is a flowchart illustrating a process according to the first embodiment.
- FIG. 14 is a flowchart illustrating a process according to the first embodiment.
- FIG. 15 is a diagram showing a ratio of the number of times of calculating the absolute difference distance in the second embodiment when five types of inversion patterns are used.
- a data processing device shows, as an example, a data processing device that implements vector quantization, which is one of pattern matching.
- vector quantization which is one of pattern matching.
- the template vector most similar to the input vector is equivalent to the operation to find the template vector with the smallest difference absolute value distance.
- FIG. 1 is a block diagram showing one configuration example of the data processing device according to the first embodiment of the present invention.
- a first search unit 110 performs a process for reducing the number of templates to be searched in a pattern matching process (vector quantization). It is constituted by a similarity calculation judging unit 102.
- the feature value calculation unit 101 calculates a feature value of the input vector using the input vector.
- the similarity calculation determining unit 102 compares the feature amount of the template vector stored in the feature amount storage unit 104 with the feature amount of the input vector calculated by the feature amount calculating unit 101. Then, it is determined whether or not to execute the similarity calculation (distance calculation).
- the similarity calculation unit 103 executes similarity calculation (distance calculation) between the input vector and the template vector according to the determination result by the similarity calculation determination unit 102.
- the feature amount storage unit 104 stores a feature amount of a template vector calculated in advance, and the template storage unit 105 stores a plurality of template vectors. is there.
- the maximum similarity search unit 106 receives the similarity calculated by the similarity calculation unit 103 and the maximum similarity stored in the maximum similarity storage unit 107 up to the present. Compare. After completing the processing for all template vectors in the template storage unit 105, the address designating unit 108 determines the similarity based on the comparison result in the maximum similarity search unit 106. Outputs the address of the largest template vector (similar to the input vector) as an index.
- FIG. 2 is a diagram for explaining processing by the first search unit shown in FIG.
- the feature amount of the input vector and the template vector is the sum of the elements of each vector (predetermined size, for example, the data value of each pixel included in a 4 ⁇ 4 pixel block).
- predetermined size for example, the data value of each pixel included in a 4 ⁇ 4 pixel block.
- Equation (1) indicates that the absolute difference distance between the two vectors is greater than the absolute value of the force ⁇ or the absolute value of the difference between the element sums of those vectors. . That is, the method using the above equation (1) is a method of extracting and using the entire density of a block as a feature value.
- Step 3 Next, to calculate the sum of the second template vector T 2 elements, the difference between the sum of the input base-vector I of elements in the similarity calculation determining unit 1 0 2. Further, the similarity calculation determination unit 102 compares the calculated difference between the sums of the elements and the minimum difference absolute value distance stored in the maximum similarity storage unit 107.
- the similarity calculation determination unit 102 determines that the difference in the sum of the elements is smaller than the minimum difference absolute value distance, or that the difference between the minimum difference absolute value distance and the sum of the elements is equal. It is determined that the similarity calculation section 103 performs similarity calculation.
- the similarity calculating unit 1 0 3 calculates the difference component absolute value distance template base vector T 2 and the input base vector I.
- the maximum similarity search unit 106 compares the difference absolute value distance calculated by the similarity calculation unit 103 with the minimum difference absolute value distance. As a result, if the difference absolute value distance calculated by the similarity calculation unit 103 is smaller than the minimum difference absolute value distance, the index of this template vector is stored, and the similarity calculation unit 103 stores the index. The difference absolute value distance calculated as described above is stored and updated in the maximum similarity storage unit 107 as the minimum difference absolute value distance.
- the similarity calculation determination unit 102 determines that the similarity calculation unit 103 does not execute the similarity calculation, and determines the difference. Starts the search for the next template vector without performing absolute distance calculation.
- Step 4 Step 3 is repeated to the end for all template vectors stored in the template storage unit 105.
- the processing is executed as described above, and the first search unit 110 searches for By selecting the template vector based on the feature amount, the number of matchings between the input vector and the template vector can be reduced.
- the calculation of the difference of the sum of the elements is performed for all template vectors.
- the difference absolute value distance operation is an n-dimensional betattle operation. In the case of 2, a large number of n-dimensional vector operations can be omitted by simple scalar operations, and the amount of operations can be reduced.
- the difference between the sum of the elements of the input vector I and the sum of the elements of the (k-1) th template vector 1 ⁇ — is "53".
- the absolute difference distance from the template vector Tk-i was "94" (in Fig. 2, it is assumed to be the minimum absolute difference distance for the (k-1) th template vector). I do.
- the difference between the sum of the elements of the input vector I and the sum of the elements of the k-th template vector T k is “1 8”, and the input vector I and the template vector T k —
- the difference absolute value distance from i is smaller than “94”, so the difference absolute value distance between input vector I and template vector T k is calculated. Since the calculated difference absolute value distance is "2 1" and smaller than the minimum difference absolute value distance, the minimum difference absolute value distance is updated to "2 1".
- the difference between the sum of the elements of the input vector I and the sum of the elements of the (k ⁇ 11) th template vector T k + 1 is “42”, and the input vector that is the minimum difference absolute value distance Since the absolute difference distance between I and the template vector T k + 1 is greater than "2 1", the absolute difference distance between the input vector 1 and the template vector T k + 1 must not be calculated. In this way, for the template vector T k + 1 , only a simple one-dimensional scalar operation is executed, and the process moves to the next template vector without executing an n-dimensional vector operation.
- the sum of the elements of the vector is used as the feature value, but the feature value is not limited to this.
- a difference between the sum of the inverted elements of the input vector and the template vector as described below may be used as the feature amount.
- Input vector An inequality such as equation (2) holds between the absolute difference distance between the template vector and the template vector and the difference between the sums of the inverted elements. ⁇ , + ⁇ (255-/-) I (n> m)
- Expression (2) above shows, as an example, a case where the data value of each pixel is a value in the range of “0” to “255”.
- inversion means inversion of black and white around an intermediate value. For example, if the data value of each pixel is in the range of "0" to "255", the center value is "127.5", so when inverted, the data value "128” becomes “127” Then, the data value "100” becomes “155”. In the example described below, inversion is performed by subtracting the data value of the inverted pixel (pixel) from "255".
- the inversion is not performed for all the pixels, but is partially performed.
- non-inversion is performed for the 0th to mth pixels, and inversion is performed for (m10th) to nth pixels.
- the numbers need not be serial numbers.
- the pixel to be inverted must have the same element number in both the input vector and the template vector.
- the method using the above equation (2) is a method of extracting and using the feature of the luminance change direction in the block.
- Fig. 3 shows an example of feature value extraction for using the relationship shown in equation (2).
- a specific inversion pattern is prepared in advance.
- the inverted pattern is represented by two values, “white” and “black”. The “white” part is non-inverted, and the “black” part is inverted.
- the feature value calculation unit 101 inverts the pixel corresponding to the input vector inversion using the inversion pattern, and obtains the sum of the elements after the inversion.
- the template vector is also inverted using a similar inversion pattern, the sum of the elements is obtained, and the sum is stored in the feature amount storage unit 104.
- the sum of the elements of the input vector after inversion obtained as described above and the The difference from the sum of the elements of the template vector is determined by the similarity calculation determining unit 102.
- the difference value of the sum of the elements is “1 5 5 2”.
- the difference absolute value distance between the input vector and the template vector is always “1 5 5 It is equal to or greater than 2 ".
- FIG. 4 shows an example in which a plurality (five types) of inverted patterns are used.
- the inversion pattern is represented by two values of “white” and “black”, where “white” is non-inversion and “black” is inversion.
- the input vector and template vector are inverted using the respective inversion patterns, and the sum of the inverted input vector elements and the inverted template vector elements are inverted.
- the difference from the total sum of is calculated for each inversion pattern.
- the absolute difference distance between the input vector and the template vector is equal to or greater than the difference value of the sum of the elements according to the above equation (2). Find the value and use it as the feature value.
- the maximum value of the difference between the sum of the elements of the input vector after inversion and the sum of the elements of the template vector after inversion is "1252".
- the difference absolute value distance between the input vector and the template vector is always equal to or larger than "1252".
- the amount of calculation can be further reduced by increasing the number of inversion patterns used for extracting the feature amount.
- Figure 5 shows the number of times the calculation was performed on the absolute difference distance between the input vector and the template vector when vector quantization was performed on the two types of still images using five types of inversion patterns.
- the codebook search rate is expressed as a percentage of the number of template vectors obtained by calculating the absolute difference distance for the number of template vectors stored in the template storage unit 105. Things.
- the size of the codebook (the total number of template vectors stored in the template storage unit 105) is 24048.
- vector quantization can significantly reduce the time required for data compression, and can speed up data compression using vector quantization.
- the similarity is considered as a difference absolute value distance.
- the calculation of the similarity that is, the difference absolute value distance, is executed by the similarity calculating section 103 shown in FIG.
- FIG. 6 is an example of a circuit block diagram for calculating a difference absolute value distance.
- reference numeral 6001 denotes a difference absolute value calculation unit, which calculates a difference absolute value distance for each element of the vector.
- Reference numeral 62 denotes an addition operation unit, which calculates and outputs the sum of the difference absolute value distances for each element calculated by the difference absolute value operation unit 601.
- the dimensions of the vector are 16 dimensions, but the present invention is not limited to this, and the number of dimensions can be increased.
- FIG. 7 is a diagram showing an example of a circuit block of the absolute difference value calculating section 600 shown in FIG.
- the difference absolute value calculation unit 6001 shown in FIG. 7 inputs data in a bit serial manner from the upper digit side, performs an arithmetic operation within an operation unit time, and outputs the difference absolute value as a calculation result from the upper digit side.
- the absolute difference calculation unit inputs binary data and outputs the result in binary ternary SD number system.
- the binary ternary SD number system is a number system that allows "1-1, 0, 1" in each digit and a number system that allows a code set.
- the difference absolute value calculation unit includes a block for performing the difference calculation and a block for performing the absolute value calculation.
- the block is composed of two blocks.
- the difference calculation block 701 performs subtraction for each digit from the upper digit, and outputs the subtraction result for each digit.
- the result of the subtraction is latched by the temporary storage register 702 and input to the state control block 703.
- the state control block 703 is a state machine that checks the subtraction result from the most significant digit and sends a sign inversion request signal to the sign inversion block if the first non-zero digit is negative. is there.
- the absolute value calculation block 704 performs inversion to obtain an absolute value because the difference value is negative.
- the sign inversion request signal is not output, the data is output through (as is) because the difference value is "0" or positive.
- the input vector and template vector are input in 16-bit bit serial from the upper digit side, and the calculation unit is The operation is performed in time, and the absolute value of the difference, which is the operation result, can be output bit-serial from the upper digit.
- the difference absolute value distance can be calculated by adding all the difference absolute values output from the higher-order digit side from the difference absolute-value calculating unit 6001 from the higher-order digit side in the addition calculating unit.
- a binary ternary SD number system is used.
- the propagation of the carry signal of the addition operation has a feature that it can be accommodated in a maximum of two digits.
- FIG. 8 is a diagram showing an example of a circuit block of the addition operation unit shown in FIG.
- the addition operation unit shown in FIG. 8 inputs data from the upper digit, performs an operation within the operation unit time, and outputs data from the upper digit.
- a structure using two delay elements internally to refer to information below two digits is adopted. I do.
- X N and Y N are input data
- Z N +2 is output data.
- the addition operation unit shown in FIG. 8 shows a case where the number of input data is two. It is also possible to increase the number of force inputs.
- FIG. 9 is a diagram illustrating an example of a circuit block of an addition operation unit in which the number of input data is expanded to four. Each adder may be designed so as to satisfy the truth table shown in FIG. By using the addition unit and the absolute value calculation unit with the number of input data expanded to 16, the absolute difference distance between two 16-dimensional vectors can be calculated from the upper digit side . In the binary ternary SD number system, the negative representation of the value can be arbitrarily designed by expressing one digit with several bits. Next, the maximum similarity search unit 106 will be described.
- the magnitude of the binary ternary SD number system described above can be determined when the numerical value is examined from the upper digit side and a difference of "2" or more appears at a certain digit or more. For two numbers, if one number is greater than the m-th digit by "2", then that number is greater than the other number.
- FIG. 12 shows a minimum value search operation unit based on the above principle.
- the sequential circuit 1201 that temporarily holds the state according to the state transition diagram 1201 is a “State Indicator”, and the circuit 1202 for finding the minimum value of the currently viewed digit of the number in the “Smaller” state is a “Minimum value observer”.
- the circuit that encodes the position of Winner J 1203 is the "Winner observer". According to this palindrome configuration, the number of inputs can be increased freely by preparing a State Indicator "1 201" for each input.
- “State Indicator” 1 201 is reset to "Smaller" in the initial state.
- the data to be compared is input serially starting from the most significant digit, and when one is in the rsmallerj state and all the others are in the "Larger” state, the "Winner observer” block 1223 outputs the comparison result and the decision end flag. Output.
- Addr is the address of the template vector
- N is the total number of template vectors
- minA is the address of the template vector of the search candidate
- min D is the absolute value of the minimum difference.
- Distance F i is the feature quantity of the input vector
- F t is the feature quantity of the template vector
- D is the absolute difference value of the feature quantity
- D ist is the difference absolute value distance being calculated.
- One of the methods of vector quantization is to remove the average value of blocks as a process before performing pattern matching.
- a vector X is divided into its average value m and the residual R from which the average value m has been removed, and pattern matching is performed using the residual R as a new vector.
- I do. m ⁇ Xi ...... (3)
- the method of reducing the amount of computation using the relationship shown in the above equation (1) is a method utilizing the large variation in the average value information of the vector. It does not work effectively when the variation in information is uniform.
- Fig. 15 shows the vector quantum in the block from which the information of the average value has been removed, using the five types of inverted patterns for one or two types of still images by using the calculation omission method according to the first embodiment described above.
- 9 is a graph showing the ratio of the calculation of the absolute difference value distance when the conversion is performed. Note that in Fig. 13 three codebook sizes are used. ing.
- the feature amount of the input pattern By using it, the number of templates that need to be matched can be reduced. Also, by performing the operation for matching from the upper digit side, the number of operation cycles can be reduced. By combining the two operation omission processes in this manner, a high operation omission rate can be obtained, and the data compression process using vector quantization can be executed at high speed.
- the present invention can be applied to a communication system that compresses and transmits image or audio data at high speed and transmits the data, and its transmission device.
- the data processing device shown in FIG. 1, particularly the first search unit for calculating the similarity is constituted by a semiconductor circuit
- the present invention does not The present invention is not limited to this, and may be configured by a CPU that executes a program that defines the operation of the data processing device. Further, the processing shown in FIGS. 13 and 14 may be realized by a computer program.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- Medical Informatics (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Databases & Information Systems (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Signal Processing (AREA)
- Compression, Expansion, Code Conversion, And Decoders (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
- Compression Or Coding Systems Of Tv Signals (AREA)
Description
Claims
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP03705222A EP1487112A4 (en) | 2002-02-20 | 2003-02-17 | DATA PROCESSING DEVICE |
US10/504,968 US20050105816A1 (en) | 2002-02-20 | 2003-02-17 | Data processing device |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2002-44054 | 2002-02-20 | ||
JP2002044054A JP2003243988A (ja) | 2002-02-20 | 2002-02-20 | データ処理装置 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2003071687A1 true WO2003071687A1 (en) | 2003-08-28 |
Family
ID=27750542
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2003/001629 WO2003071687A1 (en) | 2002-02-20 | 2003-02-17 | Data processing device |
Country Status (4)
Country | Link |
---|---|
US (1) | US20050105816A1 (ja) |
EP (1) | EP1487112A4 (ja) |
JP (1) | JP2003243988A (ja) |
WO (1) | WO2003071687A1 (ja) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111339344A (zh) * | 2020-02-25 | 2020-06-26 | 北京百度网讯科技有限公司 | 室内图像检索方法、装置及电子设备 |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2010091555A1 (zh) * | 2009-02-13 | 2010-08-19 | 华为技术有限公司 | 一种立体声编码方法和装置 |
WO2014132608A1 (ja) | 2013-02-26 | 2014-09-04 | 日本電気株式会社 | 並列処理装置、並列処理方法、および並列処理プログラム記憶媒体 |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH11289459A (ja) * | 1998-04-02 | 1999-10-19 | Matsushita Electric Ind Co Ltd | 画像符号化方法 |
JP2000004164A (ja) * | 1998-04-17 | 2000-01-07 | Tadahiro Omi | ベクトル量子化装置および方法、記録媒体 |
JP2001204024A (ja) * | 2000-01-18 | 2001-07-27 | Tadahiro Omi | データ圧縮方法、データ圧縮装置及び記憶媒体 |
Family Cites Families (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
IT1257073B (it) * | 1992-08-11 | 1996-01-05 | Ist Trentino Di Cultura | Sistema di riconoscimento, particolarmente per il riconoscimento di persone. |
JP2972657B2 (ja) * | 1997-06-20 | 1999-11-08 | 山形日本電気株式会社 | Efm信号のフレーム周期検出回路及びefm信号再生用ビット同期クロック信号の周波数制御方法 |
EP1039372A4 (en) * | 1997-12-17 | 2005-02-02 | Ultraclean Technology Res Inst | METHOD AND SEMICONDUCTOR CIRCUIT FOR PERFORMING ARITHMETIC OPERATIONS |
JP2000201078A (ja) * | 1998-04-17 | 2000-07-18 | Tadahiro Omi | デ―タ圧縮装置および方法、デ―タ伸長装置および方法、デ―タ圧縮伸長システムおよび方法、コ―ドブックの作成方法、記録媒体 |
JP2002064383A (ja) * | 2000-08-18 | 2002-02-28 | Yamaha Corp | Δς変調器 |
US6719689B2 (en) * | 2001-04-30 | 2004-04-13 | Medtronic, Inc. | Method and system for compressing and storing data in a medical device having limited storage |
US6640008B1 (en) * | 2001-06-29 | 2003-10-28 | Shih-Jong J. Lee | Rotation and scale invariant pattern matching method |
-
2002
- 2002-02-20 JP JP2002044054A patent/JP2003243988A/ja active Pending
-
2003
- 2003-02-17 EP EP03705222A patent/EP1487112A4/en not_active Ceased
- 2003-02-17 US US10/504,968 patent/US20050105816A1/en not_active Abandoned
- 2003-02-17 WO PCT/JP2003/001629 patent/WO2003071687A1/ja not_active Application Discontinuation
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH11289459A (ja) * | 1998-04-02 | 1999-10-19 | Matsushita Electric Ind Co Ltd | 画像符号化方法 |
JP2000004164A (ja) * | 1998-04-17 | 2000-01-07 | Tadahiro Omi | ベクトル量子化装置および方法、記録媒体 |
JP2001204024A (ja) * | 2000-01-18 | 2001-07-27 | Tadahiro Omi | データ圧縮方法、データ圧縮装置及び記憶媒体 |
Non-Patent Citations (1)
Title |
---|
See also references of EP1487112A4 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111339344A (zh) * | 2020-02-25 | 2020-06-26 | 北京百度网讯科技有限公司 | 室内图像检索方法、装置及电子设备 |
CN111339344B (zh) * | 2020-02-25 | 2023-04-07 | 北京百度网讯科技有限公司 | 室内图像检索方法、装置及电子设备 |
Also Published As
Publication number | Publication date |
---|---|
EP1487112A1 (en) | 2004-12-15 |
EP1487112A4 (en) | 2005-12-21 |
US20050105816A1 (en) | 2005-05-19 |
JP2003243988A (ja) | 2003-08-29 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
JP6540725B2 (ja) | 演算処理装置、方法、およびプログラム | |
US10872295B1 (en) | Residual quantization of bit-shift weights in an artificial neural network | |
CN114418105B (zh) | 一种基于量子线路处理量子应用问题的方法及装置 | |
EP3769208B1 (en) | Stochastic rounding logic | |
CN112506880A (zh) | 数据处理方法及相关设备 | |
CN114615507B (zh) | 一种图像编码方法、解码方法及相关装置 | |
Perez et al. | Efficient projection algorithms onto the weighted ℓ1 ball | |
CN114640354A (zh) | 数据压缩方法、装置、电子设备及计算机可读存储介质 | |
AU2021269358A1 (en) | Method for determining a preimage element of a cryptographic hash function, computer program, and data processing system | |
WO2003071687A1 (en) | Data processing device | |
US10271051B2 (en) | Method of coding a real signal into a quantized signal | |
CN102067108B (zh) | 并向量分数与符号对称舍入误差的乘积的快速计算 | |
US20060179098A1 (en) | System and method for reduction of leading zero detect for decimal floating point numbers | |
CN111492369A (zh) | 人工神经网络中移位权重的残差量化 | |
US20210216867A1 (en) | Information processing apparatus, neural network computation program, and neural network computation method | |
TWI708196B (zh) | 使用基於累積計數分佈之函數之用於模型參數之解壓縮之方法及處理器 | |
EP2643833B1 (en) | Low complexity target vector identification | |
CN112612762A (zh) | 数据处理方法及相关设备 | |
WO2020245877A1 (ja) | 量子アニーリング計算装置、量子アニーリング計算方法および量子アニーリング計算プログラム | |
JP4405452B2 (ja) | 逆変換回路 | |
JP3951036B2 (ja) | 動的小数点方式、動的小数点演算方法、並びに2次元逆離散コサイン変換方法及びその装置 | |
JP4243277B2 (ja) | データ処理装置 | |
WO2019155556A1 (ja) | 行列圧縮装置、圧縮行列アクセス装置、方法、及びコンピュータ可読媒体 | |
US20210334635A1 (en) | Neural network accelerator configured to perform operation on logarithm domain | |
CN110134813A (zh) | 图像检索方法、图像检索装置及终端设备 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
AK | Designated states |
Kind code of ref document: A1 Designated state(s): US |
|
AL | Designated countries for regional patents |
Kind code of ref document: A1 Designated state(s): AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HU IE IT LU MC NL PT SE SI SK TR |
|
121 | Ep: the epo has been informed by wipo that ep was designated in this application | ||
WWE | Wipo information: entry into national phase |
Ref document number: 10504968 Country of ref document: US |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2003705222 Country of ref document: EP |
|
WWP | Wipo information: published in national office |
Ref document number: 2003705222 Country of ref document: EP |
|
WWR | Wipo information: refused in national office |
Ref document number: 2003705222 Country of ref document: EP |
|
WWW | Wipo information: withdrawn in national office |
Ref document number: 2003705222 Country of ref document: EP |