CN110213582B - High-precision quantitative acceleration method for ultrahigh-resolution image analysis - Google Patents
High-precision quantitative acceleration method for ultrahigh-resolution image analysis Download PDFInfo
- Publication number
- CN110213582B CN110213582B CN201910509308.1A CN201910509308A CN110213582B CN 110213582 B CN110213582 B CN 110213582B CN 201910509308 A CN201910509308 A CN 201910509308A CN 110213582 B CN110213582 B CN 110213582B
- Authority
- CN
- China
- Prior art keywords
- calculation
- result
- resolution image
- image analysis
- variable
- 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
Images
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
- H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
- H04N19/102—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or selection affected or controlled by the adaptive coding
- H04N19/124—Quantisation
-
- 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
- H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
- H04N19/134—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or criterion affecting or controlling the adaptive coding
- H04N19/146—Data rate or code amount at the encoder output
Abstract
The invention provides an ultrahigh resolution image analysis methodThe high-precision quantization acceleration method is used for improving a large amount of intermediate data calculation processes in an ultrahigh-resolution image analysis process through quantization, and is characterized by comprising the following steps of: step S1, obtaining the accuracy alpha required by the final result, and obtaining the accuracy beta required to be reserved by intermediate calculation; step S2, obtaining an intermediate variable X, quantizing the intermediate variable X according to the accuracy beta required to be reserved in the intermediate calculation through the current calculated system gamma to obtain an offset variable X1(ii) a Step S3, based on the offset variable X1Performing calculation operation to obtain a calculation result Y1, and reserving a calculation process queue Q in the calculation process; step S4, calculating the result Y1The accuracy beta needing to be reserved in the intermediate calculation and the final result Y of the final system obtained by the calculation process queue Q.
Description
Technical Field
The invention belongs to the field of image data processing, and particularly relates to a high-precision quantitative acceleration method for ultrahigh-resolution image analysis.
Background
According to the development trend of the present day, the development of big data technology is vigorous, and the big data technology is widely applied to various fields and permeates various fields in life. Due to the development of networks, hardware equipment and the like, the scale of data collected in biomedical engineering, industrial engineering, military public security, aerospace and the like is gradually increased. Taking digital video images as an example, with the development of technology, in order to collect more information, the resolution of video images is higher and higher, the resolution is higher, and the specification of data of each frame will increase. And the number of frames that can be acquired per second is increasing, which will increase the amount of digital video data obtained in an exponential manner.
Data processing technology has been in the past for over 60 years, and as electronic computers have advanced to a certain level, computers have begun to be used to process ever larger amounts of data. However, the exponentially increased data size has not only a storage problem, a consistency problem, a call transfer problem, and the like of the large data itself, but also a problem of how to efficiently analyze and process the large data size data. The main reason is that due to the system application scenario, data processing in many fields requires higher and higher real-time performance, so an acceleration algorithm with universality is necessary in the present computing environment with large data volume. Taking ultra-high-definition image videos widely applied to the aspects of traffic security, biotechnology, life leisure and the like as an example, many steps are needed for a computer to replace human eyes to understand one frame with ultra-high resolution in one ultra-high-definition video. From video imaging, transmission to monitoring devices, display devices and analysis platforms generally first goes through a process of encoding and then decoding, and then the image is processed through large-scale calculation. The more bits of image data are stored per pixel, the more the amount of computation is, and the greater the delay that is generated.
Although the resource efficiency can be simply used, and the traditional digital video and image processing mostly depends on the powerful computing power of CPU, DSP, GPU, FPGA and the like. This approach merely enhances efficiency by using more powerful hardware acceleration resources and does not substantially reduce the amount of computation in the data processing process. With the more widespread and deeper application scenes of image data analysis, the design of data analysis algorithms is more and more complex, which puts higher requirements on the operation speed of data analysis, and the instantaneity of analysis data is faced with huge pressure, so that the computational power of huge calculation amount on hardware is not small load.
In the face of the processing requirement of large-scale image data real-time performance, how to rapidly and accurately analyze and process target data in unit time is of great importance to the practical application of image big data analysis and processing. Therefore, for limited hardware conditions, how to reduce the calculation amount in the image data processing process on the basis of ensuring the accuracy requirement has a great significance on real-time data processing.
Disclosure of Invention
In order to solve the problems, the invention provides a high-precision quantification acceleration method for ultrahigh-resolution image analysis, which can essentially improve the calculated amount by quantifying huge calculation pressure in the ultrahigh-resolution image analysis process, and adopts the following technical scheme:
the invention provides a high-precision quantization acceleration method for ultrahigh-resolution image analysis, which is used for improving the calculation process of a large amount of intermediate data in the ultrahigh-resolution image analysis process through quantization and is characterized by comprising the following steps of: step S1, obtaining the accuracy alpha required by the final result, and obtaining the accuracy beta required to be reserved by intermediate calculation; step S2, obtaining an intermediate variable X, quantizing the intermediate variable X according to the accuracy beta required to be reserved in the intermediate calculation through the current calculated system gamma to obtain an offset variable X1(ii) a Step S3, based on the offset variable X1Performing calculation operation to obtain a calculation result Y1And reserving a calculation process queue Q in the calculation process; step S4, calculating the result Y1The accuracy beta required to be reserved for the intermediate calculation and the calculation process queue Q obtain a final result Y of the final system, wherein the step S2 includes the following sub-steps: step S2-1, calculating an offset digit n according to the currently calculated system gamma and the accuracy beta needing to be reserved by the intermediate calculation; step S2-2, the intermediate variable X is shifted to the left by n bits to obtain the shift variable X according to the intermediate variable X and the shift bit number n1。
The ultrahigh-resolution image analysis-oriented high-precision quantitative acceleration method provided by the invention can also have the technical characteristics that in the step S1, the specific calculation mode of the precision beta required to be reserved by intermediate calculation is as follows:
in the formula, δ is a preset precision conversion parameter.
The ultrahigh-resolution image analysis-oriented high-precision quantitative acceleration method provided by the invention can also have the technical characteristic that the set value of the precision conversion parameter delta is 10.
The ultrahigh resolution image analysis-oriented high-precision quantitative acceleration method provided by the invention can also have the technical characteristics that the method for calculating the offset digit n in the step S2-1 comprises the following steps:
the number n of offset bits is the smallest positive integer that satisfies formula (2).
The ultrahigh resolution image analysis-oriented high-precision quantitative acceleration method provided by the invention can also have the technical characteristics that the offset variable X is calculated in the step S2-21The method comprises the following steps:
X′＝X*γn (3)
calculating the data X 'after the offset of the intermediate variable X according to the formula (3), simultaneously recording the bit of the last translation, and directly discarding the data X' after the bit of the last translation to obtain the offset variable X1。
The ultrahigh-resolution image analysis-oriented high-precision quantification acceleration method provided by the invention can also have the technical characteristics that in the step S2, the system gamma at least comprises binary, octal, decimal and hexadecimal systems.
The ultrahigh resolution image analysis-oriented high-precision quantization acceleration method provided by the invention can also have the technical characteristic that when the calculation operation is addition or subtraction, the addition and subtraction are normally carried out in step S3 to obtain the calculation result Y1And records the queue Q1 of the addition and subtraction calculation process thereof, the calculation result Y is directly written according to the queue Q1 of the addition and subtraction calculation process in step S41Shifting n bits to the right translates to the final result Y of the final scale.
The ultrahigh resolution image analysis-oriented high-precision quantization acceleration method provided by the invention can also have the technical characteristics that when the calculation operation is multiplication, the multiplication operation is carried out in step S3 to obtain the calculation result Y1At the same time, record its multiplication process queue Q2, in step S4 according to the number k of elements in the multiplication process queue Q2, obtain the multiplication number, and will calculate the result Y1Shifting n x k bits to the right converts to the final result Y of the final scale.
The ultrahigh resolution image analysis-oriented high-precision quantization acceleration method provided by the invention can also have the technical characteristics that when the calculation operation is division, the division operation is carried out in step S3 to obtain the calculation result Y1At the same time, the queue Q3 of the division calculation process is recorded, and the calculation result Y is used in step S41And directly assigning the value to be the final result Y of the final system.
Action and Effect of the invention
According to the ultrahigh resolution image analysis-oriented high-precision quantitative acceleration method, floating point-like type intermediate variables which need to be subjected to complex calculation originally are converted into integer-like type variables in the ultrahigh resolution image analysis according to the input precision of an analyst, so that the calculated amount in the calculation process is directly reduced, and the calculation speed is accelerated. According to the invention, the intermediate data is converted based on the requirement of precision requirement, so that the time for calculation and analysis can be substantially reduced by reducing the storage bit number of the data and utilizing the quantized data for calculation, and the loss of precision caused by reducing the data bit number can be avoided, so that the precision of the result is not influenced. In addition, the method has universality and can be migrated and applied to the large-data-volume analysis and calculation of various ultrahigh-resolution images, so that the calculation amount of various analysis and calculation is reduced on the premise of ensuring the precision, and some applications requiring high real-time requirements in real scenes are better served.
Drawings
FIG. 1 is a flow chart of a high precision quantization acceleration method in an embodiment of the present invention;
FIG. 2 is a diagram of a data expansion and displacement process of an embodiment of the present invention; and
fig. 3 is a diagram of a post-quantization data simulation calculation process and a data result conversion process according to an embodiment of the present invention.
Detailed Description
In order to make the technical means, the creation features, the achievement purposes and the effects of the present invention easy to understand, the following describes the high-precision quantitative acceleration method for ultrahigh resolution image analysis according to the present invention in detail with reference to the embodiments and the accompanying drawings.
< example >
The ultrahigh-resolution image analysis-oriented high-precision quantitative acceleration method is implemented on a software system Windows server 2012, an OpenCV library compiled by CUDA is used in a calculation process based on an ultrahigh-resolution video image, and an acceleration calculation unit used is an NVIDIA 1080Ti GPU.
In the embodiment, an ultrahigh resolution video image is used as a processing object, a single frame of the ultrahigh resolution video image is stored with 8 bits, the resolution is 5120 × 5120, and the frame rate is 20 frames/s. The final accuracy is required to be within 0.5 pel and the calculation process is exemplified by using convolution calculations.
For ultrahigh resolution images, given data is a well-calculated integer, but normalization is required before processing, and intermediate variables are obtained in the calculation process. Assuming that the values of the first 3 × 3 pixels of the obtained image and the value of the partial intermediate variable X are (7.09375, 1.125, 1.125; 6.25, 5.00390625, 3.125; 3.1125, 11.4327, 1.125), the intermediate variable X is stored in binary by using CUDA based on Opencv in the calculation.
Fig. 1 is a flowchart of a high-precision quantization acceleration method according to an embodiment of the present invention.
As shown in fig. 1, the flow of the high-precision quantization acceleration method for full-face ultra-high resolution image analysis can be divided into four steps:
step S1, obtaining the accuracy α required by the calculation result, and obtaining the accuracy β required to be retained by the intermediate calculation, where the specific calculation method is as follows:
in this embodiment, it is assumed that the required final accuracy is within 0.5 pixel, so the accuracy α required by the calculation result is 0.1, while in this embodiment, the accuracy conversion parameter δ is set to 10, so the accuracy β required to be retained in the middle is calculated to be 0.01 according to the formula (1).
In this embodiment, the precision translation parameter δ is set to 10, but not limited to it, and may be set according to the task itself and the application, and if the precision is a more important indicator of the application result, the precision translation parameter δ may be increased by a proper amount.
In this embodiment, the accuracy α and the accuracy conversion parameter δ may be input by an analyst through an input terminal (e.g., an input device such as a keyboard) of the software system in advance.
Step S2, obtaining an intermediate variable X, and quantizing the intermediate variable X according to the accuracy beta required to be reserved in the intermediate calculation through the current calculated system gamma to obtain an offset variable X1, wherein the step S2-2 specifically comprises the step S2-1.
Step S2-1, obtaining the offset digit n according to the current calculated system γ and the intermediate calculation precision β that needs to be preserved, and the calculation method is as follows:
the number n of offset bits is the smallest positive integer that satisfies formula (2).
In this embodiment, since the intermediate variable X is stored in binary, the binary γ is equal to 2, and the offset number n is equal to 7 by calculation using formula (2).
The number of offset bits n is required to be the smallest positive integer satisfying the formula (2) on the premise that the formula (2) is satisfied. The reason why it needs to be the smallest positive integer is that: if not the smallest positive integer, extra bits are left blank, and the presence of these bits increases the amount of computation and does not significantly improve the accuracy of the computation. Since the intermediate accuracy β to be preserved is determined strictly from the result accuracy and the accuracy transformation parameter δ, it is not necessary to move more bits in the next step, so n is the smallest positive integer satisfying the above formula (2).
Step S2-2, according to the intermediate variable X and the offset digit n, the intermediate variable X is offset to the left by n digits to obtain an offset variable X1The calculation method comprises the following steps:
X′＝X*γn (3)
calculating the data X 'after the offset of the intermediate variable X according to the formula (3), simultaneously recording the bit of the last translation, and directly discarding the data X' after the bit of the last translation to obtain the offset variable X1。
In this embodiment, the intermediate variable X is stored with 16 bits, so the CUDA-based opencv is used to expand the data from 16 bits to 32 bits first. The reason for this is that: the 16-bit storage mode is extended to 32 bits by using the opencv function because the original 16-bit storage mode cannot well store the data after the shift by shifting the number of bits.
FIG. 2 is a diagram of a data expansion and displacement process of an embodiment of the present invention.
As shown in fig. 2, the present embodiment takes the intermediate variable X as 5.00390625 as an example: after the data of the intermediate variable X is expanded into 32 bits, the number is stored in binary form to 101.00000001 (as shown in (a) in fig. 2); subsequently, according to the offset digit n calculated in step S2-2, the intermediate variable X is shifted by 7 digits to obtain a result after quantization of the shift: 1010000000.1 (shown as (b) in FIG. 2); further, the remaining bits from the left to right (n + 1) th bit of the data itself are changed to 0 to reduce the amount of calculation by quantization and obtain the offset variable X after displacement quantization1The quantization process finally results in integers, i.e. the final fraction is directly omitted on the basis of accuracy, thus obtaining the final displacement quantization result (offset variable X)1) Comprises the following steps: 1010000000 (shown in fig. 2 (c)).
Step S3, based on the offset variable X1Performing calculation operation to obtain a calculation result Y1And the calculation process queue Q is reserved in the calculation process.
In the present embodiment, since convolution calculation is used, which is a process of performing weighted summation on data on an image by convolution kernel, the calculation operation in step S3 is only addition. Setting the convolution kernel to 3 × 3 specification and the parameters of the convolution kernel to (1, 0, 0; 0, 1, 0; 0, 0, 1), the data quantization process is not performed on the parameters of the convolution kernel in this embodiment because the parameters of the convolution kernelIt is inherently a small integer that does not require quantization. Then, performing convolution operation on each quantized intermediate variable to obtain: 11100001100+1010000000+10010000, resulting in a result 11010011100 (shown in FIG. 3). At this time, the sequence and number of times of performing various operations are recorded in the calculation process queue Q, and thus the calculation result Y is passed1And retaining the calculation process queue Q in the calculation process to obtain the final result Y.
Step S4, calculating the result Y1The accuracy beta needing to be reserved in the intermediate calculation and the final result Y of the final system obtained by the calculation process queue Q.
Since the process queue Q also contains only additions, the calculation result Y can be shifted based on the process queue Q1And converting into a final result Y. The number of bits of the displacement is still the offset number of bits n calculated in step S2, so Y is added1Shifting n bits to the right can result in the final result satisfying the accuracy α. In this embodiment, n is 7 and only addition is included in the calculation. Therefore, shifting the result Y1-11010010110 to the right by 7 bits yields the final result Y-1101.001011.
In the examples, it was found that the convolution result with (1, 0, 0; 0, 1, 0; 0, 0, 0, 1) data X (7.09375, 1.125, 1.125; 6.25, 5.00390625, 3.125; 3.1125, 11.4327, 1.125) with the convolution kernel parameters (1, 0, 0; 0, 0, 1) without quantization (i.e. not calculated by the method of the present invention using the usual method) was 7.09375+5.00390625+1.125 — 13.22265625. And the result, after it has been quantified in its calculation (i.e. by the method of the invention) and converted into a decimal variable, is 13.171875. The result of the conversion of the final result Y into decimal is: 13.171875. whereas in applications the accuracy requirement set is 0.5. It can be seen that the error of the two results is 0.05078125, which completely meets the precision requirement of the application.
In terms of computational efficiency, when a computer performs four arithmetic operations of addition, subtraction, multiplication and division, there are many differences between the operation of the integer data and the operation of the floating-point data by the operating system bottom instruction set. These differences result in the floating-point operation having a larger throughput than the integer arithmetic, and the floating-point operation having a larger delay than the integer arithmetic, both of which generally result in a longer operation time for the floating-point number than the integer arithmetic. In addition, in floating-point number operation, the design of hardware logic is more complex than that of integer numerical values, and indirectly, the operation speed is reduced. On the other hand, when data is stored in the memory, a single floating point occupies more bits than an integer value, so that the speed of reading the memory data is remarkably slow. Therefore, based on the embodiment, the intermediate variables (such as floating point numbers) similar to the floating point type are converted into the variables (such as integer numbers) similar to the integer type under the condition of ensuring the required precision, so that the operation efficiency is greatly improved, and the speed and precision calculation is perfectly balanced.
The method is applied to the real-time analysis and processing process of the ultrahigh resolution image, the GPU is used as a hardware resource and is applied to the real-time calibration algorithm of the ultrahigh resolution image, the position deviation in the calibration algorithm facing the 5 Kx 5K ultrahigh resolution image is within 0.5 pixel, the angle deviation is within the accuracy of 0.5 radian, the analysis and processing efficiency reaches 20fps, and the requirement of real-time processing of 20 frames of ultrahigh resolution images is met.
Examples effects and effects
According to the ultrahigh-resolution image analysis-oriented high-precision quantitative acceleration method provided by the embodiment, floating-point-like type intermediate variables which need to be subjected to complex calculation originally are generated in the ultrahigh-resolution image analysis and are converted into integer-like type variables according to the input precision of an analyst, and the analysis calculation is performed according to the integer-like type variables, so that the calculation amount in the calculation process is directly reduced, and the calculation speed is accelerated. According to the invention, the intermediate data is converted based on the requirement of precision requirement, so that the time for calculation and analysis can be substantially reduced by reducing the storage bit number of the data and utilizing the quantized data for calculation, and the loss of precision caused by reducing the data bit number can be avoided, so that the precision of the result is not influenced. In addition, the method has universality and can be migrated and applied to the large-data-volume analysis and calculation of various ultrahigh-resolution images, so that the calculation amount of various analysis and calculation is reduced on the premise of ensuring the precision, and some applications requiring high real-time requirements in real scenes are better served.
The above-described embodiments are merely illustrative of specific embodiments of the present invention, and the present invention is not limited to the description of the above-described embodiments.
In the embodiment, the calculation operation in step S3 only employs an addition operation, but the calculation operation of the present invention may include, but is not limited to, addition, subtraction, multiplication, and division, an exponential operation, and the like, and if there is a corresponding operation in the queue, the result is converted accordingly.
For example, when the calculation operation in step S3 is multiplication, the calculation result Y may be obtained by performing multiplication in step S31Meanwhile, recording the queue Q2 of the multiplication process; further, in step S4, the number of multiplications is obtained according to the number k of elements in the queue Q2 during the multiplication process, and the calculation result Y is used1Shifting n x k bits to the right converts to the final result Y of the final scale.
For another example, when the calculation operation in step S3 is division, the calculation result Y may be obtained by performing division in step S31Meanwhile, recording a division calculation process queue Q3; further, the calculation result Y is used in step S41And directly assigning the value to be the final result Y of the final system.
In an embodiment, the intermediate variable X is derived from image video data of ultra high definition video analysis. In other embodiments, the intermediate variable X may also be intermediate data used in large data volumes of various super-resolution images, such as encoded data, video raw data, and the like.
In the embodiment, the method disclosed by the invention is combined with an NVIDIA GPU supporting CUDA development to accelerate the calculation. In other embodiments, the method of the present invention may be combined with other methods to accelerate the calculation of large-scale data volume, and the hardware acceleration resource may be an FPGA with built-in image algorithm, a neural network acceleration unit, or the like.
In the embodiment, the image analysis algorithm is realized by using CUDA programming through an OpenCV library compiled by CUDA, and the acceleration is performed by adopting the method. In other embodiments, the calculation process of the method of the present invention may also use different programming languages to convert and calculate the data, supporting but not limited to C/C + +, C #, Java, python, etc.
Claims (7)
1. A high-precision quantification acceleration method for ultrahigh-resolution image analysis is used for quantifying a large number of intermediate variables in the ultrahigh-resolution image analysis process so as to improve the calculated amount in the calculation process, and is characterized by comprising the following steps of:
step S1, obtaining the accuracy alpha required by the final result, and obtaining the accuracy beta required to be reserved by intermediate calculation;
step S2, obtaining an intermediate variable X, quantizing the intermediate variable X according to the accuracy beta required to be reserved in the intermediate calculation through the current calculated system gamma to obtain an offset variable X1；
Step S3, based on the offset variable X1Performing calculation operation to obtain a calculation result Y1, and reserving a calculation process queue Q in the calculation process;
step S4, calculating the result Y1The accuracy beta of the intermediate calculation needing to be reserved and the final result Y of the final system obtained by the calculation process queue Q,
wherein the step S2 includes the following substeps:
step S2-1, calculating an offset digit n according to the currently calculated system γ and the accuracy β required to be retained by the intermediate calculation, wherein the offset digit n is calculated by the following method:
the offset digit n is the smallest positive integer satisfying formula (2);
step S2-2, calculating the offset variable X based on the intermediate variable X and the offset digit n1The offset variable X1The calculation method of (2) is as follows:
X′＝X*γn (3)
shifting the intermediate variable X to the left by n bits according to formula (3) to obtain data X ', recording the last shifted bit, and directly discarding the data X' after the last shifted bit to obtain the shift variable X1。
2. The ultrahigh resolution image analysis-oriented high-precision quantitative acceleration method according to claim 1, characterized in that:
in step S1, the specific calculation manner of the accuracy β that needs to be retained in the intermediate calculation is as follows:
in the formula, δ is a preset precision conversion parameter.
3. The ultrahigh resolution image analysis-oriented high-precision quantitative acceleration method according to claim 2, characterized in that:
wherein the set value of the precision conversion parameter delta is 10.
4. The ultrahigh resolution image analysis-oriented high-precision quantitative acceleration method according to claim 1, characterized in that:
wherein, in the step S2, the γ includes at least binary, octal, decimal, and hexadecimal.
5. The ultrahigh resolution image analysis-oriented high-precision quantitative acceleration method according to claim 1, characterized in that:
wherein, when the calculation operation is addition or subtraction,
in step S3, addition and subtraction are normally performed to obtain a calculation result Y1And recording the addition and subtraction thereofThe queue Q1 of the legal calculation process,
the calculation result Y is directly processed according to the queue Q1 of the addition and subtraction calculation process in step S41Shifting n bits to the right translates to the final result Y of the final scale.
6. The ultrahigh resolution image analysis-oriented high-precision quantitative acceleration method according to claim 1, characterized in that:
wherein, when the calculation operation is a multiplication,
in step S3, the multiplication is performed to obtain a calculation result Y1Meanwhile, recording the queue Q2 of the multiplication process,
in step S4, the number of multiplications is obtained according to the number k of elements in the queue Q2 of the multiplication process, and the calculation result Y is used1Shifting n x k bits to the right converts to the final result Y of the final scale.
7. The ultrahigh resolution image analysis-oriented high-precision quantitative acceleration method according to claim 1, characterized in that:
wherein, when the calculation operation is a division operation,
in step S3, division is performed to obtain a calculation result Y1Meanwhile, recording the queue Q3 of the division calculation process,
the calculation result Y is processed in step S41And directly assigning the value to be the final result Y of the final system.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910509308.1A CN110213582B (en) | 2019-06-13 | 2019-06-13 | High-precision quantitative acceleration method for ultrahigh-resolution image analysis |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910509308.1A CN110213582B (en) | 2019-06-13 | 2019-06-13 | High-precision quantitative acceleration method for ultrahigh-resolution image analysis |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110213582A CN110213582A (en) | 2019-09-06 |
| CN110213582B true CN110213582B (en) | 2021-05-11 |
Family
ID=67792446
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910509308.1A Active CN110213582B (en) | 2019-06-13 | 2019-06-13 | High-precision quantitative acceleration method for ultrahigh-resolution image analysis |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110213582B (en) |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1570843A (en) * | 2003-07-15 | 2005-01-26 | 中国科学院微电子中心 | Circuit realization structure for concurrent low power dissipation software computing element |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10362307B2 (en) * | 2016-09-09 | 2019-07-23 | Hanwha Techwin Co., Ltd. | Quantization parameter determination method and image capture apparatus |
-
2019
- 2019-06-13 CN CN201910509308.1A patent/CN110213582B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1570843A (en) * | 2003-07-15 | 2005-01-26 | 中国科学院微电子中心 | Circuit realization structure for concurrent low power dissipation software computing element |
Non-Patent Citations (1)
| Title |
|---|
| 《浮点运算的FPGA实现》;陈冬等;《DSP开发与应用》;20071231;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110213582A (en) | 2019-09-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2019184823A1 (en) | Convolutional neural network model-based image processing method and device | |
| US11249721B2 (en) | Multiplication circuit, system on chip, and electronic device | |
| CN112001914A (en) | Depth image completion method and device | |
| CN111986278B (en) | Image encoding device, probability model generating device, and image compression system | |
| Barannik et al. | Video data compression methods in the decision support systems | |
| US11018692B2 (en) | Floating point data set compression | |
| CN110728350A (en) | Quantification for machine learning models | |
| CN108932715B (en) | Deep learning-based coronary angiography image segmentation optimization method | |
| CN105208394B (en) | A kind of real-time digital image compression prediction technique and system | |
| CN114615507B (en) | Image coding method, decoding method and related device | |
| CN106683041B (en) | Quantum image miscut method based on NEQR expression | |
| CN110213582B (en) | High-precision quantitative acceleration method for ultrahigh-resolution image analysis | |
| CN110363713B (en) | Hyperspectral image noise reduction method based on recursive sample scaling and bilinear factorization | |
| CN116095183A (en) | Data compression method and related equipment | |
| WO2021143634A1 (en) | Arithmetic coder, method for implementing arithmetic coding, and image coding method | |
| JP2022546774A (en) | Interpolation filtering method and device, computer program and electronic device for intra prediction | |
| Huang et al. | Czip: A Fast Lossless Compression Algorithm for Climate Data | |
| US11263517B1 (en) | Flexible weight expansion | |
| Paul et al. | Image compression scheme based on histogram equalization and convolution neural network | |
| Kawser et al. | Embedding with pixel repetition for reversible data hiding | |
| Li et al. | Design and FPGA implementation of JPEG-LS image decompression | |
| US11409356B1 (en) | Using data correlation to reduce the power consumption of signal processing systems without affecting the precision of computation | |
| CN116828184B (en) | Video encoding method, video decoding method, video encoding device, video decoding device, computer equipment and storage medium | |
| US20230161558A1 (en) | Method and apparatus with quantization scheme implementation of artificial neural network | |
| KR100207650B1 (en) | Motion estimation apparatus of image compression system |
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 |