CN113762498B - Method for quantizing RoiAlign operator - Google Patents
Method for quantizing RoiAlign operator Download PDFInfo
- Publication number
- CN113762498B CN113762498B CN202010497788.7A CN202010497788A CN113762498B CN 113762498 B CN113762498 B CN 113762498B CN 202010497788 A CN202010497788 A CN 202010497788A CN 113762498 B CN113762498 B CN 113762498B
- Authority
- CN
- China
- Prior art keywords
- assigning
- weights
- binsize
- data
- channel
- 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
- 238000000034 method Methods 0.000 title claims abstract description 31
- 230000008569 process Effects 0.000 claims abstract description 8
- 238000013139 quantization Methods 0.000 claims description 8
- 238000001514 detection method Methods 0.000 claims description 5
- 238000005070 sampling Methods 0.000 claims description 5
- 238000012545 processing Methods 0.000 abstract description 3
- 238000006243 chemical reaction Methods 0.000 abstract description 2
- 238000013527 convolutional neural network Methods 0.000 description 5
- 230000006870 function Effects 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000001133 acceleration Effects 0.000 description 1
- 230000002776 aggregation Effects 0.000 description 1
- 238000004220 aggregation Methods 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000011002 quantification Methods 0.000 description 1
- 230000002787 reinforcement Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/04—Inference or reasoning models
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Image Processing (AREA)
Abstract
The invention provides a method for quantifying a ROIAlign operator, which comprises the following steps: s1, inputting Featuremap and quantizing data to obtain low-bit data; s2, calculating coordinates and weights according to the input ROIs, quantizing the weights, obtaining position indexes and corresponding weight values of the final output feature map, and performing RoiAlign operation on quantized data, wherein when the weights are calculated, the weights are fixed when the weights are calculated because the coordinates of the ROIs are floating points and the corresponding weights are floating points; s3, obtaining indexes and weights according to the steps, calculating a final result, and obtaining the output of the poolheight×poolwidth×channel for each unit through multiple AvgPooling operations. The ROIAlign directly processes the low-bit data without conversion to full-precision processing.
Description
Technical Field
The invention relates to the technical field of neural network acceleration, in particular to a method for quantifying a RoiAlign operator.
Background
In recent years, with rapid development of technology, a large data age has come. Deep learning takes a Deep Neural Network (DNN) as a model, and has quite remarkable results in many key fields of artificial intelligence, such as image recognition, reinforcement learning, semantic analysis and the like. The Convolutional Neural Network (CNN) is used as a typical DNN structure, can effectively extract hidden layer characteristics of images, accurately classifies the images, and is widely applied to the fields of image recognition and detection in recent years.
In particular, the target detection network ROIAlign operator: ROIAlign is a region feature aggregation approach proposed in the paper Mask-RCNN (author Kaiming He, georgia Gkioxari, piotter dolla r, ross Girshick, see https:// arxiv. Org/abs/1703.06870) to generate a fixed size featuremap based on a candidate box region pro-pos map.
However, in the prior art, floating point operation is adopted for the operator, and for the quantized model, the input quantized data needs to be converted into floating point numbers and then operated by encountering the ROIAlign operator, so that the operation efficiency and the bandwidth requirement of the whole quantized model are reduced.
Furthermore, the common terminology in the prior art is as follows:
convolutional neural network (Convolutional Neural Networks, CNN): is a type of feedforward neural network that includes convolution calculations and has a depth structure.
Quantification: quantization refers to the process of approximating a continuous value (or a large number of possible discrete values) of a signal to a finite number (or fewer) discrete values.
Low bits: the data is quantized to 8bit,4bit or 2bit wide data.
Disclosure of Invention
In order to solve the technical problems, the application provides a method for quantifying the ROIAlign operator, which aims to overcome the defects in the prior art, and provides a method for accelerating the model reasoning efficiency and reducing the bandwidth requirements, and solves the problem that the existing low-bit model needs to convert input into floating point numbers in the reasoning process.
The method of the invention can carry out quantization processing on the ROIAlign operator, namely, the data which is input into quantization is directly operated without being converted into floating point numbers and then corresponding operation is carried out. The ROIAlign directly processes the low-bit data without conversion to full-precision processing.
Specifically, the invention provides a method for quantifying the ROIAlign operator, the method comprising the steps of:
s1, inputting Featuremap and quantizing data to obtain low-bit data;
s2, calculating coordinates and weights according to the input ROIs, quantizing the weights, obtaining position indexes and corresponding weight values of final output featuremap, and performing RoiAlign operation on quantized data, wherein when the weights are calculated, the weights are fixed points because the coordinates of the ROIs are floating points and the corresponding weights are floating points, and when the weights are calculated, the weights are fixed points;
s3, obtaining indexes and weights according to the steps, calculating a final result, and obtaining the output of the poolheight×poolwidth×channel for each unit through multiple AvgPooling operations.
The step S1, data quantization: quantizing the data to be quantized according to a formula shown in a formula (1) to obtain low-bit data,
formula (1)
Description of variables: w (W) f Is an array, W q Max for quantized data w
Full precision data W f Middle maximum value, min w Full precision data W f B is the quantized bit width.
The calculating coordinates and weights and quantifying weights in step S2 further includes:
s2.1, setting up a structural body point, which comprises four members of xMin, yMin, rWidth and rHeight, wherein xMin is the minimum value of parameter x, yMin is the minimum value of parameter y, rWidth is the width of parameter r, and rHeight is the height of parameter r; here, the structural body Point represents a target frame for target detection, xMin, yMin, rWidth, and rwight represent the upper left corner coordinate and the length and width of the target frame on the feature map, respectively;
s2.2, rounding the powerheight, powerwidth, binSize, downsamples, fixedWidth, width and height; wherein,
PoolHeight represents the length of the feature map after the Roi is fixed;
PoolWidth represents the width of the feature map after the Roi is fixed;
binSize represents the number of sampling points per region;
down sample indicates that the feature is obtained by sampling N times from an original image, and N is a positive integer;
fixedWidth represents the bit width of the fractional part quantization of the coordinates;
s2.3, structure list: roi
Assigning getNum (roi) to roiNum;
reassigning 1 left shift fixedWidth to fixedScale;
s2.4, sequentially assigning values from 0 to roiNum to the tag to do the following operations:
assigning the roller (tag) to xMin, xMax, yMin, yMax;
assigning rHeight/poolight to the roiBinH;
assigning rWidth/poolWidth to the roiBinW;
wherein, from 0 to poolight are assigned to ph in sequence:
wherein, the method is used for assigning the pw in sequence from 0 to poolWidth:
wherein, the following operations are performed for assigning values to bh from 0 to binSize in sequence:
assigning yMin+ph, binSize+ (bh+0.5) binSize/roiBinH to y;
wherein, the following operations are performed for assigning to bw in sequence from 0 to binSize:
assigning xmin+pw + (bw+0.5) binSize/roiBinW to x;
assigning int (x) to xLow and int (y) to yLow;
assigning xLow+1 to xHigh, and yLow+1 to yHigh;
assigning (y-yLow) fixedScale to ly, and (x-xLow) fixedScale to lx;
assigning fixedScale-ly to hy and fixedScale-lx to hx;
assigning hy to w1, hy to lx to w2, ly to w3, ly to w4;
assigning yLow with+xLow to pos1 and yLow with+xhigh to pos2;
assigning vHigh with+xLow to pos3 and yHigh with+xHigh to pos4;
assigning tag+ (ph_poolheight+pw) channel to index;
calRoiAlign (index, pos1, pos2, pos3, pos4, w1, w2, w3, w 4) was calculated.
The calculating the final output according to the position index and the weight in the step S3 further includes: s3.1, featureMap: (featureMap is a multi-dimensional data with dimensions { height, width, channel })
S3.2, outlfeatureMap: (outfeatureMap, dimension { roiNum, poohight, poolWidth channel });
s3.3, fixedWidth rounding;
s3.4, calculating function calRoiAlign (index, pos1, pos2, pos3, pos4, w1, w2, w3, w 4), wherein the function is realized as a bilinear interpolation process, and pos1, pos2, pos3, pos4 are positions w1, w2, w3, w4 of surrounding pixels of the pixel to be calculated, and the weights of the surrounding pixels are calculated;
s3.5, assigning a featureMap [ pos1 x channel ] to dataPos1;
assigning a featureMap [ pos2 x channel ] to dataPos2;
assigning a featureMap [ pos3 channel ] to dataPos3;
assigning a featureMap [ pos4 x channel ] to dataPos4;
s3.6, sequentially assigning values to the tag from 0 to channel to do the following operations:
will beAssigning a value to tmpValue;
dataPos1++，dataPos2++，dataPos3++，dataPos4++
rightShift←fixedWidth+fixedWidth+binSize*binSize/2
assigning fixedwidth+fixedwidth+binsize to lightshift;
tmpvue right shift lightshift is reassigned to the outfeatureMap [ index+tag ].
Thus, the present application has the advantages that:
(1) For the RoiAlign operator, the input is quantized data, the operation is carried out without being converted into full-precision data, and the RoiAlign operation can be directly carried out on the quantized data;
(2) The reasoning process and speed of the low-bit model are optimized, and the requirements on bandwidth and memory are reduced.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this application, illustrate and together with the description serve to explain the invention.
FIG. 1 is a schematic diagram of the ROIALign operator flow in the prior art.
FIG. 2 is a schematic diagram of the quantized ROIAlign operator flow of the present invention.
FIG. 3 is a schematic flow chart of the method of the present invention.
FIG. 4 is a flow chart of a coding method for calculating coordinates and weights and quantifying the weights in the method of the present invention.
FIG. 5 is a flow chart of a coding method for calculating a final output according to a position index and a weight in the method of the present invention.
Detailed Description
In order that the technical content and advantages of the present invention may be more clearly understood, a further detailed description of the present invention will now be made with reference to the accompanying drawings.
As shown in fig. 3, a method of quantifying the ROIAlign operator according to the present invention includes the steps of:
s1, inputting Featuremap and quantizing data to obtain low-bit data;
s2, calculating coordinates and weights according to the input ROIs, quantizing the weights, obtaining position indexes and corresponding weight values of final output featuremap, and performing RoiAlign operation on quantized data, wherein when the weights are calculated, the weights are fixed points because the coordinates of the ROIs are floating points and the corresponding weights are floating points, and when the weights are calculated, the weights are fixed points;
s3, obtaining indexes and weights according to the steps, calculating a final result, and obtaining the output of the poolheight×poolwidth×channel for each unit through multiple AvgPooling operations.
Specifically, the present invention can also be interpreted as follows:
in the prior art, the implementation of the ROIAlign operator is mainly divided into 2 steps, 1, and a position index and a corresponding weight value of a final output feature map are obtained according to an input ROI. 2. The final result is calculated from the index and weight obtained in the first step and the output of roiNum x poolighth x pointchannel is obtained for each unit by AvgPooling operation. The flow chart is shown in fig. 1.
The implementation of the quantized ROIAlign operator is also divided into 2 steps, 1, the position index of the final output feature map and the corresponding weight value are obtained according to the input ROI, however, when the weight is calculated, the coordinate of the ROI is floating point, and the corresponding weight is floating point number, so that the position of the ROI is fixed when the weight is calculated. 2. The final result is calculated from the index and weight obtained in the first step and the output of poohight x poolWidth x Channel is obtained for each cell by the AvgPooling operation. The flow chart is shown in fig. 2.
Wherein, the method of calculating coordinates and weights and quantifying the weights is as shown in fig. 4:
01: setting up a structural body point, which comprises four members of xMin, yMin, rWidth and rHeight, wherein xMin is a minimum value of a parameter x, yMin is a minimum value of a parameter y, rWidth is a width of a parameter r, and rHeight is a height of the parameter r; here, the structure body Point represents a target frame for target detection, so xMin, yMin, rWidth, and rwight represent the upper left corner coordinates and length and width of the target frame on the feature map, respectively;
02: rounding the powerheight, powerwidth, binSize, downsampled, fixedWidth, width, height; wherein,
PoolHeight represents the length of the feature map after the Roi is fixed;
PoolWidth represents the width of the feature map after the Roi is fixed;
binSize represents the number of sampling points per region;
down sample indicates how many times the feature is sampled from the original image;
fixedWidth represents the bit width of the fractional part quantization of the coordinates;
03: structure list: roi
04: assigning getNum (roi) to roiNum;
05, reassigning 1 left shift fixedWidth to fixedScale;
06: for assigning values to tags in order from 0 to roiNum:
07: assigning the roller (tag) to xMin, xMax, yMin, yMax;
08: assigning rHeight/poolight to the roiBinH;
09: assigning rWidth/poolWidth to the roiBinW;
10: wherein, the following operations are performed for assigning values to ph in sequence from 0 to poolight:
11: wherein, the method is used for assigning the pw in sequence from 0 to poolWidth:
12: wherein, the following operations are performed for assigning values to bh from 0 to binSize in sequence:
13: assigning yMin+ph, binSize+ (bh+0.5) binSize/roiBinH to y;
and 14, wherein, the following operations are performed for assigning the bw to the binSize in sequence from 0:
15: assigning xmin+pw + (bw+0.5) binSize/roiBinW to x;
assigning int (x) to xLow and int (y) to yLow;
17: assigning xLow+1 to xHigh, and yLow+1 to yHigh;
18: assigning (y-yLow) fixedScale to ly, and (x-xLow) fixedScale to lx;
19: assigning fixedScale-ly to hy and fixedScale-lx to hx;
20: assigning hy to w1, hy to lx to w2, ly to w3, ly to w4;
21: assigning yLow with+xLow to pos1 and yLow with+xhigh to pos2;
22: assigning yHigh with+xLow to pos3 and yHigh with+xhigh to pos4;
23: assigning tag+ (ph_poolheight+pw) channel to index;
24: calculating calRoiAlign (index, pos1, pos2, pos3, pos4, w1, w2, w3, w 4), wherein the function is realized as a bilinear interpolation process, pos1, pos2, pos3, pos4 is the position w1, w2, w3, w4 of the surrounding pixel points of the pixel point to be calculated, and the weight of the surrounding pixel points is calculated;
wherein, the final output method is calculated according to the position index and the weight, and the coding is as shown in fig. 5:
01: featureMap: (featureMap is a piece of multidimensional data with dimensions { height, width, channel });
02 outfeaturemap: (outfeatureMap, dimension { roiNum, poohight, poolWidth channel });
02, fixedwidth rounding;
calculating function calRoiAlign (index, pos1, pos2, pos3, pos4, w1, w2, w3, w 4);
04, assigning a featureMap [ pos1 channel ] to dataPos1;
05, assigning the featureMap [ pos2 channel ] to dataPos2;
06, assigning the featureMap [ pos3 channel ] to dataPos3;
assigning a featureMap [ pos4 channel ] to dataPos4;
08, the following operations are performed for sequentially assigning values to the tag from 0 to channel:
09 will beAssigning a value to tmpValue;
10:dataPos1++，dataPos2++，dataPos3++，dataPos4++
11:rightShift←fixedWidth+fixedWidth+binSize*binSize/2
assigning fixedwidth+fixedwidth+binsize to lightshift by binSize/2; 13: tmpvue right shift lightshift is reassigned to the outfeatureMap [ index+tag ].
The above description is only of the preferred embodiments of the present invention and is not intended to limit the present invention, and various modifications and variations can be made to the embodiments of the present invention by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (2)
1. A method of quantifying the ROIAlign operator, the method comprising the steps of:
s1, inputting Featuremap and quantizing data to obtain low-bit data;
s2, calculating coordinates and weights according to the input ROIs, quantizing the weights, obtaining position indexes and corresponding weight values of the final output feature map, and performing RoiAlign operation on quantized data, wherein when the weights are calculated, the weights are fixed when the weights are calculated because the coordinates of the ROIs are floating points and the corresponding weights are floating points;
the calculating coordinates and weights and quantifying weights further comprises:
s2.1, setting up a structural body point, which comprises four members of xMin, yMin, rWidth and rHeight, wherein xMin is the minimum value of parameter x, yMin is the minimum value of parameter y, rWidth is the width of parameter r, and rHeight is the height of parameter r; here, the structural body Point represents a target frame for target detection, xMin, yMin, rWidth, and rwight represent the upper left corner coordinate and the length and width of the target frame on the feature map, respectively;
s2.2, rounding the powerheight, powerwidth, binSize, downsamples, fixedWidth, width and height; wherein,
PoolHeight represents the length of the feature map after the Roi is fixed;
PoolWidth represents the width of the feature map after the Roi is fixed;
binSize represents the number of sampling points per region;
down sample indicates that the feature is obtained by sampling N times from an original image, and N is a positive integer; fixedWidth represents the bit width of the fractional part quantization of the coordinates;
s2.3, structure list: roi
Assigning getNum (roi) to roiNum;
reassigning 1 left shift fixedWidth to fixedScale;
s2.4, sequentially assigning values from 0 to roiNum to the tag to do the following operations:
assigning the roller (tag) to xMin, xMax, yMin, yMax;
assigning rHeight/poolight to the roiBinH;
assigning rWidth/poolWidth to the roiBinW;
wherein, from 0 to poolight are assigned to ph in sequence:
wherein, the method is used for assigning the pw in sequence from 0 to poolWidth:
wherein, the following operations are performed for assigning values to bh from 0 to binSize in sequence:
assigning yMin+ph, binSize+ (bh+0.5) binSize/roiBinH to y;
wherein, the following operations are performed for assigning to bw in sequence from 0 to binSize:
assigning xmin+pw + (bw+0.5) binSize/roiBinW to x;
assigning int (x) to xLow and int (y) to yLow;
assigning xLow+1 to xHigh, and yLow+1 to yHigh;
assigning (y-yLow) fixedScale to ly, and (x-xLow) fixedScale to lx;
assigning fixedScale-ly to hy and fixedScale-lx to hx;
assigning hy to w1, hy to lx to w2, ly to w3, ly to w4;
assigning yLow with+xLow to pos1 and yLow with+xhigh to pos2;
assigning yHigh with+xLow to pos3 and yHigh with+xhigh to pos4;
assigning tag+ (ph_poolheight+pw) channel to index;
calculating calRoiAlign (index, pos1, pos2, pos3, pos4, w1, w2, w3, w 4);
s3, obtaining indexes and weights according to the steps, calculating a final result, and obtaining the output of a poolHeight multiplied by poolWidth multiplied by Channel for each unit by multiple AvgPooling operations;
the calculating the final output according to the position index and the weight in the step S3 further includes:
s3.1, featureMap: (featureMap is a multi-dimensional data with dimensions { height, width, channel })
S3.2, outlfeatureMap: (outfeatureMap, dimension { roiNum, poohight, poolWidth channel });
s3.3, fixedWidth rounding;
s3.4, calculating function calRoiAlign (index, pos1, pos2, pos3, pos4, w1, w2, w3, w 4), wherein the function is realized as a bilinear interpolation process, and pos1, pos2, pos3, pos4 are positions w1, w2, w3, w4 of surrounding pixels of the pixel to be calculated, and the weights of the surrounding pixels are calculated;
s3.5, assigning a featureMap [ pos1 x channel ] to dataPos1;
assigning a featureMap [ pos2 x channel ] to dataPos2;
assigning a featureMap [ pos3 channel ] to dataPos3;
assigning a featureMap [ pos4 x channel ] to dataPos4;
s3.6, sequentially assigning values to the tag from 0 to channel to do the following operations:
will beAssigning a value to tmpValue;
dataPos1++，dataPos2++，dataPos3++，dataPos4++；
rightShift←fixedWidth+fixedWidth+binSize*binSize/2，
assigning fixedwidth+fixedwidth+binsize to lightshift;
tmpvue right shift lightshift is reassigned to the outfeatureMap [ index+tag ].
2. The method of quantifying the ROIAlign operator according to claim 1, wherein the step S1 is data quantization: quantizing the data to be quantized according to a formula shown in a formula (1) to obtain low-bit data,
formula (1)
Description of variables: w (W) f Is an array, W q Max for quantized data w Full precision data W f Middle maximum value, min w Full precision data W f B is the quantized bit width, S W Is a scaling factor that quantizes floating point data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010497788.7A CN113762498B (en) | 2020-06-04 | 2020-06-04 | Method for quantizing RoiAlign operator |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010497788.7A CN113762498B (en) | 2020-06-04 | 2020-06-04 | Method for quantizing RoiAlign operator |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113762498A CN113762498A (en) | 2021-12-07 |
| CN113762498B true CN113762498B (en) | 2024-01-23 |
Family
ID=78783422
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010497788.7A Active CN113762498B (en) | 2020-06-04 | 2020-06-04 | Method for quantizing RoiAlign operator |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113762498B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109005409A (en) * | 2018-07-27 | 2018-12-14 | 浙江工业大学 | A kind of intelligent video coding method based on object detecting and tracking |
| CN109685064A (en) * | 2018-10-29 | 2019-04-26 | 黑龙江科技大学 | A kind of modified RoIAlign provincial characteristics aggregation algorithms |
| CN109685208A (en) * | 2018-12-24 | 2019-04-26 | 合肥君正科技有限公司 | A kind of method and device accelerated for the dilute combization of neural network processor data |
| CN110008915A (en) * | 2019-04-11 | 2019-07-12 | 电子科技大学 | The system and method for dense human body attitude estimation is carried out based on mask-RCNN |
| CN110852416A (en) * | 2019-09-30 | 2020-02-28 | 成都恒创新星科技有限公司 | CNN accelerated computing method and system based on low-precision floating-point data expression form |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11006926B2 (en) * | 2018-02-27 | 2021-05-18 | Siemens Medical Solutions Usa, Inc. | Region of interest placement for quantitative ultrasound imaging |
-
2020
- 2020-06-04 CN CN202010497788.7A patent/CN113762498B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109005409A (en) * | 2018-07-27 | 2018-12-14 | 浙江工业大学 | A kind of intelligent video coding method based on object detecting and tracking |
| CN109685064A (en) * | 2018-10-29 | 2019-04-26 | 黑龙江科技大学 | A kind of modified RoIAlign provincial characteristics aggregation algorithms |
| CN109685208A (en) * | 2018-12-24 | 2019-04-26 | 合肥君正科技有限公司 | A kind of method and device accelerated for the dilute combization of neural network processor data |
| CN110008915A (en) * | 2019-04-11 | 2019-07-12 | 电子科技大学 | The system and method for dense human body attitude estimation is carried out based on mask-RCNN |
| CN110852416A (en) * | 2019-09-30 | 2020-02-28 | 成都恒创新星科技有限公司 | CNN accelerated computing method and system based on low-precision floating-point data expression form |
Non-Patent Citations (1)
| Title |
|---|
| 基于低码率传输的红外视频编码方法研究;陈青华;红外;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113762498A (en) | 2021-12-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110135580B (en) | Convolution network full integer quantization method and application method thereof | |
| CN115049936A (en) | High-resolution remote sensing image-oriented boundary enhancement type semantic segmentation method | |
| CN112183742B (en) | Neural network hybrid quantization method based on progressive quantization and Hessian information | |
| CN111147862B (en) | End-to-end image compression method based on target coding | |
| CN108446766A (en) | A kind of method of quick trained storehouse own coding deep neural network | |
| CN113888547A (en) | Non-supervision domain self-adaptive remote sensing road semantic segmentation method based on GAN network | |
| CN115311555A (en) | Remote sensing image building extraction model generalization method based on batch style mixing | |
| CN116958827A (en) | Deep learning-based abandoned land area extraction method | |
| CN113762498B (en) | Method for quantizing RoiAlign operator | |
| CN114332479A (en) | Training method of target detection model and related device | |
| CN114419060A (en) | Skin mirror image segmentation method and system | |
| CN110782396B (en) | Light-weight image super-resolution reconstruction network and reconstruction method | |
| CN115170807B (en) | Image segmentation and model training method, device, equipment and medium | |
| CN112487992A (en) | Stream model-based face emotion image generation method and device | |
| CN112150362A (en) | Picture preprocessing solution | |
| CN116309213A (en) | High-real-time multi-source image fusion method based on generation countermeasure network | |
| CN114463449A (en) | Hyperspectral image compression method based on edge guide | |
| CN114677281A (en) | FIB-SEM super-resolution algorithm based on generation countermeasure network | |
| CN113223038A (en) | Discrete cosine transform-based mask representation instance segmentation method | |
| CN118350984A (en) | Image style migration method based on multi-level cascade structure | |
| CN111967580B (en) | Low-bit neural network training method and system based on feature migration | |
| CN117528085B (en) | Video compression coding method based on intelligent feature clustering | |
| CN113762496B (en) | Method for reducing low-bit convolutional neural network reasoning operation complexity | |
| CN113762452B (en) | Method for quantizing PRELU activation function | |
| CN113159217B (en) | Attention mechanism target detection method based on event camera |
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 |