CN113762498A - Method for quantizing RoiAlign operator - Google Patents
Method for quantizing RoiAlign operator Download PDFInfo
- Publication number
- CN113762498A CN113762498A CN202010497788.7A CN202010497788A CN113762498A CN 113762498 A CN113762498 A CN 113762498A CN 202010497788 A CN202010497788 A CN 202010497788A CN 113762498 A CN113762498 A CN 113762498A
- Authority
- CN
- China
- Prior art keywords
- assigning
- binsize
- width
- weight
- data
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 34
- 230000008569 process Effects 0.000 claims abstract description 10
- 238000013139 quantization Methods 0.000 claims description 9
- 238000010586 diagram Methods 0.000 claims description 6
- 238000001514 detection method Methods 0.000 claims description 5
- 238000005070 sampling Methods 0.000 claims description 5
- 238000004364 calculation method Methods 0.000 claims description 4
- 238000012545 processing Methods 0.000 abstract description 3
- 238000013527 convolutional neural network Methods 0.000 description 4
- 230000006870 function Effects 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 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
- 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
Images
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 quantizing a ROIAlign operator, which comprises the following steps: s1, inputting Featuremap and quantizing the data to obtain low-bit data; s2, calculating coordinates and weight according to the input ROI and quantizing the weight, acquiring a position index of a final output feature map and a corresponding weight value, and performing RoiAlign operation on quantized data, wherein when the weight is calculated, the coordinates of the ROI are floating points, and the corresponding weight is also a floating point number, so that the weight can be fixed in point when the weight is calculated; s3, calculating the final result according to the indexes and weights obtained in the above steps and obtaining the output of poolHeight x poolWidth x Channel for each unit of Avgpoling operation. In the method, ROIAlign directly processes low-bit data without converting the low-bit data into full-precision data for processing.
Description
Technical Field
The invention relates to the technical field of neural network acceleration, in particular to a method for quantizing a RoiAlign operator.
Background
In recent years, with the rapid development of science and technology, a big data age has come. Deep learning takes a Deep Neural Network (DNN) as a model, and achieves remarkable results in key fields of many human intelligence, such as image recognition, reinforcement learning, semantic analysis and the like. The Convolutional Neural Network (CNN) is a typical DNN structure, can effectively extract hidden layer features of an image and accurately classify the image, and is widely applied to the field of image identification and detection in recent years.
In particular, the object detection network ROIAlign operator: ROIAlign is a regional feature aggregation proposed in the paper Mask-RCNN (authors Kaiming He, Georgia Gkioxari, Piotr Doll r, Ross Girshick, see https:// axiv. org/abs/1703.06870), when generating a fixed-size featuremap from a candidate box region pro-posal map.
However, in the prior art, floating point operation is adopted for the operator, and for the quantized model, the ROIAlign operator needs to convert input quantized data into floating point numbers first and then operate, so that the operating efficiency of the whole quantization model is reduced and the requirement on bandwidth is met.
Furthermore, the common terminology in the prior art is as follows:
convolutional Neural Networks (CNN): is a type of feedforward neural network that contains convolution calculations and has a depth structure.
And (3) 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) of discrete values.
Low bit rate: and quantizing the data into data with bit width of 8bit, 4bit or 2 bit.
Disclosure of Invention
In order to solve the technical problems, the application provides a method for quantizing the roiign operator, aims to overcome the defects in the prior art, provides a method for accelerating the model reasoning efficiency and reducing the bandwidth requirement, and solves the problem that the existing low-bit model needs to convert input into floating point numbers in the reasoning process.
The method can carry out quantization processing on the ROIAlign operator, namely, the input quantized data is directly operated without being converted into floating point numbers and then corresponding operation is carried out. In the method, ROIAlign directly processes low-bit data without converting the low-bit data into full-precision data for processing.
Specifically, the invention provides a method for quantizing ROIAlign operator, which comprises the following steps:
s1, inputting Featuremap and quantizing the data to obtain low-bit data;
s2, calculating coordinates and weight according to the input ROI, quantizing the weight, obtaining a position index of a final output featuremap and a corresponding weight value, and performing RoiAlign operation on quantized data, wherein when the weight is calculated, the coordinates of the ROI are floating points, and the corresponding weight is also a floating point number, so that the weight can be fixed in point when the weight is calculated;
s3, calculating the final result according to the indexes and weights obtained in the above steps and obtaining the output of poolHeight x poolWidth x Channel for each unit of Avgpoling operation.
In step S1, data quantization: quantizing the data to be quantized according to a formula shown in formula (1) to obtain low-bit data,
Description of variables: wfFor full-precision data being an array, WqFor quantized data, maxw
Full precision data WfMedian maximum value, minwFull precision data WfAnd b is the bit width after quantization.
The step S2 of calculating coordinates and weights and quantizing weights further includes:
s2.1, setting a structure point, wherein the structure point comprises four members of xMin, yMin, rWidth and rHeight, wherein xMin is the minimum value of a parameter x, yMin is the minimum value of a parameter y, rWidth is the width of a parameter r, and rHeight is the height of the parameter r; the structural body Point represents a target frame of target detection, and xMin, yMin, rWidth and rwight respectively represent the coordinate and the length and the width of the upper left corner of the target frame on the characteristic diagram;
s2.2, rounding poolHeight, poolWidth, binSize, downSample, fixedWidth, width and height; wherein,
PoolHeight indicates the length of the feature map after fixation of Roi;
PoolWidth represents the width of the feature map after fixing the Roi;
binSize represents the number of sampling points per region;
the downSample indicates that the feature of the layer is obtained by sampling N times from the original image, and N is a positive integer;
fixedWidth represents the bit width of the coordinate decimal part quantization;
s2.3, structure list: roi
Assigning getNum (roi) to roiNum;
shifting 1 left fixed width and assigning to fixed Scale;
s2.4, assigning values to tag from 0 to roiNum in sequence, and performing the following operations:
assigning the roi (tag) to xMin, xMax, yMin, yMax;
assigning rHeight/poolHeight to roiBinH;
assigning rWidth/poolWidth to roiBinW;
wherein, the values are sequentially assigned to ph from 0 to poolHeight, and the following operations are carried out:
wherein, the following operations are performed for assigning values to pw from 0 to poolWidth in sequence:
and assigning values to bh from 0 to binSize in sequence, wherein the following operations are performed:
y is assigned y min + ph binSize + (bh +0.5) binSize/roiBinH;
and assigning values to bw from 0 to binSize in sequence, wherein the following operations are performed:
assigning x to xMin + pw binSize + (bw +0.5) binSize/roiBinW;
assign int (x) to xLow, int (y) to yLow;
assigning xLow +1 to xHigh, and assigning yLow +1 to yHigh;
assigning (y-yLow) fixedScale to ly and (x-xLow) fixedScale to lx;
assigning fixedScale-ly to hy, and assigning fixedScale-lx to hx;
assigning hy x hx to w1, hy x lx to w2, ly x hx to w3, and ly x lx to w 4;
assigning yLow _ width + xLow to pos1 and yLow _ width + xHigh to pos 2;
assigning vHigh × width + xLow to pos3 and yHigh × width + xHigh to pos 4;
assigning tag + (ph _ poolHeight + pw) channel to index;
calRoiAlign (index, pos1, pos2, pos3, pos4, w1, w2, w3, w4) was calculated.
The step S3 of calculating the final output according to the position index and the weight further includes: s3.1, featureMap: (featureMap is a multidimensional data with dimensions { height, width, channel })
S3.2, outfeatureMap: (outfeatureMap, dimension { roiNum, poolhight, poolWidth channel });
s3.3, rounding the fixedWidth;
s3.4, calculating function calRoiAlign (index, pos1, pos2, pos3, pos4, w1, w2, w3 and w4), wherein the function is realized as a bilinear interpolation process, pos1, pos2, pos3 and pos4 are positions w1, w2, w3 and w4 of surrounding pixels of the pixel to be solved are weights of the surrounding pixels during calculation;
s3.5, assigning featureMap [ pos1 × channel ] to dataPos 1;
featureMap [ pos2 channel ] was assigned to dataPos 2;
featureMap [ pos3 channel ] was assigned to dataPos 3;
featureMap [ pos4 channel ] was assigned to dataPos 4;
s3.6, assigning values to tag from 0 to channel in sequence, and performing the following operations:
dataPos1++,dataPos2++,dataPos3++,dataPos4++
rightShift←fixedWidth+fixedWidth+binSize*binSize/2
assigning fixedWidth + fixedWidth + binSize × binSize/2 to rightShift;
and the tmpValue is shifted right by rightShift and is assigned to outfeatureMap [ index + tag ].
Thus, the present application has the advantages that:
(1) for the RoiAlign operator, quantized data are input, the RoiAlign operator does not need to be converted into full-precision data and then operated, and RoiAlign operation can be directly performed on the quantized data;
(2) the inference 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 embodiment(s) of the invention and together with the description serve to explain the principles of the invention.
Fig. 1 is a schematic diagram of the prior art roiallign operator process.
FIG. 2 is a schematic diagram of the process of quantifying the ROIAlign operator according to the present invention.
FIG. 3 is a schematic flow diagram of the method of the present invention.
Fig. 4 is a flow chart illustrating a coding method for calculating coordinates and weights and quantizing the weights in the method of the present invention.
FIG. 5 is a flow diagram of a method for implementing the coding method for computing final outputs based on position indices and weights in the method of the present invention.
Detailed Description
In order that the technical contents and advantages of the present invention can be more clearly understood, the present invention will now be described in further detail with reference to the accompanying drawings.
As shown in fig. 3, a method for quantizing roiign operator according to the present invention includes the following steps:
s1, inputting Featuremap and quantizing the data to obtain low-bit data;
s2, calculating coordinates and weight according to the input ROI, quantizing the weight, obtaining a position index of a final output featuremap and a corresponding weight value, and performing RoiAlign operation on quantized data, wherein when the weight is calculated, the coordinates of the ROI are floating points, and the corresponding weight is also a floating point number, so that the weight can be fixed in point when the weight is calculated;
s3, calculating the final result according to the indexes and weights obtained in the above steps and obtaining the output of poolHeight x poolWidth x Channel for each unit of Avgpoling operation.
Specifically, the present invention can also be interpreted as follows:
in the prior art, the ROIAlign operator is mainly realized in 2 steps, 1, a position index and a corresponding weight value of a final output feature map are obtained according to an input ROI. 2. And calculating a final result according to the index and the weight obtained in the first step, and obtaining the output of roiNum x poolHeight x poolWidth x Channel for each unit of the multiple AvgPoling operations. The flow chart is shown in fig. 1.
The implementation of the quantized ROIAlign operator according to the present invention is also divided into 2 steps, 1, obtaining the position index and the corresponding weight value of the final output feature map according to the input ROI, however, when calculating the weight, because the coordinates of the ROI are floating point, the corresponding weight is also floating point number, so that the ROI is fixed point when calculating the weight. 2. And calculating a final result according to the index and the weight obtained in the first step, and obtaining the output of poolHeight x poolWidth x Channel for each unit of the multiple AvgPoling operations. The flow chart is shown in fig. 2.
The method for calculating coordinates and weights and quantizing the weights is coded as shown in fig. 4:
01: setting a structure point, which comprises four members of xMin, yMin, rWidth and rHeight, wherein xMin is the minimum value of a parameter x, yMin is the minimum value of a parameter y, rWidth is the width of a parameter r, and rHeight is the height of the parameter r; the structure Point here represents a target frame for target detection, so xMin, yMin, rWidth and rwight respectively represent the coordinates and length and width of the upper left corner of the target frame on the feature map;
02: rounding poolHeight, poolWidth, binSize, downSample, fixedWidth, width, height; wherein,
PoolHeight indicates the length of the feature map after fixation of Roi;
PoolWidth represents the width of the feature map after fixing the Roi;
binSize represents the number of sampling points per region;
the downSample represents how many times the feature of the layer is sampled from the original image;
fixedWidth represents the bit width of the coordinate decimal part quantization;
03: structure list: roi
04: assigning getNum (roi) to roiNum;
05, moving the 1 left fixed Width and assigning the value to fixed Scale;
06: assigning the tag with the following operations from 0 to roiNum in sequence:
07: assigning the roi (tag) to xMin, xMax, yMin, yMax;
08: assigning rHeight/poolHeight to roiBinH;
09: assigning rWidth/poolWidth to roiBinW;
10: wherein, the following operations are carried out for assigning values to ph from 0 to poolHeight in sequence:
11: wherein, the following operations are performed for assigning values to pw from 0 to poolWidth in sequence:
12: and assigning values to bh from 0 to binSize in sequence, wherein the following operations are performed:
13: y is assigned y min + ph binSize + (bh +0.5) binSize/roiBinH;
and 14, assigning values to bw from 0 to binSize in sequence, and performing the following operations:
15: assigning x to xMin + pw binSize + (bw +0.5) binSize/roiBinW;
assign int (x) to xLow, int (y) to yLow;
17: assigning xLow +1 to xHigh, and assigning yLow +1 to yHigh;
18: assigning (y-yLow) fixedScale to ly and (x-xLow) fixedScale to lx;
19: assigning fixedScale-ly to hy, and assigning fixedScale-lx to hx;
20: assigning hy x hx to w1, hy x lx to w2, ly x hx to w3, and ly x lx to w 4;
21: assigning yLow _ width + xLow to pos1 and yLow _ width + xHigh to pos 2;
22: assigning yHigh width + xLow to pos3 and yHigh width + xHigh to pos 4;
23: assigning tag + (ph _ poolHeight + pw) channel to index;
24: calculating calRoiAlign (index, pos1, pos2, pos3, pos4, w1, w2, w3 and w4), wherein the function is realized as a bilinear interpolation process, pos1, pos2, pos3 and pos4 are positions w1, w2, w3 and w4 of surrounding pixel points of the pixel point to be solved are weights of the surrounding pixel points during calculation;
in which, the method of calculating the final output according to the position index and the weight is coded as shown in fig. 5:
01: featureMap: (featureMap is a plurality of multidimensional data with dimensions { height, width, channel });
02 outfeatureMap: (outfeatureMap, dimension { roiNum, poolhight, poolWidth channel });
02, rounding the fixedWidth;
calculating function calRoiAlign (index, pos1, pos2, pos3, pos4, w1, w2, w3, w 4);
04, assigning featureMap [ pos1 channel ] to dataPos 1;
05 assigning featureMap [ pos2 channel ] to dataPos 2;
06 assigning featureMap [ pos3 channel ] to dataPos 3;
07 assigning featureMap [ pos4 channel ] to dataPos 4;
08, assigning the value to tag from 0 to channel in sequence, and performing the following operations:
10:dataPos1++,dataPos2++,dataPos3++,dataPos4++
11:rightShift←fixedWidth+fixedWidth+binSize*binSize/2
assigning fixedWidth + fixedWidth + binSize/2 to rightShift; 13: and the tmpValue is shifted right by rightShift and is assigned to outfeatureMap [ index + tag ].
The above description is only a preferred embodiment of the present invention, and is not intended to limit the present invention, and various modifications and changes may be made to the embodiment of the present invention by those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (4)
1. A method of quantizing ROIAlign operators, the method comprising the steps of:
s1, inputting Featuremap and quantizing the data to obtain low-bit data;
s2, calculating coordinates and weight according to the input ROI and quantizing the weight, acquiring a position index of a final output feature map and a corresponding weight value, and performing RoiAlign operation on quantized data, wherein when the weight is calculated, the coordinates of the ROI are floating points, and the corresponding weight is also a floating point number, so that the weight can be fixed in point when the weight is calculated;
s3, calculating the final result according to the indexes and weights obtained in the above steps and obtaining the output of poolHeight x poolWidth x Channel for each unit of Avgpoling operation.
2. The method of claim 1, wherein the step S1 is implemented by quantizing data: quantizing the data to be quantized according to a formula shown in formula (1) to obtain low-bit data,
Description of variables: wfFor full-precision data being an array, WqFor quantized data, maxwFull precision data WfMedian maximum value, minwFull precision data WfAnd b is the bit width after quantization.
3. The method of claim 1, wherein the step of calculating coordinates and weights and quantizing weights in step S2 further comprises:
s2.1, setting a structure point, wherein the structure point comprises four members of xMin, yMin, rWidth and rHeight, wherein xMin is the minimum value of a parameter x, yMin is the minimum value of a parameter y, rWidth is the width of a parameter r, and rHeight is the height of the parameter r; the structural body Point represents a target frame of target detection, and xMin, yMin, rWidth and rwight respectively represent the coordinate and the length and the width of the upper left corner of the target frame on the characteristic diagram;
s2.2, rounding poolHeight, poolWidth, binSize, downSample, fixedWidth, width and height; wherein,
PoolHeight indicates the length of the feature map after fixation of Roi;
PoolWidth represents the width of the feature map after fixing the Roi;
binSize represents the number of sampling points per region;
the downSample indicates that the feature of the layer is obtained by sampling N times from the original image, and N is a positive integer; fixedWidth represents the bit width of the coordinate decimal part quantization;
s2.3, structure list: roi
Assigning getNum (roi) to roiNum;
shifting 1 left fixed width and assigning to fixed Scale;
s2.4, assigning values to tga from 0 to roiNum in sequence to perform the following operations:
assigning the roi (tag) to xMin, xMax, yMin, yMax;
assigning rHeight/poolHeight to roiBinH;
assigning rWidth/poolWidth to roiBinW;
wherein, the values are sequentially assigned to ph from 0 to poolHeight, and the following operations are carried out:
wherein, the following operations are performed for assigning values to pw from 0 to poolWidth in sequence:
and assigning values to bh from 0 to binSize in sequence, wherein the following operations are performed:
y is assigned y min + ph binSize + (bh +0.5) binSize/roiBinH;
and assigning values to bw from 0 to binSize in sequence, wherein the following operations are performed:
assigning x to xMin + pw binSize + (bw +0.5) binSize/roiBinW;
assign int (x) to xLow, int (y) to yLow;
assigning xLow +1 to xHigh, and assigning yLow +1 to yHigh;
assigning (y-yLow) fixddScale to ly and (x-xLow) fixedScale to lx;
assigning fixedScale-ly to hy, and assigning fixedScale-lx to hx;
assigning hy x hx to w1, hy x lx to w2, ly x hx to w3, and ly x lx to w 4;
assigning yLow _ width + xLow to pos1 and yLow _ width + xHigh to pos 2;
assigning yHigh width + xLow to pos3 and yHigh width + xHigh to pos 4;
assigning tag (ph _ poolHeight + pw) channel to index;
calRoiAlign (index, pos1, pos2, pos3, pos4, w1, w2, w3, w4) was calculated.
4. The method of claim 1, wherein the step of calculating the final output according to the position index and the weight in step S3 further comprises:
s3.1, featureMaps (featureMaps are multidimensional data with dimensions { height, width, channel })
S3.2, outfeatureMap: (outfeatureMap, dimension { roiNum, poolhight, poolWidth channel });
s3.3, rounding the fixedWidth;
s3.4, calculating function calRoiAlign (index, pos1, pos2, pos3, pos4, w1, w2, w3 and w4), wherein the function is realized as a bilinear interpolation process, pos1, pos2, pos3 and pos4 are positions w1, w2, w3 and w4 of surrounding pixels of the pixel to be solved are weights of the surrounding pixels during calculation;
s3.5, assigning featureMap [ pos1 × channel ] to dataPos 1;
featureMap [ pos2 channel ] was assigned to dataPos 2;
featureMap [ pos3 channel ] was assigned to dataPos 3;
featureMap [ pos4 channel ] was assigned to dataPos 4;
s3.6, assigning values to tag from 0 to channel in sequence, and performing the following operations:
dataPos1++,dataPos2++,dataPos3++,dataPos4++rightShift←fixedWidth+fixedWidth+binSize*binSize/2
assigning fixedWidth + fixedWidth + binSize × binSize/2 to rightShift;
and the tmpValue is shifted right by rightShift and is assigned to outfeatureMap [ index + tag ].
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 true CN113762498A (en) | 2021-12-07 |
CN113762498B 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 (6)
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 |
US20190261949A1 (en) * | 2018-02-27 | 2019-08-29 | Siemens Medical Solutions Usa, Inc. | Region of interest placement for quantitative ultrasound imaging |
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 |
-
2020
- 2020-06-04 CN CN202010497788.7A patent/CN113762498B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190261949A1 (en) * | 2018-02-27 | 2019-08-29 | Siemens Medical Solutions Usa, Inc. | Region of interest placement for quantitative ultrasound imaging |
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 |
---|
陈青华: "基于低码率传输的红外视频编码方法研究", 红外 * |
Also Published As
Publication number | Publication date |
---|---|
CN113762498B (en) | 2024-01-23 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108446766A (en) | A kind of method of quick trained storehouse own coding deep neural network | |
CN113688822A (en) | Time sequence attention mechanism scene image identification method | |
CN111147862B (en) | End-to-end image compression method based on target coding | |
CN115759237A (en) | End-to-end deep neural network model compression and heterogeneous conversion system and method | |
CN114240955A (en) | Semi-supervised cross-domain self-adaptive image segmentation method | |
CN116958827A (en) | Deep learning-based abandoned land area extraction method | |
CN115913247A (en) | Deep lossless compression method and system for high-frequency power data | |
CN114332479A (en) | Training method of target detection model and related device | |
CN112487992B (en) | Stream model-based face emotion image generation method and device | |
CN113762498A (en) | Method for quantizing RoiAlign operator | |
CN110782396B (en) | Light-weight image super-resolution reconstruction network and reconstruction method | |
CN115170807B (en) | Image segmentation and model training method, device, equipment and medium | |
CN117197632A (en) | Transformer-based electron microscope pollen image target detection method | |
CN114299252B (en) | Method and device for reconstructing universal three-dimensional model, storage medium and electronic equipment | |
CN115836849A (en) | Arrhythmia classification method based on comparison test time adaptation | |
CN115063364A (en) | Intelligent traditional Chinese medicine tongue diagnosis method based on artificial intelligence and computer storage medium | |
CN113223038A (en) | Discrete cosine transform-based mask representation instance segmentation method | |
CN114463449A (en) | Hyperspectral image compression method based on edge guide | |
CN113988154A (en) | Unsupervised decoupling image generation method based on invariant information distillation | |
Ling et al. | A Study of Quantisation-aware Training on Time Series Transformer Models for Resource-constrained FPGAs | |
Suzuki et al. | ProgressiveNN: Achieving Computational Scalability without Network Alteration by MSB-first Accumulative Computation | |
CN113762496B (en) | Method for reducing low-bit convolutional neural network reasoning operation complexity | |
CN116934594A (en) | Ultrahigh resolution image calculation method based on transducer | |
CN113159217B (en) | Attention mechanism target detection method based on event camera | |
Zhou et al. | Multi-scales feature integration single shot multi-box detector on small object detection |
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 |