CN108960251A - A kind of images match description generates the hardware circuit implementation method of scale space - Google Patents
A kind of images match description generates the hardware circuit implementation method of scale space Download PDFInfo
- Publication number
- CN108960251A CN108960251A CN201810495451.5A CN201810495451A CN108960251A CN 108960251 A CN108960251 A CN 108960251A CN 201810495451 A CN201810495451 A CN 201810495451A CN 108960251 A CN108960251 A CN 108960251A
- Authority
- CN
- China
- Prior art keywords
- image
- gaussian
- scale
- generates
- scale space
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of images match to describe the hardware circuit implementation method that son generates scale space, includes the following steps: under concrete application occasion, loading need to do matched image；Gaussian smoothing is done to the image and generates first group of image；Next group of image is generated using the method for down-sampling combination gaussian filtering；Difference of Gaussian pyramid is generated according to obtained gaussian pyramid.The present invention is redesigned by the scale space generation method to SIFT image matching algorithm, the characteristics of to adapt to FPGA operation fixed point, parallel, assembly line, it accelerates the speed of scale space generation and saves hardware resource, to effectively optimize the performance of entire characteristic matching module.
Description
Technical field
The present invention relates to the graphical rule space generation methods based on SIFT algorithm family, and in particular to a kind of images match
Description generates the hardware circuit implementation method of scale space.
Background technique
Currently, graphical rule space, which generates, is broadly divided into two kinds: one kind is multi-scale Representation (multi-scale
Representation), using the mode for adding low-pass filtering based on down-sampling；Another kind of is that scale space indicates
(scale-space representation), using the mode of Laplce's filtering under identical sample rate.It is based on
The graphical rule space of SIFT algorithm family generates, and is that the combination of two kinds of above methods is taken its advantage, keeps away its disadvantage, will scheme
As scale space divides (octave) and layer (layer) in groups[1], so as to form the multiscale space of image.
Classical SIFT graphical rule space generation method is developed in Image Multiscale Space Theory, is had very
Strong theories integration also has similar structure with human visual system, research shows that detection obtains under SIFT scale space
Feature Descriptor comprehensive performance be optimal[2], including scale invariability (scale invariant), rotational invariance
(rotate invariant), illumination invariant (illumination invariant) and certain affine-invariant features
(affine invariant)。
However, classical SIFT graphical rule space is had the disadvantage that based on Software Create
1) computationally intensive.For the image of a 512*512, generates for 3 groups 6 layers of scale space, only filter just
6*3=18 group filter is needed, carries out convolution algorithm, it is assumed that average filter template size is 16*16, needs 512*512*15*
15=226Secondary multiplication, calculation amount are surprising；
2) arithmetic speed is low.Scale space based on filtering operation generates, and relates generally to summation operation and product calculation, and
The space complexity and input data of complexity (time complexity and the space complexity) and input data of both operations
How much proportional, images of a 512*512, be intel core i5,8G DDR, 64 win7 operation system in CPU
On the desktop computer of system, 3 groups 6 layers of scale space is established, runing time was also above 3 seconds.
Summary of the invention
Goal of the invention: the figure in the graphical rule space generation method based on SIFT algorithm family, for dimension calculation
As being at least also required to two groups.But scale space establishes huge calculating (based on the smooth of convolution), becomes obstruction and improves image
Matching speed, one of the principal element for promoting business throughput provide a kind of based on SIFT algorithm in order to overcome this problem
Images match description generates the hardware circuit implementation method of scale space, accelerates the speed that scale space is established, and reduces and calculates
Guarantee data precision while amount, efficiently use the characteristic of hardware configuration, including assembly line and computation capability, reduces algorithm
Hardware resource occupancy, effectively optimize the performance of entire characteristic matching module, enhance the portability of algorithm.
Technical solution: to achieve the above object, the present invention provides a kind of images match description and generates the hard of scale space
Part circuit implementing method, includes the following steps:
1) under concrete application occasion, loading need to do matched image；
2) Gaussian smoothing is done to the image and generates first group of image；
3) next group of image is generated using the method for down-sampling combination gaussian filtering；
4) difference of Gaussian pyramid (Difference of Gaussian, DOG) is generated according to obtained gaussian pyramid.
The detailed process of the step 1 are as follows:
1.1) the gray scale floating data of image is read by Ethernet into DDR；
1.2) image data in DDR is successively read according to the sequence of Row Column into Buffer.
The detailed process of the step 2 are as follows:
2.1) separability for utilizing Gaussian function, splits into two one-dimensional Gaussian filters for 2-d gaussian filters device
Convolution will need the one-dimensional filtering coefficients used to carry out fixed point processing, and wherein the size of Gaussian template width should be with
Scale where the image is directly proportional；
2.2) column filtering is carried out to image, it is using the symmetry of Gaussian template when column filter, the pixel of symmetric position is first
It adds up, the utilization of less multiplier is then multiplied to template；
2.3) capable filtering is carried out to row filter result, using the reuse of data when row filters, is filtered using transposition FIR
Device saves hardware resource with less output delay；
2.4) mode for smoothly obtaining next layer in change conventional method to upper one layer, meets semigroup knot using Gaussian function
The property of structure, all layers all smoothly obtain the image under this layer of scale to first layer, and make its synchronism output of result, to save place
Manage the time.
The detailed process of the step 3 are as follows:
While step 2 smoothly obtains first group of each tomographic image, to two sampling of input picture drop, sampled result is carried out
With the consistent processing of process described in step 2, the image of other more large scale groups is obtained.
The detailed process of the step 4 are as follows:
For step 3 obtain as a result, i.e. gaussian pyramid, while a upper tomographic image is subtracted into next tomographic image and obtains height
This difference pyramid, is equally parallel processing herein.
Mesoscale space of the present invention generating principle are as follows:
SIFT algorithm combines the respective advantage of scale space representation and pyramidal representation, innovatively proposes height
This pyramid is for constructing scale space.
Gaussian pyramid is made of the different size of gaussian filtering layer of multiple groups, and first group since original image with specific ruler
Spend σ0Gaussian convolution core filter to obtain first group of the first floor, then with Gaussian convolution core by filtering layer by layer obtain one group it is discrete
Scale space；Next group again by under the layer third from the bottom in upper one group two it is down-sampled obtain this group of first layer, and with same volume
Product stratum nucleare layer filter to obtain the scale space of this group, and so on obtain other groups.
The scale image of sequential sampling in order to obtain, enables scale be incremented by with k times.If every group of needs are sought in S layers of scale
Characteristic point is looked for, then needs the scale image of at least (S+3) layer.The empirical value obtained is tested according to David Lowe, taking S is 3, again
Due to requiring every group of layer third from the bottom is down-sampled to obtain next group of first layer, then the scale that layer third from the bottom is S layers should be
2 times of first layer scale, then k=21/S。
Gauss quintar keeps picture size and pixel quantity constant in same group, does according to Scale-space theory to image
Gaussian convolution simulates dimensional variation.When image is excessively fuzzy in this group, i.e., when having bulk redundancy information in image, then according to image
Pyramidal representation generates next group of the first floor after down-sampled, and so on finally obtain gaussian pyramid.Gaussian pyramid
It is exactly based on and combines the method for Scale-space theory and image pyramid, that is, ensure that the scale continuity of image, and big
Reduce calculation amount greatly.
Laplce's Gaussian function σ of dimension normalization2▽2The Local Extremum that G is responded in scale space is very stable
Characteristics of image.In order to reduce the calculation amount of the operator, difference Gaussian function DoG (Difference of is used in SIFT
Gaussian) carry out Laplce's Gaussian function of approximate dimension normalization.Difference gaussian pyramid is by the Gauss gold word that has constructed
Difference obtains tower two-by-two, and every group of S+2 layers of Gauss scale layer obtain S+1 layers of difference Gauss layer.
The utility model has the advantages that compared with prior art, the present invention having following advantage:
1, on the basis of previous method for producing software, by parallel computation and the pipeline design, by it is traditional based on
Floating number, the scale space serially executed generation are converted into fixed point, parallel generation method, and filtering method is turned by two-dimensional convolution
One-dimensional convolution is turned to, the speed of scale space generation is accelerated, the power consumption and hardware resource of floating-point processing are reduced, thus effectively
Ground optimizes the performance of entire characteristic matching module；
2, optimization method is generated using the SIFT scale space based on FPGA, because having carried out parallel and flowing water line computation,
And certain optimization has been carried out in process, each tomographic image can be with synchronism output, while greatly reducing calculation amount, Er Qie
Wherein most stable of a collection of characteristic point has been retained in the extreme point that former algorithm detects, is reduced the pressure of subsequent match, is mentioned
The high arithmetic speed of images match, and hardware resource occupancy is greatly reduced, it is with a wide range of applications.
Detailed description of the invention
Fig. 1 is overall flow schematic diagram of the invention；
Fig. 2 is that parallel filtering obtains the schematic diagram of gaussian pyramid；
Fig. 3 is gaussian filtering structural schematic diagram；
Fig. 4 is to obtain the pyramidal schematic diagram of difference of Gaussian by gaussian pyramid.
Specific embodiment
In the following with reference to the drawings and specific embodiments, the present invention is furture elucidated.
Embodiment 1:
As shown in Figure 1, the present invention provides a kind of hardware circuit implementation method of images match description generation scale space,
Include the following steps:
1) under concrete application occasion, loading need to do matched image, detailed process are as follows:
1.1) the gray scale floating data of image is read by Ethernet into DDR；
1.2) image data in DDR is successively read according to the sequence of Row Column into Buffer.
2) Gaussian smoothing is done to the image and generates first group of image, detailed process are as follows:
2.1) separability for utilizing Gaussian function, splits into two one-dimensional Gaussian filters for 2-d gaussian filters device
Convolution will need the one-dimensional filtering coefficients used to carry out fixed point processing, and wherein the size of Gaussian template width should be with
Scale where the image is directly proportional；
2.2) column filtering is carried out to image, it is using the symmetry of Gaussian template when column filter, the pixel of symmetric position is first
It adds up, the utilization of less multiplier is then multiplied to template；
2.3) capable filtering is carried out to row filter result, using the reuse of data when row filters, is filtered using transposition FIR
Device saves hardware resource with less output delay；
2.4) mode for smoothly obtaining next layer in change conventional method to upper one layer, meets semigroup knot using Gaussian function
The property of structure, all layers all smoothly obtain the image under this layer of scale to first layer, and make its synchronism output of result, to save place
Manage the time.
3) next group of image, detailed process are generated using the method for down-sampling combination gaussian filtering are as follows:
While step 2 smoothly obtains first group of each tomographic image, to two sampling of input picture drop, sampled result is carried out
With the consistent processing of process described in step 2, the image of other more large scale groups is obtained.
4) difference of Gaussian pyramid, detailed process are generated according to obtained gaussian pyramid are as follows:
For step 3 obtain as a result, i.e. gaussian pyramid, while a upper tomographic image is subtracted into next tomographic image and obtains height
This difference pyramid, is equally parallel processing herein.
According to above method step, it is assumed that input is the picture of a 512*512 size, one byte (8 of each pixel
A bit) it is indicated, the gaussian pyramid of 2 groups 6 layers of building, first layer graphical rule positioning 1.6.First by filter system
Number fixed point processing, scale is respectively 1.6,1.6*21/3,1.6*22/3, 3.2,3.2*21/3,3.2*22/3, 1024 times of expansion, four
House five finely tunes the sum of assurance coefficient after entering be 1024.Then Gaussian smoothing is carried out to 6*2=12 tomographic image respectively.Second group of image
It is to be filtered on the basis of first group of first layer image drop sampling, specific gaussian filtering structure is as shown in figure 3, filtering is
Rank of advanced units filtering, then carries out capable filtering, obtains gaussian pyramid as shown in Figure 2, the gaussian pyramid that filtering is obtained
Difference of Gaussian pyramid is obtained, it is specific as shown in Figure 4.
Scale space hardware circuit realizes that step is specific as follows:
Gaussian pyramid is suitble to use spatial parallelism method, while doing gaussian filtering to all images in one group, so not only
It can greatly accelerate the algorithm speed of service, moreover it is possible to while filter result is obtained, conveniently do the assembly line arrangement of algorithm.Due to same group
In, every layer of Gauss scale image carries out convolution operation simultaneously, and every group of Gaussian convolution core is enabled to be superimposed convolution since first layer scale
Remaining each layer, and separately design the coefficient of the convolution kernel of following each layer.
By the superposability of gaussian kernel function, from every group of first layer (relative scalar σ0) filtering obtain remaining each layer, count
Calculate the scale coefficient of the gaussian kernel function of other each layers:
By scale coefficient, the Filtering Template coefficient of each layer of gaussian kernel function finally can be obtained.Again by gaussian kernel function
Detachable property be conducive to substantially reduce calculation amount, each 2-d gaussian filters device can be enabled to split two one-dimensional filtering devices, respectively
For Gauss line filter and Gauss column filter, and their coefficients are identical.
It is much larger due to compare fixed-point calculation to the resource consumption of floating-point operation in logical device, it will be in algorithm
Floating-point operation do fixed point processing.Under the premise of taking into account resource consumption and algorithm performance, this method is to each raw Gaussian
Each template parameter of filter is multiplied by 1024, then rounds up, and obtains the gaussian filtering template of only integer.
Detachable dimensional Gaussian convolution kernel function is the identical row convolution kernel function of coefficients and column convolution kernel letter
Number.Two-dimensional gaussian filtering is done equal to accumulating once row filtering on the image and once arranging filtering to image, splits dimensional Gaussian
Calculation amount can be substantially reduced after core.
With G (x, y, σ1) for, do the fixed point of gaussian filtering, it is known that σ0=1.6, it can obtain:
In order to guarantee the adequacy of gaussian filtering, i.e. according to the energy for being included within the scope of 3 σ of radius of gaussian kernel function
Integral in the radius is the 99.7% of entire gaussian kernel function, therefore Filtering Template width should also follow the principle, to protect
Demonstrate,prove scale invariability.Then G (x, y, σ1) corresponding to template width should be 6 σ1≈ 7.358 rounds up and obtains template for odd number
Width is 9.By Gaussian filter design factor:
[0.0016 0.0164 0.0861 0.2333 0.3252 0.2333 0.0861 0.0164 0.0016]
1024 are multiplied by each coefficients:
[1.6266 16.6945 88.0879 238.9582 333.2656 238.9582 88.0879 16.6945
1.6266]
Each coefficient is rounded up again:
[2 17 88 239 333 239 88 17 2]
But this can not as final coefficients because the coefficient of the template and be 1025, greater than the amplification of script
Coefficient 1024.It is demonstrated experimentally that if coefficient between each layer gaussian filtering template and inconsistent, it will be each interlayer in difference Gauss
Error amplification causes characteristic point detection that can not detect the characteristic point of real stabilization and Scale invariant.
Therefore, which also needs to adjust.It is as small as possible for principle to be changed to template, if coefficients and being less than
1024, then it finds and is less than before rounding up but rounds up closest to coefficients X.5, if coefficients and if being greater than 1024
Conversely, to take into account the central symmetry principle of template simultaneously.Due to 1025-1024=1, it is contemplated that central symmetry and change to template
The smallest principle is moved, is changed to 333-1=2 for the 333 of template center, obtains final G (x, y, σ1) template:
[2 17 88 239 332 239 88 17 2]
The Filtering Template coefficient of other scales obtains in the same way.
The centre symmetry of gaussian filtering template is utilized when carrying out convolution.It is filtered using a width for the Gauss of w (w is odd number)
For wave template, convolution operation needs w-1 sub-addition and w times each time:
G (i)=G (w-1-i) is known by the centre symmetry of gaussian filtering template, then:
Finally, will successively be subtracted each other to obtain difference of Gaussian pyramid by each tomographic image of gaussian pyramid obtained by the above method.
Embodiment 2:
Scale invariability test, specific testing procedure are carried out to this method below are as follows:
A) the scale space generation step for optimizing source figure and matching figure respectively obtains belonging to respective graphical rule
Space, i.e., final difference of Gaussian pyramid；
B) two groups of difference of Gaussian pyramids are obtained into description of characteristic point with the method for classics SIFT respectively；
C) description is schemed into matching and carries out Euclidean distance matching with source figure description, method is: description, recently
Neighborhood distance/time nearest neighbor distance < 0.6 is determined as a matching pair, statistical match is to quantity and when completing consumed by matching
Between.
By under different platform and scale space generating mode to the scale invariability of the images match of two 640*512
It is tested, obtains Tables 1 and 2:
Scale invariability is tested under 1 FPGA of table
Note: FPGA clock frequency 100MHz
Scale invariability is tested under 2 CPU of table
Note: CPU model core i5, dominant frequency 3.2GHz
Through Tables 1 and 2 as can be seen that at hardware platform FPGA, the scale space generation method after optimization is maintained
SIFT possesses the characteristic of feature-rich point quantity, has very big promotion in processing speed, substantially meets the requirement of real-time.
Claims (5)
1. the hardware circuit implementation method that a kind of images match description generates scale space, it is characterised in that: including walking as follows
It is rapid:
1) under concrete application occasion, loading need to do matched image；
2) Gaussian smoothing is done to the image and generates first group of image；
3) next group of image is generated using the method for down-sampling combination gaussian filtering；
4) difference of Gaussian pyramid is generated according to obtained gaussian pyramid.
2. a kind of images match description according to claim 1 generates the hardware circuit implementation method of scale space,
It is characterized in that: the detailed process of the step 1 are as follows:
1.1) the gray scale floating data of image is read into memory into (DDR) by Ethernet
1.2) image data in DDR is successively read according to the sequence of Row Column into buffer (Buffer).
3. a kind of images match description according to claim 1 generates the hardware circuit implementation method of scale space,
It is characterized in that: the detailed process of the step 2 are as follows:
2.1) 2-d gaussian filters device, is split into the volume of two one-dimensional Gaussian filters by the separability for utilizing Gaussian function
The one-dimensional filtering coefficients for needing to use are carried out fixed point processing by product, and wherein the size of Gaussian template width should be with this
Scale where image is directly proportional；
2.2) column filtering is carried out to image first to have added the pixel of symmetric position using the symmetry of Gaussian template when column filter
Come, the utilization of less multiplier is then multiplied to template；
2.3) capable filtering is carried out to row filter result, using the reuse of data when row filters, using transposition FIR filter；
2.4) mode for smoothly obtaining next layer in change conventional method to upper one layer, meets Semigroup Structure using Gaussian function
Property, all layers all smoothly obtain the image under this layer of scale to first layer, and make its synchronism output of result.
4. a kind of images match description according to claim 3 generates the hardware circuit implementation method of scale space,
It is characterized in that: the detailed process of the step 3 are as follows:
While step 2 smoothly obtains first group of each tomographic image, to two sampling of input picture drop, sampled result is carried out and is walked
The rapid 2 consistent processing of process, obtains the image of other more large scale groups.
5. a kind of images match description according to claim 1 or 4 generates the hardware circuit implementation method of scale space,
It is characterized by: the detailed process of the step 4 are as follows:
For step 3 obtain as a result, i.e. gaussian pyramid, while a upper tomographic image is subtracted into next tomographic image and obtains Gaussian difference
Divide pyramid, is equally parallel processing herein.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810495451.5A CN108960251A (en) | 2018-05-22 | 2018-05-22 | A kind of images match description generates the hardware circuit implementation method of scale space |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810495451.5A CN108960251A (en) | 2018-05-22 | 2018-05-22 | A kind of images match description generates the hardware circuit implementation method of scale space |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108960251A true CN108960251A (en) | 2018-12-07 |
Family
ID=64499488
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810495451.5A Pending CN108960251A (en) | 2018-05-22 | 2018-05-22 | A kind of images match description generates the hardware circuit implementation method of scale space |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108960251A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108334882A (en) * | 2018-01-05 | 2018-07-27 | 佛山市顺德区中山大学研究院 | A kind of system and method for parallel generation SIFT description |
| CN109919825A (en) * | 2019-01-29 | 2019-06-21 | 北京航空航天大学 | A kind of ORB-SLAM hardware accelerator |
| CN111738920A (en) * | 2020-06-12 | 2020-10-02 | 山东大学 | FPGA (field programmable Gate array) framework for panoramic stitching acceleration and panoramic image stitching method |
| CN112132914A (en) * | 2019-06-25 | 2020-12-25 | 珠海市一微半导体有限公司 | Image scale space establishing method and image processing chip |
| CN112541507A (en) * | 2020-12-17 | 2021-03-23 | 中国海洋大学 | Multi-scale convolutional neural network feature extraction method, system, medium and application |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106504229A (en) * | 2016-09-30 | 2017-03-15 | 上海联影医疗科技有限公司 | The detection method of characteristic point in image |
| CN107392215A (en) * | 2017-08-02 | 2017-11-24 | 焦点科技股份有限公司 | A kind of multigraph detection method based on SIFT algorithms |
| WO2020181506A1 (en) * | 2019-03-12 | 2020-09-17 | 深圳市大疆创新科技有限公司 | Image processing method, apparatus and system |
-
2018
- 2018-05-22 CN CN201810495451.5A patent/CN108960251A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106504229A (en) * | 2016-09-30 | 2017-03-15 | 上海联影医疗科技有限公司 | The detection method of characteristic point in image |
| CN107392215A (en) * | 2017-08-02 | 2017-11-24 | 焦点科技股份有限公司 | A kind of multigraph detection method based on SIFT algorithms |
| WO2020181506A1 (en) * | 2019-03-12 | 2020-09-17 | 深圳市大疆创新科技有限公司 | Image processing method, apparatus and system |
Non-Patent Citations (1)
| Title |
|---|
| 薛顺瑞等: "《基于FPGA并行处理SIFT算法特征点检测》", 《电视技术》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108334882A (en) * | 2018-01-05 | 2018-07-27 | 佛山市顺德区中山大学研究院 | A kind of system and method for parallel generation SIFT description |
| CN109919825A (en) * | 2019-01-29 | 2019-06-21 | 北京航空航天大学 | A kind of ORB-SLAM hardware accelerator |
| CN112132914A (en) * | 2019-06-25 | 2020-12-25 | 珠海市一微半导体有限公司 | Image scale space establishing method and image processing chip |
| CN112132914B (en) * | 2019-06-25 | 2024-07-12 | 珠海一微半导体股份有限公司 | Image scale space establishment method and image processing chip |
| CN111738920A (en) * | 2020-06-12 | 2020-10-02 | 山东大学 | FPGA (field programmable Gate array) framework for panoramic stitching acceleration and panoramic image stitching method |
| CN112541507A (en) * | 2020-12-17 | 2021-03-23 | 中国海洋大学 | Multi-scale convolutional neural network feature extraction method, system, medium and application |
| CN112541507B (en) * | 2020-12-17 | 2023-04-18 | 中国海洋大学 | Multi-scale convolutional neural network feature extraction method, system, medium and application |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108960251A (en) | A kind of images match description generates the hardware circuit implementation method of scale space | |
| Cai et al. | Yolobile: Real-time object detection on mobile devices via compression-compilation co-design | |
| Pestana et al. | A full featured configurable accelerator for object detection with YOLO | |
| EP4036810A1 (en) | Neural network processing method and apparatus, computer device and storage medium | |
| Ghaffari et al. | FPGA-based convolutional neural network accelerator design using high level synthesize | |
| WO2019007406A1 (en) | Data processing apparatus and method | |
| CN110383300A (en) | A kind of computing device and method | |
| Chang et al. | An efficient implementation of 2D convolution in CNN | |
| CN110163358A (en) | A kind of computing device and method | |
| CN110163333B (en) | Parallel optimization method of convolutional neural network | |
| CN110543936B (en) | Multi-parallel acceleration method for CNN full-connection layer operation | |
| WO2022111002A1 (en) | Method and apparatus for training neural network, and computer readable storage medium | |
| CN113887710A (en) | Digital format selection in recurrent neural networks | |
| CN114757347A (en) | Method and system for realizing low bit quantization neural network accelerator | |
| Zhang et al. | Graph-PBN: Graph-based parallel branch network for efficient point cloud learning | |
| CN102970545A (en) | Static image compression method based on two-dimensional discrete wavelet transform algorithm | |
| Gao et al. | Multi-branch aware module with channel shuffle pixel-wise attention for lightweight image super-resolution | |
| CN115545175A (en) | Running bidirectional recurrent neural networks in hardware | |
| CN114782336A (en) | Method and device for predicting fiber bundle orientation distribution based on graph convolution neural network | |
| Jiang et al. | An energy-efficient and high-PSNR image denoising pre-processor based on optimized non-local means algorithm for real-time FHD video applications | |
| CN113887714A (en) | Method and system for running dynamic recurrent neural networks in hardware | |
| Gu et al. | A Swin Transformer based Framework for Shape Recognition | |
| Zeng et al. | A Hybrid-Pipelined Architecture for FPGA-based Binary Weight DenseNet with High Performance-Efficiency | |
| Cheng | Design and implementation of convolutional neural network accelerator based on fpga | |
| Li et al. | Fpga-based object detection acceleration architecture design |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181207 |