CN110659721A - Method and system for constructing target detection network - Google Patents
Method and system for constructing target detection network Download PDFInfo
- Publication number
- CN110659721A CN110659721A CN201910713035.2A CN201910713035A CN110659721A CN 110659721 A CN110659721 A CN 110659721A CN 201910713035 A CN201910713035 A CN 201910713035A CN 110659721 A CN110659721 A CN 110659721A
- Authority
- CN
- China
- Prior art keywords
- search
- network
- determining
- target detection
- 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.)
- Granted
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 67
- 238000000034 method Methods 0.000 title claims abstract description 40
- 238000011156 evaluation Methods 0.000 claims abstract description 34
- 230000006870 function Effects 0.000 claims abstract description 21
- 238000012549 training Methods 0.000 claims description 13
- 238000012360 testing method Methods 0.000 claims description 8
- 230000002787 reinforcement Effects 0.000 claims description 5
- 230000002068 genetic effect Effects 0.000 claims description 2
- 230000004044 response Effects 0.000 claims description 2
- 238000011478 gradient descent method Methods 0.000 claims 2
- 238000010586 diagram Methods 0.000 description 11
- 230000008901 benefit Effects 0.000 description 4
- 238000012545 processing Methods 0.000 description 4
- 230000008569 process Effects 0.000 description 3
- 238000006467 substitution reaction Methods 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000007689 inspection Methods 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 238000010845 search algorithm Methods 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- PCHJSUWPFVWCPO-UHFFFAOYSA-N gold Chemical compound [Au] PCHJSUWPFVWCPO-UHFFFAOYSA-N 0.000 description 1
- 239000010931 gold Substances 0.000 description 1
- 229910052737 gold Inorganic materials 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 239000002994 raw material Substances 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000012795 verification Methods 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
- G06N3/04—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
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- General Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Neurology (AREA)
- Data Exchanges In Wide-Area Networks (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention provides a method and a system for constructing a target detection network, wherein the method comprises the following steps: determining a network architecture search space of the target detection network, wherein the search space defines a network architecture search range and a detection submodule; determining a search strategy, wherein the search strategy defines a method for searching the search space; determining a model evaluation index, wherein the model evaluation index is an objective function of the search strategy. According to the invention, various search strategies are combined, and the network architecture is searched from coarse to fine, so that the search efficiency is improved, and a better local optimal solution is obtained; multiple model evaluation indexes are fused, and meanwhile, the speed, the precision and the memory are predicted by using small network learning, so that better evaluation indexes can be obtained, and the evaluation efficiency is improved; automated structure searching, by utilizing a large amount of computing power, reduces the time cost of requiring a large number of in-field experts to perform a large number of trial and error prior to the search.
Description
Technical Field
The invention relates to the technical field of computer vision and the field of embedded chips, in particular to a method and a system for constructing a target detection network based on chip operation.
Background
At present, object recognition and detection segmentation are widely used in the technical fields of image processing, face recognition, monitoring and tracking and the like. The deep neural network can effectively learn hierarchical representation from data, and is widely applied to the technical field of computer vision, such as object recognition, detection segmentation and the like. In the target detection task, a wide variety of target detection architectures are developed, such as a single-stage detection model, a two-stage detection model, a cascaded multi-stage detection model, and an anchor-free (anchor-free) detection model. By replacing the basic network architecture of the detection models, different tradeoffs of speed, precision and memory consumption can be realized. At present, detection models are also increasingly deployed in edge devices, such as mobile phone chips.
First, compared with the GPU, the chip in the edge device has a small memory and the chip architecture is very different. However, most of the existing detection model architectures are designed for GPU equipment, and the GPU and the edge equipment chip have different operation efficiencies for different operators; secondly, there are some basic models, such as MobileNet, ShuffleNet, etc., designed for edge devices, but they are mainly based on target identification tasks and public data sets, and the target detection and target identification tasks have different requirements for network architectures, etc., and different data sets also have different requirements for different detection architectures.
Disclosure of Invention
In order to solve the above problems, the present invention provides a method and a system for constructing a target detection network based on chip operation, which can automatically and custom design a corresponding detection model architecture according to different edge device chips and different detection data sets by using the existing GPU computing power.
According to an aspect of the present invention, there is provided a method for constructing an object detection network, including:
determining a network architecture search space of the target detection network, wherein the search space defines a network architecture search range and a detection submodule;
determining a search strategy, wherein the search strategy defines a method for searching the search space;
determining a model evaluation index, wherein the model evaluation index is an objective function of the search strategy.
According to another aspect of the present invention, there is also provided a system for constructing an object detection network, including:
the search space determining module is used for determining a network architecture search space of the target detection network, and the search space defines a network architecture search range and a detection submodule;
a search strategy determination module for determining a search strategy defining a method of searching the search space;
and the evaluation index determining module is used for determining a model evaluation index, and the model evaluation index is an objective function of the search strategy.
Compared with the prior art, the method for automatically constructing the target detection network based on the chip and the detection task, disclosed by the invention, has the following beneficial effects: (1) the method avoids the defects of the prior method that one search strategy is adopted, various search strategies are combined, and the network architecture is searched from coarse to fine, so that the search efficiency is improved, and a better local optimal solution is obtained. (2) The method and the device have the advantages that various model evaluation indexes are fused, the accuracy is only adopted as a measurement index in the conventional network model architecture search, or the evaluation index of the substitution speed or the memory is designed based on experience. (3) Automated structure searching, by utilizing a large amount of computing power, reduces the time cost of requiring a large number of in-field experts to perform a large number of trial and error prior to the search.
Drawings
Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are only for purposes of illustrating the preferred embodiments and are not to be construed as limiting the invention. Also, like reference numerals are used to refer to like parts throughout the drawings. In the drawings:
FIG. 1 is a flow chart of a method for constructing a target detection network according to the present invention;
FIG. 2 is a flow diagram illustrating steps for a coarse-to-fine search in accordance with an embodiment of the present invention;
FIG. 3 illustrates a basic flow diagram of an automated search according to an embodiment of the present invention;
FIG. 4 is a block flow diagram illustrating a coarse search according to an embodiment of the present invention;
FIG. 5 is a schematic diagram illustrating a fine search according to an embodiment of the present invention;
FIG. 6 illustrates a detailed flow diagram of different metric prediction according to an embodiment of the present invention;
fig. 7 is a diagram showing a construction system of the object detection network of the present invention.
Detailed Description
Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.
As shown in fig. 1, a method for constructing a target detection network according to the present invention includes the following steps:
s1, determining a network architecture search space of the target detection network, wherein the search space defines a network architecture search range and a detection submodule;
the search space defines the scope of the network architecture search, i.e., what detection architecture model can be characterized in the search space. The invention can use different network structure operators and detection submodules, such as depth separable convolution, space separable convolution, multi-space convolution, pyramid feature extractor and the like. A good search space can effectively improve the search efficiency.
S2, determining a search strategy, wherein the search strategy defines a method for searching the search space;
the search strategy defines how to effectively search the search space, such as methods based on genetic algorithm, reinforced learning, gradient descent and the like.
S3, determining a model evaluation index, wherein the model evaluation index is an objective function of the search strategy.
The model evaluation index, i.e. the objective function of the search algorithm, may indicate what model is a good model.
As shown in fig. 2, the present invention discloses a search strategy of a chip-based automated target detection network architecture, which includes:
rough searching: this step can be performed on a public detection data set, and in combination with various evaluation indexes, basic detection submodules adapted to a specific chip architecture are searched out, and these basic detection submodule classes include: some like the basic operator module of the depth separable convolution, and some like the hierarchical feature extraction module of the gold pyramid structure. These sub-modules searched on the public data set can provide raw materials for designing more complex detection architectures.
A differentiable macro network architecture search space is designed for a specific data set.
Fine searching: and optimizing the differentiable network architecture space by using a gradient method to obtain a final detection network architecture suitable for a specific task.
As shown in fig. 3, a basic search strategy of an automated network architecture according to an embodiment of the present invention is illustrated, and the rough search and the fine search both follow the strategy, including:
defining a network search space: in the coarse search, the search space is mainly different detection sub-modules, and in the fine search, the search space is a microminiature macro detection model architecture formed by the combination of the sub-modules.
Defining a search strategy: the search strategy adopted in the rough search is reinforcement learning, such as a Reinforce method, and a Reinforce algorithm is a classic strategy gradient algorithm in reinforcement learning. The search strategy in the fine search is a gradient method, namely the whole network search space is a differentiable and can be directly trained through a back propagation algorithm, and the detailed process is shown in fig. 6.
The model was evaluated: the invention uses various indexes to evaluate the searched model, for example, the various indexes include the precision of the model on the test data set, the running speed on the edge device and the memory occupation on the edge device. However, the conventional automatic search algorithm usually only adopts the precision as the model evaluation index. In the rough search, the invention can evaluate the model on the public data set, and in the fine search, the invention evaluates the model by using the specific data set corresponding to the task.
As shown in fig. 4, a block diagram of a coarse search process according to an embodiment of the present invention is shown.
The steps of the rough search are consistent with the basic steps shown in FIG. 3, for example, in the rough search, the RNN controller is used to generate a training submodule. The sampled training submodule is trained in a training set and tested in a testing set, performance evaluation obtained in the testing set is returned to the RNN controller as reward (reward), and the network weight of the RNN controller is driven to be updated by using a REINFORCE algorithm.
In the process of generating the sub-network, an RNN (cyclic network) iterates t time steps, and under each time step, the RNN outputs different parameters corresponding to a network structure, such as the size of a convolution kernel, the number of feature layers and the like, and a training sub-module can be defined by the different network architecture parameters.
As shown in FIG. 5, a schematic diagram of a crawl according to an embodiment of the invention is shown.
In fig. 5, the connection between the features of different layers is not clear at first, and the connection between two feature layers may have a variety of operation possibilities, such as convolution, upsampling, depth separable convolution, or sub-module operations obtained by searching. The connection between the feature layers adopts a mixed form of various operations, different operations are endowed with an initialized connection probability, the connection probabilities corresponding to some operations are strengthened through gradient training, and finally the target detection network architecture after fine search is obtained. For example, as for feature layer 1 and feature layer 2, it is unclear what connection type is good, and here, 3x3 convolution, 5x5 convolution or multi-space convolution can be used. In the figure, different convolution operations are represented by different colored lines, the thickness represents the probability α of different convolution operations, and the initial value probabilities are the same, and are learnable variables. Feature layer 1 gets different inputs to feature layer 2 under different convolution operations, and the total input to feature layer 2 is the result of the alpha weighted input provided by feature layer 1. While the network structure is optimized with gradients, the distribution of α also changes. The probability of some convolutions increases and the probability of some convolutions decreases. After training is completed, we only keep the convolution operation with the highest probability.
As shown in fig. 6, a detailed diagram of different index predictions according to an embodiment of the present invention is shown.
For a certain network sub-module or architecture, the operation speed or the memory size on the chip can be obtained only by actually deploying the network sub-module or architecture on the chip for testing. For the automated network architecture search, a great deal of time is consumed for acquiring the speed and the memory through interaction with a chip. The invention samples some structures, obtains corresponding training data by actually measuring the speed and the memory ratio on the chip, and trains the small network to learn and predict the corresponding performance of a certain sub-structure or module. In one example, for a 3 × 3 convolution module, different input sizes (including the number of signatures, the spatial size of the signatures) are run on-chip at different times. The corresponding chip processing speed and memory consumption can be actually measured on the chip by sampling the number of different characteristic diagrams and the space sizes of the different characteristic diagrams. Therefore, data of different module sizes, processing speeds and memory consumption are obtained, a corresponding small network can be trained by using the data, and the network can help to predict the chip processing time and the memory consumption of the 3x3 convolution module under different input sizes.
As shown in the following equations, the composition of an objective function according to an embodiment of the present invention is shown.
L＝Lreg+Lcls+λ1Llatency+λ2Lmemory
The objective function L comprises basic detection task performance indexes, and the speed and memory indexes are predicted by a network when the edge device runs. L isclsAs a loss function of the target classification result, LregIs a loss function of the regression results of the target box, LmemoryFor the time response delay loss function of the network structure on the chip, LmemoryFor a network architecture memory consumption loss function on a chip, λ1，λ2Is the weight of the corresponding loss function. Wherein. It should be noted that the inspection task performance index is obtained in the inspection verification dataset.
As shown in fig. 7, the present invention further provides a system 100 for constructing an object detection network, including:
a search space determining module 101, configured to determine a network architecture search space of the target detection network, where the search space defines a network architecture search range and a detection sub-module;
a search strategy determination module 102, configured to determine a search strategy, where the search strategy defines a method for searching the search space;
an evaluation index determining module 103, configured to determine a model evaluation index, where the model evaluation index is an objective function of the search policy.
The invention discloses a method and a system for automatically constructing a target detection network based on a chip and a detection task, which have the following beneficial effects: (1) the method avoids the defects of the prior method that one search strategy is adopted, various search strategies are combined, and the network architecture is searched from coarse to fine, so that the search efficiency is improved, and a better local optimal solution is obtained. (2) The method and the device have the advantages that various model evaluation indexes are fused, the accuracy is only adopted as a measurement index in the conventional network model architecture search, or the evaluation index of the substitution speed or the memory is designed based on experience. (3) Automated structure searching, by utilizing a large amount of computing power, reduces the time cost of requiring a large number of in-field experts to perform a large number of trial and error prior to the search.
The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the appended claims.
Claims (10)
1. A method for constructing a target detection network is characterized by comprising the following steps:
determining a network architecture search space of the target detection network, wherein the search space defines a network architecture search range and a detection submodule;
determining a search strategy, wherein the search strategy defines a method for searching the search space;
determining a model evaluation index, wherein the model evaluation index is an objective function of the search strategy.
2. The method of claim 1,
the detection submodule includes one or more of: depth separable convolution, space separable convolution, multi-space convolution, pyramid feature extractor.
3. The method of claim 1,
the search strategy includes one or more of: genetic algorithm-based searches, reinforcement learning-based searches, gradient descent-based searches.
4. The method of claim 1,
the search strategy comprises the following steps:
combining a plurality of evaluation indexes, and performing rough search detection on the basis of reinforcement learning;
combining the detection sub-modules to form a micro network searching space;
and performing fine search on the micro network search space based on a gradient descent method to obtain a final target detection network.
5. The method of claim 4,
the combined multiple evaluation indexes and reinforcement learning-based coarse search detection submodule comprises:
the RNN controller generates a training submodule;
training the training submodule in a training set, testing the training set in a testing set, returning the performance evaluation obtained in the testing set to the RNN controller, and driving the update of the network weight of the RNN controller by using a REINFORCE algorithm.
6. The method of claim 4,
the fine searching of the micro-network search space based on the gradient descent method to obtain a final target detection network comprises the following steps:
assigning initialized connection probabilities to different operations of the micro-web searchable space;
and (4) performing gradient training, reserving the operation with the highest connection probability and abandoning other operations to obtain the final target detection network.
7. The method of claim 1,
the model assessment indicators include one or more of: accuracy of the model on the test data set, running speed on the edge device, memory footprint on the edge device.
8. The method of claim 7,
the determining of the model evaluation index comprises:
fusing various model evaluation indexes;
and predicting different model evaluation indexes by using the small network.
9. The method according to claim 7 or 8,
the objective function L is calculated as follows:
L＝Lreg+Lcls+λ1Llatency+λ2Lmemory
wherein L isclsAs a loss function of the target classification result, LregIs a loss function of the regression results of the target box, LlatencyFor the time response delay loss function, L, of the target detection network on chipmemoryFor the memory consumption loss function of the target detection network on the chip, lambda1，λ2Is the weight of the corresponding loss function.
10. A system for constructing an object detection network, comprising:
the search space determining module is used for determining a network architecture search space of the target detection network, and the search space defines a network architecture search range and a detection submodule;
a search strategy determination module for determining a search strategy defining a method of searching the search space;
and the evaluation index determining module is used for determining a model evaluation index, and the model evaluation index is an objective function of the search strategy.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910713035.2A CN110659721B (en) | 2019-08-02 | 2019-08-02 | Method and system for constructing target detection network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910713035.2A CN110659721B (en) | 2019-08-02 | 2019-08-02 | Method and system for constructing target detection network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110659721A true CN110659721A (en) | 2020-01-07 |
| CN110659721B CN110659721B (en) | 2022-07-22 |
Family
ID=69036882
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910713035.2A Active CN110659721B (en) | 2019-08-02 | 2019-08-02 | Method and system for constructing target detection network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110659721B (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111325328A (en) * | 2020-03-06 | 2020-06-23 | 上海商汤临港智能科技有限公司 | Neural network generation method, data processing method and device |
| CN111428224A (en) * | 2020-04-02 | 2020-07-17 | 苏州杰锐思智能科技股份有限公司 | Computer account login method based on face recognition |
| CN112101525A (en) * | 2020-09-08 | 2020-12-18 | 南方科技大学 | Method, device and system for designing neural network through NAS |
| CN112115914A (en) * | 2020-09-28 | 2020-12-22 | 北京市商汤科技开发有限公司 | Target detection method and device, electronic equipment and storage medium |
| CN112258486A (en) * | 2020-10-28 | 2021-01-22 | 汕头大学 | Retinal vessel segmentation method for fundus image based on evolutionary neural architecture search |
| CN113128308A (en) * | 2020-01-10 | 2021-07-16 | 中南大学 | Pedestrian detection method, device, equipment and medium in port scene |
| WO2021143883A1 (en) * | 2020-01-15 | 2021-07-22 | 华为技术有限公司 | Adaptive search method and apparatus for neural network |
| CN113283426A (en) * | 2021-04-30 | 2021-08-20 | 南京大学 | Embedded target detection model generation method based on multi-target neural network search |
| CN113408692A (en) * | 2020-03-16 | 2021-09-17 | 顺丰科技有限公司 | Network structure searching method, device, equipment and storage medium |
| CN113591795A (en) * | 2021-08-19 | 2021-11-02 | 西南石油大学 | Lightweight face detection method and system based on mixed attention feature pyramid structure |
| CN113743605A (en) * | 2021-06-16 | 2021-12-03 | 温州大学 | Method for searching smoke and fire detection network architecture based on evolution method |
| CN113780542A (en) * | 2021-09-08 | 2021-12-10 | 北京航空航天大学杭州创新研究院 | FPGA-oriented multi-target network structure construction method |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107610091A (en) * | 2017-07-31 | 2018-01-19 | 阿里巴巴集团控股有限公司 | Vehicle insurance image processing method, device, server and system |
| CN108536791A (en) * | 2018-03-30 | 2018-09-14 | 北京百度网讯科技有限公司 | Searching method, equipment based on neural network and storage medium |
| CN108595301A (en) * | 2018-03-26 | 2018-09-28 | 中国科学院计算技术研究所 | A kind of server energy consumption prediction technique and system based on machine learning |
| CN108985458A (en) * | 2018-07-23 | 2018-12-11 | 东北大学 | A kind of double tree monte carlo search algorithms of sequential synchronous game |
| CN109615073A (en) * | 2018-12-03 | 2019-04-12 | 郑州云海信息技术有限公司 | A kind of construction method of neural network model, equipment and storage medium |
| CN109948425A (en) * | 2019-01-22 | 2019-06-28 | 中国矿业大学 | A kind of perception of structure is from paying attention to and online example polymerize matched pedestrian's searching method and device |
-
2019
- 2019-08-02 CN CN201910713035.2A patent/CN110659721B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107610091A (en) * | 2017-07-31 | 2018-01-19 | 阿里巴巴集团控股有限公司 | Vehicle insurance image processing method, device, server and system |
| CN108595301A (en) * | 2018-03-26 | 2018-09-28 | 中国科学院计算技术研究所 | A kind of server energy consumption prediction technique and system based on machine learning |
| CN108536791A (en) * | 2018-03-30 | 2018-09-14 | 北京百度网讯科技有限公司 | Searching method, equipment based on neural network and storage medium |
| CN108985458A (en) * | 2018-07-23 | 2018-12-11 | 东北大学 | A kind of double tree monte carlo search algorithms of sequential synchronous game |
| CN109615073A (en) * | 2018-12-03 | 2019-04-12 | 郑州云海信息技术有限公司 | A kind of construction method of neural network model, equipment and storage medium |
| CN109948425A (en) * | 2019-01-22 | 2019-06-28 | 中国矿业大学 | A kind of perception of structure is from paying attention to and online example polymerize matched pedestrian's searching method and device |
Non-Patent Citations (2)
| Title |
|---|
| HANXIAO LIU ET AL.: ""DARTS: DIFFERENTIABLE ARCHITECTURE SEARCH"", 《ARXIV》 * |
| THOMAS ELSKEN ET AL.: ""Neural Architecture Search: A Survey"", 《ARXIV》 * |
Cited By (21)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113128308B (en) * | 2020-01-10 | 2022-05-20 | 中南大学 | Pedestrian detection method, device, equipment and medium in port scene |
| CN113128308A (en) * | 2020-01-10 | 2021-07-16 | 中南大学 | Pedestrian detection method, device, equipment and medium in port scene |
| JP2023510566A (en) * | 2020-01-15 | 2023-03-14 | 華為技術有限公司 | Adaptive search method and apparatus for neural networks |
| WO2021143883A1 (en) * | 2020-01-15 | 2021-07-22 | 华为技术有限公司 | Adaptive search method and apparatus for neural network |
| JP7366274B2 (en) | 2020-01-15 | 2023-10-20 | 華為技術有限公司 | Adaptive search method and device for neural networks |
| CN111325328A (en) * | 2020-03-06 | 2020-06-23 | 上海商汤临港智能科技有限公司 | Neural network generation method, data processing method and device |
| CN111325328B (en) * | 2020-03-06 | 2023-10-24 | 上海商汤临港智能科技有限公司 | Neural network generation method, data processing method and device |
| CN113408692A (en) * | 2020-03-16 | 2021-09-17 | 顺丰科技有限公司 | Network structure searching method, device, equipment and storage medium |
| CN111428224B (en) * | 2020-04-02 | 2023-10-13 | 苏州杰锐思智能科技股份有限公司 | Face recognition-based computer account login method |
| CN111428224A (en) * | 2020-04-02 | 2020-07-17 | 苏州杰锐思智能科技股份有限公司 | Computer account login method based on face recognition |
| CN112101525A (en) * | 2020-09-08 | 2020-12-18 | 南方科技大学 | Method, device and system for designing neural network through NAS |
| CN112115914A (en) * | 2020-09-28 | 2020-12-22 | 北京市商汤科技开发有限公司 | Target detection method and device, electronic equipment and storage medium |
| CN112115914B (en) * | 2020-09-28 | 2023-04-07 | 北京市商汤科技开发有限公司 | Target detection method and device, electronic equipment and storage medium |
| CN112258486B (en) * | 2020-10-28 | 2023-04-07 | 汕头大学 | Retinal vessel segmentation method for fundus image based on evolutionary neural architecture search |
| CN112258486A (en) * | 2020-10-28 | 2021-01-22 | 汕头大学 | Retinal vessel segmentation method for fundus image based on evolutionary neural architecture search |
| CN113283426A (en) * | 2021-04-30 | 2021-08-20 | 南京大学 | Embedded target detection model generation method based on multi-target neural network search |
| CN113743605A (en) * | 2021-06-16 | 2021-12-03 | 温州大学 | Method for searching smoke and fire detection network architecture based on evolution method |
| CN113591795B (en) * | 2021-08-19 | 2023-08-08 | 西南石油大学 | Lightweight face detection method and system based on mixed attention characteristic pyramid structure |
| CN113591795A (en) * | 2021-08-19 | 2021-11-02 | 西南石油大学 | Lightweight face detection method and system based on mixed attention feature pyramid structure |
| CN113780542A (en) * | 2021-09-08 | 2021-12-10 | 北京航空航天大学杭州创新研究院 | FPGA-oriented multi-target network structure construction method |
| CN113780542B (en) * | 2021-09-08 | 2023-09-12 | 北京航空航天大学杭州创新研究院 | Method for constructing multi-target network structure facing FPGA |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110659721B (en) | 2022-07-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110659721B (en) | Method and system for constructing target detection network | |
| CN109947086B (en) | Mechanical fault migration diagnosis method and system based on counterstudy | |
| Xu et al. | Identification framework for cracks on a steel structure surface by a restricted Boltzmann machines algorithm based on consumer‐grade camera images | |
| CN109120462B (en) | Method and device for predicting opportunistic network link and readable storage medium | |
| Wang et al. | Pixel-wise fabric defect detection by CNNs without labeled training data | |
| CA3037941A1 (en) | Method and system for generating and using vehicle pricing models | |
| US20240005164A1 (en) | Neural Network Training Method and Related Device | |
| CN111861909A (en) | Network fine-grained image denoising and classifying method | |
| CN116151319A (en) | Method and device for searching neural network integration model and electronic equipment | |
| CN114881343A (en) | Short-term load prediction method and device of power system based on feature selection | |
| CN114048817A (en) | Deep learning input set priority test method based on variation strategy | |
| CN113065634B (en) | Image processing method, neural network training method and related equipment | |
| CN112733724B (en) | Relativity relationship verification method and device based on discrimination sample meta-digger | |
| Katam et al. | A review on technological advancements in the field of data driven structural health monitoring | |
| Fan | Data mining model for predicting the quality level and classification of construction projects | |
| Alfaz et al. | A deep convolutional neural network based approach to classify and detect crack in concrete surface using xception | |
| CN115858388A (en) | Test case priority ordering method and device based on variation model mapping chart | |
| CN113393023B (en) | Mold quality evaluation method, apparatus, device and storage medium | |
| CN112859034B (en) | Natural environment radar echo amplitude model classification method and device | |
| CN114298290A (en) | Neural network coding method and coder based on self-supervision learning | |
| CN109376243A (en) | File classification method and device | |
| CN117693754A (en) | Training masked automatic encoders for image restoration | |
| Kavitha et al. | Explainable AI for Detecting Fissures on Concrete Surfaces Using Transfer Learning | |
| Ceylan et al. | Siamese inception time network for remaining useful life estimation | |
| El Khatib et al. | Neural network-based parking system object detection and predictive modeling |
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 | ||
| CB02 | Change of applicant information |
Address after: Room 101, building 1, block C, Qianjiang Century Park, ningwei street, Xiaoshan District, Hangzhou City, Zhejiang Province Applicant after: Hangzhou Weiming Information Technology Co.,Ltd. Applicant after: Institute of Information Technology, Zhejiang Peking University Address before: Room 288-1, 857 Xinbei Road, Ningwei Town, Xiaoshan District, Hangzhou City, Zhejiang Province Applicant before: Institute of Information Technology, Zhejiang Peking University Applicant before: Hangzhou Weiming Information Technology Co.,Ltd. |
|
| CB02 | Change of applicant information | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| EE01 | Entry into force of recordation of patent licensing contract |
Application publication date: 20200107 Assignee: Zhejiang smart video security Innovation Center Co.,Ltd. Assignor: Institute of Information Technology, Zhejiang Peking University Contract record no.: X2022330000930 Denomination of invention: A method and system for constructing target detection network Granted publication date: 20220722 License type: Common License Record date: 20221229 |
|
| EE01 | Entry into force of recordation of patent licensing contract |