CN108171117A - Electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing - Google Patents
Electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing Download PDFInfo
- Publication number
- CN108171117A CN108171117A CN201711268417.6A CN201711268417A CN108171117A CN 108171117 A CN108171117 A CN 108171117A CN 201711268417 A CN201711268417 A CN 201711268417A CN 108171117 A CN108171117 A CN 108171117A
- Authority
- CN
- China
- Prior art keywords
- image
- layer
- task
- input
- module
- 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
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
- G06V10/94—Hardware or software architectures specially adapted for image or video understanding
- G06V10/955—Hardware or software architectures specially adapted for image or video understanding using specific electronic processors
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/94—Hardware or software architectures specially adapted for image or video understanding
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of electric power artificial intelligence visual analysis systems based on multinuclear heterogeneous Computing, which is characterized in that including using lightweight neural network as core, including a kind of multinuclear heterogeneous Computing module and business application module;It is accessed between business application module and multinuclear heterogeneous Computing module by network service interface, transmits data;The multinuclear heterogeneous Computing module includes GPU calculate nodes, CPU storage managements node, CPU calculate nodes, is connected between each node by interchanger;Business application module includes image management module, image labeling module, model training module, algorithm application module;The present invention is based on multinuclear heterogeneous Computing frames, it can efficiently realize the training of online or offline image data, form lightweight rapid image categorization model, realize electric power intranet and extranet graphic service application, the value for playing image data can be maximized, has preferable application prospect.
Description
Technical field
The present invention relates to Signal and Information Processing fields, and in particular to a kind of electric power people based on multinuclear heterogeneous Computing
Work intelligent vision analysis system.
Background technology
Artificial intelligence and deep learning are advanced automatic field scientific and technological means, at present in image procossing, nature
The fields such as language identification are applied widely.Currently, artificial intelligence analysis's method has begun to apply to all conglomeraties, including
Intelligent transportation, intelligent medical treatment, smart home, automatic Pilot, Intelligent hardware etc..Various depth convolutional neural networks are (such as in recent years
Lenet, Alexnet, VGG, Resnet, Xnception etc.) it emerges in an endless stream, it is widely used in the meters such as image identification, target detection
Calculation machine visual field.Headend equipment under power industry, the business scenarios such as substation, transmission line of electricity can all acquire largely daily
Image and video data, related service department is there are urgent image and video data analysis and identification demand, both at home and abroad
There is more mature deep learning image recognition technology.But most of deep learning image recognition technology calculates consumption
Greatly, the speed of service is slow.Heterogeneous Computing can allow different types of computing device that can share the process of calculating and as a result, same
When continue to optimize and accelerate calculate process, it is made to have higher calculating efficiency.Heterogeneous Computing is at home and abroad just fast
Hail exhibition, especially CPU+GPU Heterogeneous Computings frame becomes research hotspot in recent years.But it has not yet to see different based on multinuclear
The lightweight neural network of structure parallel computation frame and its application in the identification of electric power image.
To solve the problems, such as that intelligent service application, invention provide one existing for existing power industry gathered data
Kind new, effective solution.
Invention content
To solve deficiency of the prior art, the present invention provides a kind of artificial intelligence of electric power based on multinuclear heterogeneous Computing
Energy visual analysis system, towards electric power image data, using lightweight neural network as core, based on multinuclear heterogeneous Computing frame
Frame can efficiently realize the training of online or offline image data, form lightweight rapid image categorization model, it can be achieved that inside and outside electric power
Net image service application.
In order to realize above-mentioned target, the present invention adopts the following technical scheme that:A kind of electricity based on multinuclear heterogeneous Computing
Power artificial intelligence visual analysis system, which is characterized in that using lightweight neural network as core, including multinuclear heterogeneous Computing
Module and business application module;It is visited between business application module and multinuclear heterogeneous Computing module by network service interface
It asks, transmits data;
The multinuclear heterogeneous Computing module includes GPU calculate nodes, CPU storage managements node, CPU calculate nodes,
It is connected between each node by interchanger;GPU calculate nodes perform for model training and intelligent task, complete intensive calculations;
CPU storage managements node stores for data, embeds relevant database and non-relational database;CPU calculate nodes are used for
Scientific algorithm, while GPU calculate nodes is assisted to complete part intensive calculations;
Business application module includes image management module, image labeling module, model training module, algorithm application module;
Image management module is used to manage electric power intranet and extranet graphic service;Image labeling module is used to provide instruction for lightweight neural network
Practice the markup information of data set;Model training module is used for the training lightweight neural network in multinuclear heterogeneous Computing module
Model;Algorithm application module utilizes lightweight god towards electric power intranet and extranet graphic service in multinuclear heterogeneous Computing module
Intellectual analysis task is performed through network model.
A kind of aforementioned electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing, it is characterized in that:Institute
Stating image management module, other business platforms communicate by web service and power industry, other business platforms of power industry
Obtain the image handled through system and video data and its related information;Image management module obtains the original image of electric system
With video data and its related information;Image management module also receives the image and video data locally uploaded;The electric power row
Other business platforms of industry include unified video monitoring platform and fortune inspection, scheduling, marketing, capital construction power informatization platform.
A kind of aforementioned electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing, it is characterized in that:Institute
The attribute that related information includes image and video is stated, including view classification, device tree, scene tree, tag tree, defect tree, profession
Type, document source;
The view classification includes the headend equipment of two class of image and video, the device tree description shooting image or video
The tool of the electric power scene of address, the scene tree description shooting image or video, the tag tree description shooting image or video
To hold in vivo, formulated by system manager, the defect tree is option, defect problem existing for description shooting image or video,
It is formulated by system manager, the power specialty title of many types description shooting image or video, the document source is retouched
State the means of shooting image or video.
A kind of aforementioned electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing, it is characterized in that:Institute
Relevant database is stated for storing the attribute of image and video, the non-relational database is used to store image and video counts
According to.
A kind of aforementioned electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing, it is characterized in that:Institute
Lightweight neural network model is stated, is specifically included:
Whole network includes 20 layers, and wherein convolutional layer has 17 layers, and pond layer has 1 layer, and full articulamentum has 1 layer, convolutional layer position
In network front end, pond layer and full articulamentum are located at network backend;
The convolution kernel used is 1 × 1 small convolution kernel and 1 × 3,3 × 1 asymmetric convolution kernel;
Have residual error structure, accelerate convergence while network depth is kept;
Network have abstention, batch standardization can arrangement parameter, introduce regularization mechanism;
Network is provided for the normalization of input data and enhancing function, and all input pictures are normalized to 224 × 224
Pixel value, and provide mirror image reversal, cutting, a variety of data enhancements of tone reversal, expanded training dataset;
Network last one layer i.e. loss function layer have can suitability, the use using softmax functions as loss function
Single labeling is classified using entropy function is intersected as multi-tag.
A kind of aforementioned electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing, it is characterized in that:Institute
Stating lightweight neural network framework is:
1st layer is convolutional layer, and step-length 2, the size of input is 224 × 224 × 3, using two cascade convolution kernels 1 ×
3 and 3 × 1;
2nd layer is convolutional layer, and step-length 1, the size of input is 112 × 112 × 32, using two cascade convolution kernels 1
× 3 and 3 × 1;
3rd layer is convolutional layer, and step-length 1, the size of input is 112 × 112 × 32, using 1 × 1 convolution kernel;
4th layer is convolutional layer, and step-length 2, the size of input is 112 × 112 × 64, using two cascade convolution kernels 1
× 3 and 3 × 1;
5th layer is convolutional layer, and step-length 1, the size of input is 56 × 56 × 64, using two cascade convolution kernels 1 × 3
With 3 × 1;
6th layer is convolutional layer, and step-length 1, the size of input is 56 × 56 × 128, using 1 × 1 convolution kernel;
7th layer is convolutional layer, and step-length 2, the size of input is 56 × 56 × 128, using two cascade convolution kernels 1 ×
3 and 3 × 1;
8th layer is convolutional layer, and step-length 1, the size of input is 28 × 28 × 128, using two cascade convolution kernels 1 ×
3 and 3 × 1;
9th layer is convolutional layer, and step-length 1, the size of input is 28 × 28 × 256, using 1 × 1 convolution kernel;
10th layer is convolutional layer, and step-length 2, the size of input is 28 × 28 × 256, using two cascade convolution kernels 1
× 3 and 3 × 1;
11th layer is convolutional layer, and step-length 1, the size of input is 14 × 14 × 256, using two cascade convolution kernels 1
× 3 and 3 × 1;
12nd layer is convolutional layer, and step-length 1, the size of input is 14 × 14 × 512, using 1 × 1 convolution kernel;
13rd layer is convolutional layer, and step-length 2, the size of input is 14 × 14 × 512, using two cascade convolution kernels 1
× 3 and 3 × 1;
14th layer is convolutional layer, and step-length 1, the size of input is 7 × 7 × 512, using two cascade convolution kernels 1 × 3
With 3 × 1;
15th layer is convolutional layer, and step-length 1, the size of input is 7 × 7 × 1024, using 1 × 1 convolution kernel;
16th layer is convolutional layer, and step-length 2, the size of input is 7 × 7 × 1024, using two cascade convolution kernels 1 ×
3 and 3 × 1;
17th layer is convolutional layer, and step-length 1, the size of input is 7 × 7 × 1024, using two cascade convolution kernels 1 ×
3 and 3 × 1;
18th layer is average pond layer, and step-length 1, the size of input is 7 × 7 × 1024, and pond size is 7 × 7;
19th layer is full articulamentum, and the size of input is 1 × 1 × 1024, includes 1000 neurons;
20th layer is loss function layer, can be adapted to, and using softmax functions as loss function with single labeling, is made
Classified by the use of entropy function is intersected as multi-tag.
A kind of aforementioned electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing, it is characterized in that:Institute
Electric power intranet and extranet graphic service is stated by following grade classification:
Level-one task includes image general utility functions task and graphic service application task;
Second task include image general utility functions task under image duplicate removal, low-quality image rejecting, video code conversion, video
Task towards fortune inspection under compression, fast browsing and graphic service application task, the task towards scheduling, towards capital construction
Task, the task of marketing oriented;
Three-level task include towards under the task of fortune inspection working truck detection, wire foreign matter detection, tree bamboo detection of the growth,
Wire icing detection, pyrotechnics detection, gold utensil rust detection, gold utensil loss detection, insulator Safety check-up, insulator lose inspection
Survey, insulator contamination detection, substation's meter number identification, the inspection of installation for transformer Oil Leakage Detecting, transformer station personnel abnormal behaviour
It surveys, the detection of transformer station personnel dressing specification, transformer station personnel discrepancy detects, pyrotechnics detects;Transformation towards under the task of scheduling
Device on off state identifies and disconnecting switch state recognition;Towards under the task of capital construction inlet and outlet vehicle detection, inlet and outlet car plate
Identification, inlet and outlet personnel detection, inlet and outlet personnel unusual checking, the detection of personnel's dressing specification, open fire detection;Marketing oriented
Task under business environmental quality detection, attendant to hilllock leave the post situation detection, the quality testing of attendant's appearance instrument,
Attendant works' behavior quality testing, customer behavior analysis and anomalous identification.
A kind of aforementioned electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing, it is characterized in that:Institute
It states image labeling and is divided into two kinds, one kind is system automatic marking, i.e. input picture is defeated with model to lightweight neural network model
The classification results gone out are markup information;Another kind is that user knows icon note, i.e., user checks image data, marks out image manually
The object classification inside included and region.
A kind of aforementioned electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing, it is characterized in that:Institute
Model training module is stated using the image and video data by image labeling as data set, towards image classification, target detection, figure
As the machine learning task of segmentation, carry out training in multinuclear heterogeneous Computing module;Can suspend in training process, cancel,
Continue training mission;If training result is not ideal enough, network parameter is changed, iterations continue to train, until being satisfied with
Model, satisfied model is stored in GPU calculate nodes, replaceable original model;Support on-line training and off-line training
Two ways;Off-line training refers to that model training task will be opened after the disposable input system of all data;On-line training refers to mould
After type training mission starts, new data can be inputted to system, addition just continues to train after the model of iteration.
A kind of aforementioned electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing, it is characterized in that:Institute
The algorithm model of algorithm application module is stated there are two source, when the lightweight neural network model that model training module obtains,
Second is that it is built in other maturity models of GPU calculate nodes;It can suspend, cancel, continuing training mission during tasks carrying, appointing
Business can delete task after terminating;After tasks carrying, user can be according to results operation image and video data
The advantageous effect that the present invention is reached:
1. present invention employs the multinuclear heterogeneous Computing frame of suitable this system, wherein the division of labor of each node is clear and definite, patrol
It is clear to collect, and interaction is reasonable, and GPU calculate nodes undertake model training and intelligent task performs function, completion intensive calculations;CPU is deposited
Storage management node undertakes data storage function, embeds a kind of relevant database and a kind of non-relational database, and CPU calculates section
Point undertakes scientific algorithm function, while GPU calculate nodes is assisted to complete part intensive calculations function;
2. proposing a kind of algorithm core of lightweight neural network as system, network includes 20 layers, employs rouleau
Product core and asymmetric convolution nuclear structure, residual error structure, abstention mechanism, batch standardized mechanism, can effectively improve classification accuracy, together
When accelerate convergence, enhance network generalization, keep accuracy rate while improve training speed and algorithm execution speed;Simultaneously
Network is provided for the normalization of input data and enhancing function, can effective EDS extended data set;Lightweight rapid image categorization mould
Type is based on existing electric power data, and parameter is got based on electric power data training, is not based on general data set, has
Innovative and practical value;
3. system has scalability and universality, the data transfer of other business platforms of support and power industry is supported
The GPU nodes of multinuclear isomery and cpu node parallel computation, have expansible algorithm network model, and suitable power scene is set
Standby, personnel, a variety of electric power analysis occasions of event category.
This system is based on multinuclear heterogeneous Computing frame, can efficiently realize the training of online or offline image data, be formed
Lightweight rapid image categorization model can maximize the valency for playing image data, it can be achieved that electric power intranet and extranet graphic service application
Value, has preferable application prospect.
Description of the drawings
Fig. 1 is present system logical architecture block diagram.
Specific embodiment
The invention will be further described below in conjunction with the accompanying drawings.Following embodiment is only used for clearly illustrating the present invention
Technical solution, and be not intended to limit the protection scope of the present invention and limit the scope of the invention.
As shown in Figure 1, a kind of electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing frame, with light
Magnitude neural network be core, the business application module including a kind of multinuclear heterogeneous Computing module and based on the frame;
Multinuclear heterogeneous Computing module includes GPU (Graphic Processing Unit graphics processors) and calculates section
Point, CPU (Central Processing Unit central processing units) storage management node, CPU calculate nodes, between each node
It is connected by a kind of infiniband (" Convertion cable " technology for supporting mostly concurrently to link) interchanger;GPU calculate nodes are used for
Model training and intelligent task perform, and complete intensive calculations;CPU storage managements node stores for data, embeds a kind of relationship
Type database and a kind of non-relational database, CPU calculate nodes are used for scientific algorithm, while GPU calculate nodes is assisted to complete
Part intensive calculations function.
Business application module includes image management module, image labeling module, model training module, algorithm application module;
Image management module is used to manage electric power intranet and extranet graphic service;Image labeling module is used to provide instruction for lightweight neural network
Practice the markup information of data set;Model training module is used for the training lightweight neural network in multinuclear heterogeneous Computing module
Model;Algorithm application module utilizes lightweight god towards electric power intranet and extranet graphic service in multinuclear heterogeneous Computing module
Intellectual analysis task is performed through network model.
It is visited between business application module and multinuclear heterogeneous Computing module by web service (network service) interface
It asks, transmits data.
Other business platform (electric power letters such as unified video monitoring platform and fortune inspection, scheduling, marketing, capital construction of power industry
Breathization platform) communicated by web service with image management module, obtain the image handled through this system and video data and
Its related information;By web service and power industry, other business platforms communicate image management module, obtain electric system
Original image and video data and its related information, the data source as subsequent analysis;Image management module also receives local
The image and video data of upload, effective supplement as subsequent analysis data source.
(1) CPU storage managements node
CPU storage management nodes embed relevant database and non-relational database.In this system, image and video counts
According to for unstructured data, the related information of image and video is structural data, these related informations include image and video
Attribute.The unstructured data being mentioned below is equivalent to image and video data, and structural data is equivalent to image and video
Related information.Relevant database is used for the attribute of structured data, i.e. image and video, and non-relational database is used
In storage unstructured data, i.e. image and video.Image and video data in this system is not of uniform size, small then tens KB,
MB big then up to a hundred, need a database that storage unit size is supported to cross over larger section, and this system selects HBase databases
(building the distributed column storage system on HDFS) stores.Relevant database selects a mySQL database (relationship type number
According to base management system).The related information of same image or video is contacted by unique ID.
(2) GPU calculate nodes
GPU calculate nodes include the GPU of grouping, and every group includes 2 GPU, to complete parallel computation.This system is using in group
GPU does the mode that data parallel is done between model parallel computation, group, and two GPU respectively hold lightweight neural network model in group
Half parameter, cooperation completes the training of single model.
(3) CPU calculate nodes
CPU calculate nodes include the CPU of grouping, and every group includes 2 CPU, to complete parallel computation.This system uses 2 groups
CPU calculate nodes complete scientific algorithm, such as retrieval, lookup, and in addition 2 groups of CPU are effective supplements of GPU calculate nodes, are used for
Parallel computation is completed, in a manner that CPU in group does and does data parallel between model parallel computation, group, two CPU are respectively held in group
There is the half parameter of lightweight neural network model, the training of single model is completed in cooperation.
The physical node quantity that " multinuclear " is mainly reflected in Computational frame is expansible, and " isomery ", which is mainly reflected in, is
The node of system is made of GPU and CPU two categories, and " parallel " is mainly reflected in data parallel, model parallel computation and appoints
Business parallel computation.
For this system, data parallel and model parallel computation can realize perfect efficient data transmission.With
For model training, it is assumed that there are 4 GPU calculate nodes to complete model training task, this 4 nodes are divided into two groups, respectively
Group 1 and group 2 include node GPU1 and GPU2, node GPU3 and GPU4 are included in group 2 in organizing 1.The model of lightweight neural network
Parallel computation and data parallel are established in packet by packet basis, and GPU does model parallel computation in group, and GPU does data simultaneously between group
Row calculates.Two nodes respectively hold the half parameter of lightweight neural network model in group, and the training of single model is completed in cooperation, i.e.,
GPU1 and GPU2 holds the half parameter of model, and GPU3 and GPU4 hold the other half parameter of model, this process is called model
Parallel computation.Data parallel is trained by synchronous stochastic gradient descent between group, and completing parameter using topology exchanges, i.e. group
GPU1 and GPU2 in 1 complete data exchange with the GPU3 in group 2 and GPU4, this process is called data parallel.
After introducing data parallel and model parallel computation, training data is read from disk, training data pre-processes, gently
Magnitude neural metwork training occupies disk, CPU, GPU resource respectively, and takes larger.Therefore assembly line is introduced so that magnetic
Disk, CPU, GPU resource can be utilized simultaneously, promote overall performance.
Perform the parallel procedure of the parallel procedure analog model training of intelligent task.Tasks in parallel calculating then refers in task pipe
Reason level, multiple intelligent tasks can be presented as data parallel by scheduling on demand, concurrently perform specific to each calculate node
With model parallel computation.
Electric power image data has the professional of its own, and the information categories such as equipment therein, personnel, scene are not deposited usually
It is that the common large data such as ImageNet, Pascal VOC is concentrated.Electric power graphic service also has that itself is professional, is related to list
The plurality of classes such as label image classification, multi-tag image classification, target detection, and it is more demanding to system real time.For with
On business demand, this system proposes a kind of neural network of lightweight, which can be adapted to the calculation of multiple business needs
Method, while model calculation amount is moderate, fast response time, precision is higher, there is very strong practical value.
The present invention proposes a kind of algorithm core of lightweight neural network as system, and main feature and advantage have:
(1) whole network includes 20 layers, and wherein convolutional layer has 17 layers, and pond layer has 1 layer, and full articulamentum has 1 layer, convolutional layer
Positioned at network front end, pond layer and full articulamentum are located at network backend.Network structure is moderate compared with the deep but number of plies, can effectively improve
Classification accuracy.
(2) convolution kernel used is 1 × 1 small convolution kernel and 1 × 3,3 × 1 asymmetric convolution kernel, wherein 1 × 3,3 × 1
Asymmetric convolution kernel can play the equivalent effects of 3 × 3 convolution kernels, but parameter greatly reduces, and is keeping network depth
Increase the non-linear of network simultaneously, reduce calculation amount and number of parameters.
(3) have residual error structure, accelerate convergence while network depth is kept, effectively prevent deep neural network
Gradient disappearance problem.
(4) network have abstention, batch standardization can arrangement parameter, introduce regularization mechanism, accelerate training speed
Over-fitting can be mitigated simultaneously, enhance network generalization.
(5) network is provided for the normalization of input data and enhancing function, all input pictures are normalized to 224 ×
224 pixel value, and provide a variety of data enhancements such as mirror image reversal, cutting, tone reversal, expand training dataset,
Over-fitting can be mitigated, enhance network generalization.
(6) network last one layer i.e. loss function layer have can suitability, using softmax functions as loss function
With single labeling, classified using entropy function is intersected as multi-tag, meet a variety of power business demands.
Compared to other neural networks, the network model number of plies that this system proposes is moderate, keeps carrying while accuracy rate
Training speed and algorithm execution speed are risen, training used time and algorithm performs used time are less than the common volumes such as ResNet, GoogLeNet
Product neural network.
The lightweight neural network framework that this system utilizes is as follows:
The lightweight neural network framework that this system utilizes is:
1st layer is convolutional layer, and step-length 2, the size of input is 224 × 224 × 3, using two cascade convolution kernels 1 ×
3 and 3 × 1;
2nd layer is convolutional layer, and step-length 1, the size of input is 112 × 112 × 32, using two cascade convolution kernels 1
× 3 and 3 × 1;
3rd layer is convolutional layer, and step-length 1, the size of input is 112 × 112 × 32, using 1 × 1 convolution kernel;
4th layer is convolutional layer, and step-length 2, the size of input is 112 × 112 × 64, using two cascade convolution kernels 1
× 3 and 3 × 1;
5th layer is convolutional layer, and step-length 1, the size of input is 56 × 56 × 64, using two cascade convolution kernels 1 × 3
With 3 × 1;
6th layer is convolutional layer, and step-length 1, the size of input is 56 × 56 × 128, using 1 × 1 convolution kernel;
7th layer is convolutional layer, and step-length 2, the size of input is 56 × 56 × 128, using two cascade convolution kernels 1 ×
3 and 3 × 1;
8th layer is convolutional layer, and step-length 1, the size of input is 28 × 28 × 128, using two cascade convolution kernels 1 ×
3 and 3 × 1;
9th layer is convolutional layer, and step-length 1, the size of input is 28 × 28 × 256, using 1 × 1 convolution kernel;
10th layer is convolutional layer, and step-length 2, the size of input is 28 × 28 × 256, using two cascade convolution kernels 1
× 3 and 3 × 1;
11th layer is convolutional layer, and step-length 1, the size of input is 14 × 14 × 256, using two cascade convolution kernels 1
× 3 and 3 × 1;
12nd layer is convolutional layer, and step-length 1, the size of input is 14 × 14 × 512, using 1 × 1 convolution kernel;
13rd layer is convolutional layer, and step-length 2, the size of input is 14 × 14 × 512, using two cascade convolution kernels 1
× 3 and 3 × 1;
14th layer is convolutional layer, and step-length 1, the size of input is 7 × 7 × 512, using two cascade convolution kernels 1 × 3
With 3 × 1;
15th layer is convolutional layer, and step-length 1, the size of input is 7 × 7 × 1024, using 1 × 1 convolution kernel;
16th layer is convolutional layer, and step-length 2, the size of input is 7 × 7 × 1024, using two cascade convolution kernels 1 ×
3 and 3 × 1;
17th layer is convolutional layer, and step-length 1, the size of input is 7 × 7 × 1024, using two cascade convolution kernels 1 ×
3 and 3 × 1;
18th layer is average pond layer, and step-length 1, the size of input is 7 × 7 × 1024, and pond size is 7 × 7;
19th layer is full articulamentum, and the size of input is 1 × 1 × 1024, includes 1000 neurons;
20th layer is loss function layer, can be adapted to, and using softmax functions as loss function with single labeling, is made
Classified by the use of entropy function is intersected as multi-tag.
The training of lightweight neural network uses the training method of transfer learning, is first instructed on large database ImageNet
Pre-training model is got, training is then being finely adjusted based on the image inside this system and video data, is obtaining lightweight
Rapid image categorization model.The image classification category of model rate of accuracy reached 93.19% that test training finishes meets practical electric power life
Produce the demand of scene.
Image management module, for managing electric power intranet and extranet graphic service, service-oriented is provided on electric power outer net file
The basic functions such as biography, the importing of electric power Intranet file, document retrieval, file download, statistics displaying.
File can utilize the association letter of the attribute of file, i.e. image and video when uploading, import, retrieving, downloading, showing
Breath.The attribute of file includes view classification, device tree, scene tree, tag tree, defect tree, many types, document source etc..Depending on
Figure classification includes the headend equipment address of two class of image and video, device tree description shooting image or video, and scene tree description is clapped
The particular content of the electric power scene of image or video, tag tree description shooting image or video is taken the photograph, is formulated by system manager, is lacked
Sunken tree is option, and defect problem existing for description shooting image or video is formulated by system manager, and many types description is clapped
The power specialty title of image or video is taken the photograph, such as transmission line of electricity, substation, current conversion station, computer room, capital construction center, business hall, file
Carry out the means of Source Description shooting figure picture or video, such as unmanned plane shooting, fixing camera shooting, robot shooting, handheld terminal
Shooting etc..
The electric power intranet and extranet graphic service of this system processing presses following grade classification:
Level-one task includes image general utility functions task and graphic service application task;
Second task include image general utility functions task under image duplicate removal, low-quality image rejecting, video code conversion, video
Task towards fortune inspection under compression, fast browsing and graphic service application task, the task towards scheduling, towards capital construction
Task, the task of marketing oriented;
Three-level task include towards under the task of fortune inspection working truck detection, wire foreign matter detection, tree bamboo detection of the growth,
Wire icing detection, pyrotechnics detection, gold utensil rust detection, gold utensil loss detection, insulator Safety check-up, insulator lose inspection
Survey, insulator contamination detection, substation's meter number identification, the inspection of installation for transformer Oil Leakage Detecting, transformer station personnel abnormal behaviour
Survey, the detection of transformer station personnel dressing specification, transformer station personnel discrepancy detection, pyrotechnics detection etc.;Change towards under the task of scheduling
The identification of depressor on off state and disconnecting switch state recognition etc.;Towards inlet and outlet vehicle detection, the inlet and outlet under the task of capital construction
Car license recognition, inlet and outlet personnel detection, inlet and outlet personnel unusual checking, the detection of personnel's dressing specification, open fire detection etc.;Face
Business environmental quality under the task of marketing detects, attendant to hilllock leaves the post situation detection, attendant's appearance instrument matter
Measure detection, attendant works' behavior quality testing, customer behavior analysis and anomalous identification etc..
Image labeling module provides automanual image labeling tool, for providing trained number for lightweight neural network
According to the markup information of collection.Image labeling is divided into two kinds, and one kind is system automatic marking, i.e. input picture to lightweight neural network
Model, using the classification results of model output as markup information.Another kind is that user knows icon note, i.e., user checks image data,
Mark out the object classification included in image and region manually.User knows effective benefit of the icon note as system automatic marking
It fills, the inaccurate information of system automatic marking can be filtered out.
Model training module, the training lightweight god in the GPU calculate nodes in multinuclear heterogeneous Computing module
Through network model.Using the image and video data by image labeling as data set, towards image classification, target detection, image
The machine learning tasks such as segmentation carry out training in multinuclear heterogeneous Computing module.
It can suspend in training process, cancel, continue training mission.If training result is not ideal enough, network ginseng is changed
Number, iterations continue to train, and until obtaining satisfied model, satisfied model is stored in GPU calculate nodes, replaceable former
Some models.
System supports on-line training and off-line training two ways.Off-line training refers to disposably input all data
Model training task is opened after system.After on-line training refers to model training task start, new data can be inputted to system, added in
Just continue to train after the model of iteration.
Algorithm application module, towards electric power intranet and extranet graphic service, the GPU in multinuclear heterogeneous Computing module is calculated
On node intellectual analysis task is performed using lightweight neural network model.System is provided towards electric power scene, equipment, people
Member, many algorithms model of event category, to be adapted to different intellectual analysis tasks.Specific task names are shown in image management mould
Block.
The algorithm model of intellectual analysis task is performed there are two source, first, the lightweight nerve that model training module obtains
Network model, second is that being built in other maturity models of GPU calculate nodes.
It can suspend, cancel, continuing training mission during tasks carrying, task can delete task after terminating.Appoint
After business is finished, user can according to results operation image and video data, based on function mentioned above, including but it is unlimited
In deleting the image and low-quality image of repetition, compress long video, the video of transcoding different-format, to the target in image point
Class, the target in detection and segmentation image etc..
This system has scalability and universality:
The scalability of system is presented as:
1. system level, power industry other business platform (unified video monitoring platform and fortune inspection, scheduling, marketing, bases
Build and wait power informatizations platforms) communicated by web service with image management module, obtain the image that is handled through this system with
Video data and its related information;By web service and power industry, other business platforms communicate image management module, obtain
Take the original image of electric system and video data and its related information, the data source as subsequent analysis;Image management module
Also the image and video data locally uploaded is received, effective supplement as subsequent analysis data source.
2. hardware view, system is based on multinuclear heterogeneous Computing frame, and different cpu nodes or GPU nodes can access
System realizes multi-core parallel concurrent operation and data efficient storage, completes the training of offline or on-time model and intelligent task performs function.
3. software view, a kind of lightweight neural network is devised, the network model in algorithm is expansible, according to different
Task characteristic is adapted to different neural network models.
The universality of system is presented as:
1. data are applicable in aspect, the structural data of platform internal memory storage is according to the quasi- Unified coding form of state's network mark, data
Form can be transmitted without conversion between power industry business platform;
2. algorithm be applicable in aspect, system provide towards electric power scene, equipment, personnel, event category many algorithms, can
To be applicable in a variety of electric power analysis occasions.
This system novelty is mainly reflected in:
1. the multinuclear heterogeneous Computing frame of suitable this system is employed, wherein the division of labor of each node is clear and definite, clear logic,
Interaction is reasonable.GPU calculate nodes undertake model training and intelligent task performs function, complete intensive calculations.CPU storage management knots
Point undertakes data storage function, embeds a kind of relevant database and a kind of non-relational database, CPU calculate nodes undertake section
Computing function is learned, while GPU calculate nodes is assisted to complete part intensive calculations function.
2. proposing a kind of algorithm core of lightweight neural network as system, network includes 20 layers, employs rouleau
Product core and asymmetric convolution nuclear structure, residual error structure, abstention mechanism, batch standardized mechanism, can effectively improve classification accuracy, together
When accelerate convergence, enhance network generalization, keep accuracy rate while improve training speed and algorithm execution speed.Simultaneously
Network is provided for the normalization of input data and enhancing function, can effective EDS extended data set.Lightweight rapid image categorization mould
Type is based on existing electric power data, and parameter is got based on electric power data training, is not based on general data set, has
Innovative and practical value.
3. system has scalability and universality, the data transfer of other business platforms of support and power industry is supported
The GPU nodes of multinuclear isomery and cpu node parallel computation, have expansible algorithm network model, and suitable power scene is set
Standby, personnel, a variety of electric power analysis occasions of event category.
This system is based on multinuclear heterogeneous Computing frame, can efficiently realize the training of online or offline image data, be formed
Lightweight rapid image categorization model can maximize the valency for playing image data, it can be achieved that electric power intranet and extranet graphic service application
Value, has preferable application prospect.
The above is only the preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art
For member, without departing from the technical principles of the invention, several improvement and deformation can also be made, these are improved and deformation
Also it should be regarded as protection scope of the present invention.
Claims (10)
1. a kind of electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing, which is characterized in that with lightweight
Neural network is core, including multinuclear heterogeneous Computing module and business application module;Business application module and multinuclear isomery
It is accessed between parallel computation module by network service interface, transmits data;
The multinuclear heterogeneous Computing module includes GPU calculate nodes, CPU storage managements node, CPU calculate nodes, each to save
It is connected between point by interchanger;GPU calculate nodes perform for model training and intelligent task, complete intensive calculations;CPU is deposited
It stores up management node to store for data, embeds relevant database and non-relational database;CPU calculate nodes are based on science
It calculates, while GPU calculate nodes is assisted to complete part intensive calculations;
Business application module includes image management module, image labeling module, model training module, algorithm application module;Image
Management module is used to manage electric power intranet and extranet graphic service;Image labeling module is used to provide trained number for lightweight neural network
According to the markup information of collection;Model training module is used for the training lightweight neural network mould in multinuclear heterogeneous Computing module
Type;Algorithm application module utilizes lightweight nerve towards electric power intranet and extranet graphic service in multinuclear heterogeneous Computing module
Network model performs intellectual analysis task.
2. a kind of electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing according to claim 1,
It is characterized in that:By web service and power industry, other business platforms communicate described image management module, power industry its
The image and video data and its related information that his business platform acquisition is handled through system;Image management module obtains electric system
Original image and video data and its related information;Image management module also receives the image and video data locally uploaded;
Other business platforms of the power industry include unified video monitoring platform and fortune inspection, scheduling, marketing, capital construction power informatization
Platform.
3. a kind of electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing according to claim 2,
It is characterized in that:The related information includes the attribute of image and video, including view classification, device tree, scene tree, tag tree,
Defect tree, many types, document source;
The view classification includes two class of image and video, and the device tree description is with shooting the headend equipment of image or video
The electric power scene of location, scene tree description shooting image or video, the tag tree description shooting image or video it is specific
Content is formulated by system manager, and the defect tree is option, defect problem existing for description shooting image or video, by
System manager formulates, the power specialty title of many types description shooting image or video, the document source description
Shoot the means of image or video.
4. a kind of electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing according to claim 1,
It is characterized in that:The relevant database is used to store the attribute of image and video, and the non-relational database is used to store
Image and video data.
5. a kind of electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing according to claim 1,
It is characterized in that:The lightweight neural network model, specifically includes:
Whole network includes 20 layers, and wherein convolutional layer has 17 layers, and pond layer has 1 layer, and full articulamentum has 1 layer, and convolutional layer is located at net
Network front end, pond layer and full articulamentum are located at network backend;
The convolution kernel used is 1 × 1 small convolution kernel and 1 × 3,3 × 1 asymmetric convolution kernel;
Have residual error structure, accelerate convergence while network depth is kept;
Network have abstention, batch standardization can arrangement parameter, introduce regularization mechanism;
Network provides the normalization for being directed to input data and enhancing function, and all input pictures are normalized to 224 × 224 picture
Element value, and mirror image reversal, cutting, a variety of data enhancements of tone reversal are provided, expand training dataset;
Network last one layer i.e. loss function layer have can suitability, using softmax functions as loss function with single mark
Label classification is classified using entropy function is intersected as multi-tag.
6. a kind of electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing according to claim 1,
It is characterized in that:The lightweight neural network framework is:
1st layer is convolutional layer, and step-length 2, the size of input is 224 × 224 × 3, using two cascade convolution kernels 1 × 3 and 3
×1;
2nd layer is convolutional layer, and step-length 1, the size of input is 112 × 112 × 32, using two cascade 1 × 3 Hes of convolution kernel
3×1;
3rd layer is convolutional layer, and step-length 1, the size of input is 112 × 112 × 32, using 1 × 1 convolution kernel;
4th layer is convolutional layer, and step-length 2, the size of input is 112 × 112 × 64, using two cascade 1 × 3 Hes of convolution kernel
3×1;
5th layer is convolutional layer, and step-length 1, the size of input is 56 × 56 × 64, using two cascade convolution kernels 1 × 3 and 3
×1;
6th layer is convolutional layer, and step-length 1, the size of input is 56 × 56 × 128, using 1 × 1 convolution kernel;
7th layer is convolutional layer, and step-length 2, the size of input is 56 × 56 × 128, using two cascade convolution kernels 1 × 3 and 3
×1;
8th layer is convolutional layer, and step-length 1, the size of input is 28 × 28 × 128, using two cascade convolution kernels 1 × 3 and 3
×1;
9th layer is convolutional layer, and step-length 1, the size of input is 28 × 28 × 256, using 1 × 1 convolution kernel;
10th layer is convolutional layer, and step-length 2, the size of input is 28 × 28 × 256, using two cascade 1 × 3 Hes of convolution kernel
3×1;
11th layer is convolutional layer, and step-length 1, the size of input is 14 × 14 × 256, using two cascade 1 × 3 Hes of convolution kernel
3×1;
12nd layer is convolutional layer, and step-length 1, the size of input is 14 × 14 × 512, using 1 × 1 convolution kernel;
13rd layer is convolutional layer, and step-length 2, the size of input is 14 × 14 × 512, using two cascade 1 × 3 Hes of convolution kernel
3×1;
14th layer is convolutional layer, and step-length 1, the size of input is 7 × 7 × 512, using two cascade convolution kernels 1 × 3 and 3
×1;
15th layer is convolutional layer, and step-length 1, the size of input is 7 × 7 × 1024, using 1 × 1 convolution kernel;
16th layer is convolutional layer, and step-length 2, the size of input is 7 × 7 × 1024, using two cascade convolution kernels 1 × 3 and 3
×1;
17th layer is convolutional layer, and step-length 1, the size of input is 7 × 7 × 1024, using two cascade convolution kernels 1 × 3 and 3
×1;
18th layer is average pond layer, and step-length 1, the size of input is 7 × 7 × 1024, and pond size is 7 × 7;
19th layer is full articulamentum, and the size of input is 1 × 1 × 1024, includes 1000 neurons;
20th layer is loss function layer, can be adapted to, and using softmax functions as loss function with single labeling, uses friendship
Entropy function is pitched as multi-tag to classify.
7. a kind of electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing according to claim 1,
It is characterized in that:The electric power intranet and extranet graphic service presses following grade classification:
Level-one task includes image general utility functions task and graphic service application task;
Second task includes image duplicate removal, low-quality image rejecting, video code conversion, video pressure under image general utility functions task
Task towards fortune inspection, the task towards scheduling, appointing towards capital construction under contracting, fast browsing and graphic service application task
Business, the task of marketing oriented;
Three-level task is included towards the working truck under the task of fortune inspection detects, wire foreign matter detects, tree bamboo detection of the growth, conducting wire
Icing detection, pyrotechnics detection, gold utensil rust detection, gold utensil loss detection, insulator Safety check-up, insulator loss detection, absolutely
The detection of edge pollution flashover, installation for transformer Oil Leakage Detecting, transformer station personnel unusual checking, becomes substation's meter number identification
The detection of power station personnel's dressing specification, transformer station personnel discrepancy detection, pyrotechnics detection;Transformer switch towards under the task of scheduling
State recognition and disconnecting switch state recognition;Towards under the task of capital construction inlet and outlet vehicle detection, inlet and outlet Car license recognition, into
Outlet personnel detection, inlet and outlet personnel unusual checking, the detection of personnel's dressing specification, open fire detection;The task of marketing oriented
Under business environmental quality detection, attendant to hilllock leave the post situation detection, the quality testing of attendant's appearance instrument, service people
The quality testing of member's work behavior, customer behavior analysis and anomalous identification.
8. a kind of electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing according to claim 1,
It is characterized in that:Described image mark is divided into two kinds, and one kind is system automatic marking, i.e. input picture to lightweight neural network mould
Type, using the classification results of model output as markup information;Another kind is that user knows icon note, i.e., user checks image data, hand
It is dynamic to mark out the object classification included in image and region.
9. a kind of electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing according to claim 1,
It is characterized in that:The model training module using by image labeling image and video data as data set, towards image classification,
Target detection, the machine learning task of image segmentation, carry out training in multinuclear heterogeneous Computing module;It can in training process
Pause, continues training mission at revocation;If training result is not ideal enough, network parameter is changed, iterations continue to train, directly
To satisfied model is obtained, satisfied model is stored in GPU calculate nodes, replaceable original model;Support on-line training
With off-line training two ways;Off-line training refers to that model training task will be opened after the disposable input system of all data;
After line training refers to model training task start, new data can be inputted to system, addition just continues to instruct after the model of iteration
Practice.
10. a kind of electric power artificial intelligence visual analysis system based on multinuclear heterogeneous Computing according to claim 1,
It is characterized in that:The algorithm model of the algorithm application module is there are two source, first, the lightweight god that model training module obtains
Through network model, second is that being built in other maturity models of GPU calculate nodes;It can suspend, cancel, continuing during tasks carrying
Training mission, task can delete task after terminating;After tasks carrying, user can be according to results operation image and video
Data.
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711268417.6A CN108171117B (en) | 2017-12-05 | 2017-12-05 | Electric power artificial intelligence visual analysis system based on multicore heterogeneous Computing |
PCT/CN2018/114389 WO2019109771A1 (en) | 2017-12-05 | 2018-11-07 | Power artificial-intelligence visual-analysis system on basis of multi-core heterogeneous parallel computing |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711268417.6A CN108171117B (en) | 2017-12-05 | 2017-12-05 | Electric power artificial intelligence visual analysis system based on multicore heterogeneous Computing |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108171117A true CN108171117A (en) | 2018-06-15 |
CN108171117B CN108171117B (en) | 2019-05-21 |
Family
ID=62524353
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711268417.6A Active CN108171117B (en) | 2017-12-05 | 2017-12-05 | Electric power artificial intelligence visual analysis system based on multicore heterogeneous Computing |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN108171117B (en) |
WO (1) | WO2019109771A1 (en) |
Cited By (28)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109255345A (en) * | 2018-08-21 | 2019-01-22 | 国网浙江省电力有限公司电力科学研究院 | A kind of cable tunnel iron rust recognition methods based on convolutional neural networks |
CN109727376A (en) * | 2018-12-29 | 2019-05-07 | 北京沃东天骏信息技术有限公司 | Generate the method, apparatus and selling apparatus of configuration file |
CN109829542A (en) * | 2019-01-29 | 2019-05-31 | 武汉星巡智能科技有限公司 | Polynary depth network model reconstructing method and device based on multi-core processor |
WO2019109771A1 (en) * | 2017-12-05 | 2019-06-13 | 南京南瑞信息通信科技有限公司 | Power artificial-intelligence visual-analysis system on basis of multi-core heterogeneous parallel computing |
CN109947573A (en) * | 2019-03-26 | 2019-06-28 | 北京智芯微电子科技有限公司 | Intelligence suitable for electric system edge calculations accelerates chip |
CN110175571A (en) * | 2019-05-28 | 2019-08-27 | 华翔翔能电气股份有限公司 | The intellectual monitoring of substation equipment state and recognition methods |
CN110705425A (en) * | 2019-09-25 | 2020-01-17 | 广州西思数字科技有限公司 | Tongue picture multi-label classification learning method based on graph convolution network |
CN110717574A (en) * | 2018-07-11 | 2020-01-21 | 杭州海康威视数字技术股份有限公司 | Neural network operation method and device and heterogeneous intelligent chip |
CN111198760A (en) * | 2018-11-20 | 2020-05-26 | 北京搜狗科技发展有限公司 | Data processing method and device |
CN111260553A (en) * | 2020-01-13 | 2020-06-09 | 哈尔滨工程大学 | Domestic vision computing system based on remote lossless video transmission |
CN111582073A (en) * | 2020-04-23 | 2020-08-25 | 浙江大学 | Transformer substation violation identification method based on ResNet101 characteristic pyramid |
CN111738084A (en) * | 2020-05-21 | 2020-10-02 | 山东大学 | Real-time target detection method and system based on CPU-GPU heterogeneous multiprocessor system on chip |
CN111931713A (en) * | 2020-09-21 | 2020-11-13 | 成都睿沿科技有限公司 | Abnormal behavior detection method and device, electronic equipment and storage medium |
CN112101266A (en) * | 2020-09-25 | 2020-12-18 | 重庆电政信息科技有限公司 | Multi-ARM-based distributed inference method for action recognition model |
CN112164446A (en) * | 2020-10-13 | 2021-01-01 | 电子科技大学 | Medical image report generation method based on multi-network fusion |
CN112528983A (en) * | 2020-12-16 | 2021-03-19 | 国网江苏省电力有限公司检修分公司 | GIS isolation/grounding switch video image acquisition system under dark light condition |
WO2021063026A1 (en) * | 2019-09-30 | 2021-04-08 | 华为技术有限公司 | Inference service networking method and apparatus |
CN112784989A (en) * | 2019-11-08 | 2021-05-11 | 阿里巴巴集团控股有限公司 | Inference system, inference method, electronic device, and computer storage medium |
CN113096002A (en) * | 2021-04-02 | 2021-07-09 | 联捷计算科技(深圳)有限公司 | Video frame scale transformation heterogeneous system and method and storage medium |
CN113222134A (en) * | 2021-07-12 | 2021-08-06 | 深圳市永达电子信息股份有限公司 | Brain-like computing system, method and computer readable storage medium |
CN113225537A (en) * | 2021-05-10 | 2021-08-06 | 合肥中科类脑智能技术有限公司 | Video image analysis system and method |
CN113515829A (en) * | 2021-05-21 | 2021-10-19 | 华北电力大学(保定) | Situation sensing method for transmission line hardware defects under extremely cold disasters |
CN113590321A (en) * | 2021-07-30 | 2021-11-02 | 西安电子科技大学 | Task configuration method for heterogeneous distributed machine learning cluster |
CN113762120A (en) * | 2021-08-27 | 2021-12-07 | 南京南瑞信息通信科技有限公司 | Insulator image segmentation method and device, electronic equipment and storage medium |
CN114757307A (en) * | 2022-06-14 | 2022-07-15 | 中国电力科学研究院有限公司 | Artificial intelligence automatic training method, system, device and storage medium |
CN114915665A (en) * | 2022-07-13 | 2022-08-16 | 香港中文大学(深圳) | Heterogeneous task scheduling method based on hierarchical strategy |
CN115114963A (en) * | 2021-09-24 | 2022-09-27 | 中国劳动关系学院 | Intelligent streaming media video big data analysis method based on convolutional neural network |
WO2023015500A1 (en) * | 2021-08-11 | 2023-02-16 | Baidu.Com Times Technology (Beijing) Co., Ltd. | Multiple-model heterogeneous computing |
Families Citing this family (39)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110334642A (en) * | 2019-07-01 | 2019-10-15 | 河南牧业经济学院 | The machine vision recognition method and system of one boar behavior |
CN110658725B (en) * | 2019-08-19 | 2023-05-02 | 周口师范学院 | Energy supervision and prediction system and method based on artificial intelligence |
CN110766604B (en) * | 2019-09-04 | 2023-06-06 | 国网浙江省电力有限公司湖州供电公司 | Expansion method for image data of double-column horizontal rotary isolating switch |
CN110910440B (en) * | 2019-09-30 | 2023-04-07 | 中国电力科学研究院有限公司 | Power transmission line length determination method and system based on power image data |
CN111027388A (en) * | 2019-11-12 | 2020-04-17 | 国网天津市电力公司 | Method and system for monitoring safety operation behaviors of constructors based on image recognition |
CN111126138A (en) * | 2019-11-18 | 2020-05-08 | 施博凯 | AI image recognition method for garbage classification |
CN113780312B (en) * | 2019-11-21 | 2024-04-12 | 同济大学 | Highway road surface condition detecting system |
CN111126196A (en) * | 2019-12-10 | 2020-05-08 | 安徽银河物联通信技术有限公司 | Equipment oil leakage detection method |
CN111079645A (en) * | 2019-12-16 | 2020-04-28 | 国网重庆市电力公司永川供电分公司 | Insulator self-explosion identification method based on AlexNet network |
CN111080652B (en) * | 2019-12-23 | 2023-03-10 | 西安电子科技大学 | Optical remote sensing image segmentation method based on multi-scale lightweight cavity convolution |
CN111159095B (en) * | 2020-01-02 | 2023-05-12 | 中国航空工业集团公司西安航空计算技术研究所 | Heterogeneous fusion embedded intelligent computing implementation method |
CN111310899B (en) * | 2020-02-19 | 2023-07-11 | 山东大学 | Power defect identification method based on symbiotic relation and small sample learning |
CN111353432B (en) * | 2020-02-28 | 2023-08-01 | 安徽华润金蟾药业股份有限公司 | Rapid clean selection method and system for honeysuckle medicinal materials based on convolutional neural network |
CN111368978B (en) * | 2020-03-02 | 2023-03-24 | 开放智能机器(上海)有限公司 | Precision improving method for offline quantization tool |
CN111339977B (en) * | 2020-03-03 | 2024-09-17 | 河南中光学集团有限公司 | Small target intelligent recognition system based on remote video monitoring and recognition method thereof |
CN111537515B (en) * | 2020-03-31 | 2023-05-12 | 国网辽宁省电力有限公司朝阳供电公司 | Iron tower bolt defect display method and system based on three-dimensional live-action model |
CN111462167A (en) * | 2020-04-21 | 2020-07-28 | 济南浪潮高新科技投资发展有限公司 | Intelligent terminal video analysis algorithm combining edge calculation and deep learning |
CN111858465B (en) * | 2020-06-29 | 2023-06-06 | 西南电子技术研究所(中国电子科技集团公司第十研究所) | Large-scale matrix QR decomposition parallel computing system |
CN111783968B (en) * | 2020-06-30 | 2024-05-31 | 山东信通电子股份有限公司 | Power transmission line monitoring method and system based on cloud edge cooperation |
CN112598142B (en) * | 2020-12-16 | 2024-02-02 | 明阳智慧能源集团股份公司 | Wind turbine maintenance working quality inspection auxiliary method and system |
CN113326873B (en) * | 2021-05-19 | 2024-07-23 | 云南电网有限责任公司电力科学研究院 | Automatic classification method for switching-on and switching-off states of power equipment based on data enhancement |
CN113343392A (en) * | 2021-07-06 | 2021-09-03 | 浙江天铂云科光电股份有限公司 | Intelligent detection and positioning method for oil conservator |
CN113505064B (en) * | 2021-07-07 | 2022-05-17 | 广东电力信息科技有限公司 | Heterogeneous information flow-based electric power big data service system testing method |
CN113657348A (en) * | 2021-08-31 | 2021-11-16 | 江苏中科云墨数字科技有限公司 | Intelligent analysis method and system for operation violation behaviors of transformer substation |
CN114244975A (en) * | 2021-11-15 | 2022-03-25 | 国能大渡河革什扎水电开发有限公司 | Mobile operation process control method and system based on video edge calculation |
CN114187156A (en) * | 2021-12-17 | 2022-03-15 | 江西洪都航空工业集团有限责任公司 | Intelligent recognition method for city management affair component under mobile background |
CN114760168A (en) * | 2022-03-22 | 2022-07-15 | 广东电力通信科技有限公司 | AI intelligent edge computing gateway |
CN114827614B (en) * | 2022-04-18 | 2024-03-22 | 重庆邮电大学 | Method for realizing LCEVC video coding optimization |
CN114760445A (en) * | 2022-04-21 | 2022-07-15 | 北京新童瑞科技有限公司 | Remote control video double-confirmation method for primary electric power equipment |
CN115116005A (en) * | 2022-06-29 | 2022-09-27 | 济南银捷实业有限公司 | Method for closing cash box, training field unmanned algorithm model and judging logic |
CN114937247B (en) * | 2022-07-21 | 2022-11-01 | 四川金信石信息技术有限公司 | Transformer substation monitoring method and system based on deep learning and electronic equipment |
CN115129479A (en) * | 2022-08-24 | 2022-09-30 | 南方电网数字电网研究院有限公司 | Method and device for processing edge data of power chip in partition mode and computer equipment |
CN115240075B (en) * | 2022-09-22 | 2022-12-13 | 山东大学 | Construction and training method of electric power vision multi-granularity pre-training large model |
CN115379286A (en) * | 2022-10-24 | 2022-11-22 | 通号通信信息集团有限公司 | Intelligent video analysis box, intelligent video analysis system and method |
CN115984675B (en) * | 2022-12-01 | 2023-10-13 | 扬州万方科技股份有限公司 | System and method for realizing multipath video decoding and AI intelligent analysis |
CN116309520B (en) * | 2023-04-03 | 2024-09-13 | 江南大学 | Strip steel surface defect detection system |
CN116450486B (en) * | 2023-06-16 | 2023-09-05 | 浪潮电子信息产业股份有限公司 | Modeling method, device, equipment and medium for nodes in multi-element heterogeneous computing system |
CN117389742B (en) * | 2023-11-10 | 2024-05-31 | 深圳市天鹤科技有限公司 | Edge computing method, device and storage medium for machine vision |
CN118334663B (en) * | 2024-06-13 | 2024-08-13 | 杭州宇泛智能科技股份有限公司 | One-stop artificial intelligent image processing model construction method and device |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102741858A (en) * | 2009-10-28 | 2012-10-17 | 数字标记公司 | Sensor-based mobile search, related methods and systems |
CN106504233A (en) * | 2016-10-18 | 2017-03-15 | 国网山东省电力公司电力科学研究院 | Image electric power widget recognition methodss and system are patrolled and examined based on the unmanned plane of Faster R CNN |
CN106527455A (en) * | 2017-01-03 | 2017-03-22 | 北京博瑞空间科技发展有限公司 | UAV landing control method and device |
US20170148431A1 (en) * | 2015-11-25 | 2017-05-25 | Baidu Usa Llc | End-to-end speech recognition |
CN107341127A (en) * | 2017-07-05 | 2017-11-10 | 西安电子科技大学 | Convolutional neural networks accelerated method based on OpenCL standards |
CN107430677A (en) * | 2015-03-20 | 2017-12-01 | 英特尔公司 | Based on the target identification lifted to binary system convolutional neural networks feature |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108171117B (en) * | 2017-12-05 | 2019-05-21 | 南京南瑞信息通信科技有限公司 | Electric power artificial intelligence visual analysis system based on multicore heterogeneous Computing |
-
2017
- 2017-12-05 CN CN201711268417.6A patent/CN108171117B/en active Active
-
2018
- 2018-11-07 WO PCT/CN2018/114389 patent/WO2019109771A1/en active Application Filing
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102741858A (en) * | 2009-10-28 | 2012-10-17 | 数字标记公司 | Sensor-based mobile search, related methods and systems |
CN107430677A (en) * | 2015-03-20 | 2017-12-01 | 英特尔公司 | Based on the target identification lifted to binary system convolutional neural networks feature |
US20170148431A1 (en) * | 2015-11-25 | 2017-05-25 | Baidu Usa Llc | End-to-end speech recognition |
CN106504233A (en) * | 2016-10-18 | 2017-03-15 | 国网山东省电力公司电力科学研究院 | Image electric power widget recognition methodss and system are patrolled and examined based on the unmanned plane of Faster R CNN |
CN106527455A (en) * | 2017-01-03 | 2017-03-22 | 北京博瑞空间科技发展有限公司 | UAV landing control method and device |
CN107341127A (en) * | 2017-07-05 | 2017-11-10 | 西安电子科技大学 | Convolutional neural networks accelerated method based on OpenCL standards |
Non-Patent Citations (1)
Title |
---|
王万国 等: "基于RCNN的无人机巡检图像电力小部件识别研究", 《地球信息科学学报》 * |
Cited By (40)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2019109771A1 (en) * | 2017-12-05 | 2019-06-13 | 南京南瑞信息通信科技有限公司 | Power artificial-intelligence visual-analysis system on basis of multi-core heterogeneous parallel computing |
CN110717574A (en) * | 2018-07-11 | 2020-01-21 | 杭州海康威视数字技术股份有限公司 | Neural network operation method and device and heterogeneous intelligent chip |
CN110717574B (en) * | 2018-07-11 | 2023-07-07 | 杭州海康威视数字技术股份有限公司 | Neural network operation method and device and heterogeneous intelligent chip |
CN109255345A (en) * | 2018-08-21 | 2019-01-22 | 国网浙江省电力有限公司电力科学研究院 | A kind of cable tunnel iron rust recognition methods based on convolutional neural networks |
CN111198760A (en) * | 2018-11-20 | 2020-05-26 | 北京搜狗科技发展有限公司 | Data processing method and device |
CN109727376A (en) * | 2018-12-29 | 2019-05-07 | 北京沃东天骏信息技术有限公司 | Generate the method, apparatus and selling apparatus of configuration file |
CN112784987B (en) * | 2019-01-29 | 2024-01-23 | 武汉星巡智能科技有限公司 | Target nursing method and device based on multistage neural network cascade |
CN109829542B (en) * | 2019-01-29 | 2021-04-16 | 武汉星巡智能科技有限公司 | Multi-core processor-based multi-element deep network model reconstruction method and device |
CN109829542A (en) * | 2019-01-29 | 2019-05-31 | 武汉星巡智能科技有限公司 | Polynary depth network model reconstructing method and device based on multi-core processor |
CN112784987A (en) * | 2019-01-29 | 2021-05-11 | 武汉星巡智能科技有限公司 | Target nursing method and device based on multistage neural network cascade |
CN109947573A (en) * | 2019-03-26 | 2019-06-28 | 北京智芯微电子科技有限公司 | Intelligence suitable for electric system edge calculations accelerates chip |
CN110175571A (en) * | 2019-05-28 | 2019-08-27 | 华翔翔能电气股份有限公司 | The intellectual monitoring of substation equipment state and recognition methods |
CN110705425A (en) * | 2019-09-25 | 2020-01-17 | 广州西思数字科技有限公司 | Tongue picture multi-label classification learning method based on graph convolution network |
CN110705425B (en) * | 2019-09-25 | 2022-06-28 | 广州西思数字科技有限公司 | Tongue picture multi-label classification method based on graph convolution network |
WO2021063026A1 (en) * | 2019-09-30 | 2021-04-08 | 华为技术有限公司 | Inference service networking method and apparatus |
CN112784989B (en) * | 2019-11-08 | 2024-05-03 | 阿里巴巴集团控股有限公司 | Inference system, inference method, electronic device, and computer storage medium |
CN112784989A (en) * | 2019-11-08 | 2021-05-11 | 阿里巴巴集团控股有限公司 | Inference system, inference method, electronic device, and computer storage medium |
CN111260553A (en) * | 2020-01-13 | 2020-06-09 | 哈尔滨工程大学 | Domestic vision computing system based on remote lossless video transmission |
CN111582073A (en) * | 2020-04-23 | 2020-08-25 | 浙江大学 | Transformer substation violation identification method based on ResNet101 characteristic pyramid |
CN111738084A (en) * | 2020-05-21 | 2020-10-02 | 山东大学 | Real-time target detection method and system based on CPU-GPU heterogeneous multiprocessor system on chip |
CN111931713A (en) * | 2020-09-21 | 2020-11-13 | 成都睿沿科技有限公司 | Abnormal behavior detection method and device, electronic equipment and storage medium |
CN111931713B (en) * | 2020-09-21 | 2021-01-29 | 成都睿沿科技有限公司 | Abnormal behavior detection method and device, electronic equipment and storage medium |
CN112101266A (en) * | 2020-09-25 | 2020-12-18 | 重庆电政信息科技有限公司 | Multi-ARM-based distributed inference method for action recognition model |
CN112164446A (en) * | 2020-10-13 | 2021-01-01 | 电子科技大学 | Medical image report generation method based on multi-network fusion |
CN112528983A (en) * | 2020-12-16 | 2021-03-19 | 国网江苏省电力有限公司检修分公司 | GIS isolation/grounding switch video image acquisition system under dark light condition |
CN112528983B (en) * | 2020-12-16 | 2023-12-26 | 国网江苏省电力有限公司检修分公司 | GIS isolation/grounding switch video image acquisition system under dim light condition |
CN113096002A (en) * | 2021-04-02 | 2021-07-09 | 联捷计算科技(深圳)有限公司 | Video frame scale transformation heterogeneous system and method and storage medium |
CN113225537A (en) * | 2021-05-10 | 2021-08-06 | 合肥中科类脑智能技术有限公司 | Video image analysis system and method |
CN113515829A (en) * | 2021-05-21 | 2021-10-19 | 华北电力大学(保定) | Situation sensing method for transmission line hardware defects under extremely cold disasters |
CN113222134A (en) * | 2021-07-12 | 2021-08-06 | 深圳市永达电子信息股份有限公司 | Brain-like computing system, method and computer readable storage medium |
CN113222134B (en) * | 2021-07-12 | 2021-10-26 | 深圳市永达电子信息股份有限公司 | Brain-like computing system, method and computer readable storage medium |
CN113590321B (en) * | 2021-07-30 | 2024-02-27 | 西安电子科技大学 | Task configuration method for heterogeneous distributed machine learning cluster |
CN113590321A (en) * | 2021-07-30 | 2021-11-02 | 西安电子科技大学 | Task configuration method for heterogeneous distributed machine learning cluster |
WO2023015500A1 (en) * | 2021-08-11 | 2023-02-16 | Baidu.Com Times Technology (Beijing) Co., Ltd. | Multiple-model heterogeneous computing |
CN113762120A (en) * | 2021-08-27 | 2021-12-07 | 南京南瑞信息通信科技有限公司 | Insulator image segmentation method and device, electronic equipment and storage medium |
CN115114963A (en) * | 2021-09-24 | 2022-09-27 | 中国劳动关系学院 | Intelligent streaming media video big data analysis method based on convolutional neural network |
CN114757307B (en) * | 2022-06-14 | 2022-09-06 | 中国电力科学研究院有限公司 | Artificial intelligence automatic training method, system, device and storage medium |
CN114757307A (en) * | 2022-06-14 | 2022-07-15 | 中国电力科学研究院有限公司 | Artificial intelligence automatic training method, system, device and storage medium |
CN114915665B (en) * | 2022-07-13 | 2022-10-21 | 香港中文大学(深圳) | Heterogeneous task scheduling method based on hierarchical strategy |
CN114915665A (en) * | 2022-07-13 | 2022-08-16 | 香港中文大学(深圳) | Heterogeneous task scheduling method based on hierarchical strategy |
Also Published As
Publication number | Publication date |
---|---|
CN108171117B (en) | 2019-05-21 |
WO2019109771A1 (en) | 2019-06-13 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108171117B (en) | Electric power artificial intelligence visual analysis system based on multicore heterogeneous Computing | |
CN106022245A (en) | Multi-source remote sensing satellite data parallel processing system and method based on algorithm classification | |
Liu et al. | A proposed radicality evaluation method for design ideas at conceptual design stage | |
CN105469376A (en) | Method and device for determining picture similarity | |
CN108055529A (en) | Electric power unmanned plane and robot graphics' data normalization artificial intelligence analysis's system | |
CN111178177A (en) | Cucumber disease identification method based on convolutional neural network | |
CN113177366B (en) | Comprehensive energy system planning method and device and terminal equipment | |
CN114462623B (en) | Data analysis method, system and platform based on edge calculation | |
CN109784161A (en) | Video monitoring intelligent analysis system and method in a kind of shop | |
CN112749663B (en) | Agricultural fruit maturity detection system based on Internet of things and CCNN model | |
CN115115830A (en) | Improved Transformer-based livestock image instance segmentation method | |
CN111429424A (en) | Heating furnace inlet abnormity identification method based on deep learning | |
CN110633684A (en) | Tobacco purchasing grading system and grading method based on deep learning | |
CN117351473A (en) | Deep learning-based fabric identification method, system, equipment and medium | |
CN112788110A (en) | Product appearance detection method based on cloud edge collaborative model optimization and implementation system thereof | |
CN111723742A (en) | Crowd density analysis method, system and device and computer readable storage medium | |
CN114494916A (en) | Black-neck crane monitoring and tracking method based on YOLO and DeepsORT | |
Jia et al. | Polar-Net: Green fruit instance segmentation in complex orchard environment | |
CN111651890B (en) | Data-driven aluminum electrolysis digital twin factory, control method and system | |
CN113408630A (en) | Transformer substation indicator lamp state identification method | |
CN113469938A (en) | Pipe gallery video analysis method and system based on embedded front-end processing server | |
CN107122472A (en) | Extensive unstructured data extracting method, its system, DDM platform | |
CN115631402B (en) | AI algorithm service platform construction method suitable for intelligent cultivation | |
CN111047731A (en) | AR technology-based telecommunication room inspection method and system | |
CN116189076A (en) | Observation and identification system and method for bird observation station |
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 |