CN107506695A - Video monitoring equipment failure automatic detection method - Google Patents
Video monitoring equipment failure automatic detection method Download PDFInfo
- Publication number
- CN107506695A CN107506695A CN201710629187.5A CN201710629187A CN107506695A CN 107506695 A CN107506695 A CN 107506695A CN 201710629187 A CN201710629187 A CN 201710629187A CN 107506695 A CN107506695 A CN 107506695A
- Authority
- CN
- China
- Prior art keywords
- image
- convolution
- layer
- video monitoring
- monitoring equipment
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
-
- 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
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N7/00—Television systems
- H04N7/18—Closed-circuit television [CCTV] systems, i.e. systems in which the video signal is not broadcast
Abstract
The invention discloses a kind of video monitoring equipment failure automatic detection method, this method comprises the following steps, convolution self-encoding encoder uses the heterogeneous networks number of plies and the feature of concealed nodes number extraction monitoring image, in pond, layer carries out image characteristics extraction, and iterative convolution self-encoding encoder restrains until the accuracy rate of disaggregated model；The feature for the monitoring image that convolutional neural networks extract convolution self-encoding encoder as image classification foundation, so as to realize supervised learning；Convolutional neural networks carry out network model training with parallelization computing framework, after network model training terminates, the model are applied to the assorting process of test image, judges video monitoring equipment damaged condition according to classification results.Video monitoring equipment failure automatic detection method of the present invention, corresponding algorithm improvement is proposed for video monitoring equipment fault detect, so as to propose the method for automatic fast detecting failure.
Description
Technical field
The invention belongs to deep learning algorithm field, more particularly to a kind of video monitoring equipment failure automatic detection method.
Background technology
Traditional video monitoring equipment fault detect is mainly realized by machinery equipment sensor, still, in reality detection
Sensor can have failure in itself, lead to not the failure situation for checking video equipment in time.The present invention is using monitoring image point
The mode of class detects video monitoring equipment failure.Monitoring image classification refers to the screenshot capture reserved during monitoring according to black
It is screen, colour cast, normal and the feature such as block and be divided into 4 classes.It is mainly used for the damage that monitoring device is judged according to the result of classification
Situation.The classification of image needs enough training samples, but because training sample does not have label, and convolutional Neural can not be completed
The foundation of network model.Unsupervised learning can solve this problem.The feature that unsupervised learning provides is convolutional neural networks
Classification provide the foundation of classification.
2006, Hinton was improved to prototype autocoder structure, first with unsupervised successively greedy training algorithm
The pre-training to hidden layer is completed, parameter optimization adjustment then is carried out to whole neutral net with BP algorithm, is effectively improved BP calculations
Method is absorbed in the situation of Local Minimum.2007, Benjio proposed the concept of sparse autocoder, had further deepened without prison
The research that educational inspector practises.2011, Jonathan proposed convolution autocoder, for building convolutional neural networks.Pass through
The research of the neutral net scholar such as Hinton, Bengio and Jonathan, the self-encoding encoder of unsupervised learning are automatic comprising prototype
The models such as encoder, sparse autocoder, noise reduction autocoder, convolution autocoder and RBM.It is unsupervised to compare these
Learning model, common self-encoding encoder can not effectively solve the problems, such as the pondization in view data and albefaction, and can produce when training
Raw substantial amounts of nuisance parameter reduces operation efficiency.
The content of the invention
The technical problem to be solved by the invention is to provide a kind of video monitoring equipment failure automatic detection method, for
Video monitoring equipment fault detect proposes corresponding algorithm improvement, so as to propose the method for automatic fast detecting failure.
The technical solution adopted for the present invention to solve the technical problems is：A kind of video monitoring equipment failure is provided to examine automatically
Survey method, step 1, convolution self-encoding encoder is using the heterogeneous networks number of plies and the feature of concealed nodes number extraction monitoring image, in pond
Change layer and carry out image characteristics extraction, iterative convolution self-encoding encoder restrains until the accuracy rate of disaggregated model；Step 2, convolutional Neural
The feature for the monitoring image that network extracts convolution self-encoding encoder as image classification foundation, so as to realize supervised learning；Step
Rapid three, convolutional neural networks carry out network model training with parallelization computing framework, after network model training terminates, by the mould
Type is applied to the assorting process of test image；Step 4, judge video monitoring equipment damaged condition according to classification results.Convolution god
It is to be combined together parallel through network and convolution self-encoding encoder.Convolution self-encoding encoder is unsupervised Feature Selection Model.
By above-mentioned technical proposal, in the step 2, during the batch convolution of convolutional neural networks and pond, using not
The image classification accuracy rate of convolutional neural networks is analyzed with the convolution core of size and pond region, model accuracy rate
Scope is 79.8%-82.3%.Meanwhile the framework of CUDA concurrent operations is added into training process in convolutional neural networks, it will instruct
Practice speed and improve 1.1-1.2 times.
By above-mentioned technical proposal, convolution self-encoding encoder by input layer, convolutional layer, pond layer, anti-pond layer, warp lamination,
Output layer forms successively, and input layer carries out convolution algorithm with 6 8*8 convolution kernel and obtains convolutional layer；Then convolutional layer is made
For input, downscaled images matrix is operated using 2*2 average pondization；Then the image array after diminution is subjected to being averaged for 2*2
Anti- pondization operation, partial reduction image array；Finally original matrix is gone back using matrix zero padding operation.Propose 5 layers of self-encoding encoder knot
Structure, add the operation of anti-pond layer and warp lamination, greatly sparse network structure, experiment prove the CAE of 5 Rotating fields have compared with
Good feature extraction effect.
By above-mentioned technical proposal, in the training of network model described in step 3, automatically extracted based on convolution self-encoding encoder
The feature of video monitoring image, label is added for training sample.Convolutional neural networks belong to supervised learning network.
By above-mentioned technical proposal, convolutional neural networks by input layer, the first convolutional layer, the first pond layer, the second convolutional layer,
Second pond layer, full articulamentum, output layer form successively.Input layer carries out convolution algorithm with 6 8*8 convolution kernel and rolled up
Lamination, then using convolutional layer as input, downscaled images matrix, so circulation 3 times are operated using 2*2 average pondization, finally
Image array is converted into 120 yuan of one-dimensional vector, one-dimensional vector is input in grader and realizes classification.
By above-mentioned technical proposal, in back-propagation process, convolutional neural networks carry out the training side of network model training
Method uses class gradient descent method, i.e., propagates adjustment weights and threshold value by the algorithm of minimization error, learning rate is in training at the beginning of meeting
The number to begin between (0,1), but when learning rate value is larger, in training the problem of convergence is shaken can occur for network；When
When learning rate is smaller, network convergence can be extremely slow.Select momentum method to change learning rate, in the consistent place of gradient direction, learn
The selection of habit rate (0.65,0.95] between, again be changed into just being changed into negative again from negative from being just changed into negative in gradient direction, such gradient side
To when changing more than 2 times, learning rate is changed according to formula (1)：
Δηt=ρ Δs ηt-1-θ[▽C(ηt-1)]T (1)
ΔηtFor the changing value of learning rate after the t times iteration, Δ ηt-1Represent the changing value of the t-1 times iterative learning rate, t
Represent iterations.Wherein learning rate η ∈ (0,1), initial value takes 0.9.ρ ∈ (0,1) are factor of momentum, determine what learning rate changed
Degree, value is bigger, and learning rate change is smaller, takes 0.8.θ=0.2 is the result that gradient direction change frequency removes 10, is constant, ladder
Degree direction change then needs for 2 times to carry out learning rate adjustment.▽C(ηt-1) it is image array changing value, it is worth to be adjusted with the t-1 times iteration
The result that image array subtracts each other after whole, wherein C represent image array, and T represents transposition.
It so can both avoid convergence point from shaking problem, and also solve the problems, such as that convergence was slow.Convolution kernel initial value is chosen
It is improper easily to produce locally optimal solution.Use micro- batch gradient descent algorithm herein, every time 50 Sample Refreshment gradients of selection and
Sample.It can so avoid disposably initializing all convolution kernels, problem be refined, Decomposition iteration.
The beneficial effect comprise that：Convolutional neural networks are proposed to be supervised with reference to improved convolution self-encoding encoder
Image classification is controlled, parallelization computing framework is carried out in convolutional neural networks training process, so as to propose automatic quick detection event
The method of barrier.
Brief description of the drawings
Below in conjunction with drawings and Examples, the invention will be further described, in accompanying drawing：
Fig. 1 is the flow chart of video monitoring equipment failure automatic detection method of the embodiment of the present invention；
Fig. 2 is CAE structural representations in video monitoring equipment failure automatic detection method of the embodiment of the present invention；
Fig. 3 is the micro- batch gradient descent algorithm flow chart of the embodiment of the present invention；
Fig. 4 is deep layer of embodiment of the present invention CNN basic structure；
Fig. 5 is that concurrent operation efficiency comparative schemes in the embodiment of the present invention.
Embodiment
In order to make the purpose , technical scheme and advantage of the present invention be clearer, it is right below in conjunction with drawings and Examples
The present invention is further elaborated.It should be appreciated that specific embodiment described herein is only to explain the present invention, not
For limiting the present invention.
In the embodiment of the present invention, there is provided a kind of video monitoring equipment failure automatic detection method, as shown in figure 1, step 1,
Convolution self-encoding encoder is using the heterogeneous networks number of plies and the feature of concealed nodes number extraction monitoring image, and in pond, it is special to carry out image for layer
Sign extraction, iterative convolution self-encoding encoder restrain until the accuracy rate of disaggregated model；Step 2, convolutional neural networks are self-editing by convolution
Foundation of the feature of the monitoring image of code device extraction as image classification, so as to realize supervised learning；Step 3, convolutional Neural net
Network carries out network model training with parallelization computing framework, and after network model training terminates, the model is applied into test chart
The assorting process of picture；Step 4, judge video monitoring equipment damaged condition according to classification results.Convolutional neural networks and convolution are certainly
Encoder is to be combined together parallel.Convolution self-encoding encoder is unsupervised Feature Selection Model.
Further, in the step 2, during the batch convolution of convolutional neural networks and pond, different size is used
Convolution core and pond region the image classification accuracy rate of convolutional neural networks is analyzed, the scope of model accuracy rate is
79.8%-82.3%.Meanwhile the framework of CUDA concurrent operations is added into training process in convolutional neural networks, by training speed
Improve 1.1-1.2 times.
Further, as shown in Fig. 2 convolution self-encoding encoder is by input layer, convolutional layer, pond layer, anti-pond layer, deconvolution
Layer, output layer form successively, and input layer carries out convolution algorithm with 6 8*8 convolution kernel and obtains convolutional layer；Then by convolutional layer
As input, downscaled images matrix is operated using 2*2 average pondization；Then the image array after diminution is subjected to the flat of 2*2
Anti- pondization operation, partial reduction image array；Finally original matrix is gone back using matrix zero padding operation.Propose 5 layers of self-encoding encoder
Structure, the operation of anti-pond layer and warp lamination is added, greatly sparse network structure, experiment proves that the CAE of 5 Rotating fields has
Preferable feature extraction effect.
Further, in the training of network model described in step 3, supervised based on the video that convolution self-encoding encoder automatically extracts
The feature of image is controlled, label is added for training sample.Convolutional neural networks belong to supervised learning network.
Further, convolutional neural networks are by input layer, the first convolutional layer, the first pond layer, the second convolutional layer, the second pond
Change layer, full articulamentum, output layer to form successively.Input layer carries out convolution algorithm with 6 8*8 convolution kernel and obtains convolutional layer,
Then using convolutional layer as input, downscaled images matrix, so circulation 3 times are operated using 2*2 average pondization, finally by image
Matrix is converted into 120 yuan of one-dimensional vector, and one-dimensional vector is input in grader and realizes classification.
Further, in back-propagation process, the training method that convolutional neural networks carry out network model training uses
Class gradient descent method, i.e., propagate adjustment weights and threshold value by the algorithm of minimization error, learning rate can be initially in training (0,
1) number between, but when learning rate value is larger, in training the problem of convergence is shaken can occur for network；When learning rate compared with
Hour, network convergence can be extremely slow.As shown in Figure 3.Momentum method is selected to change learning rate, in the consistent place of gradient direction,
Learning rate selection (0.65,0.95] between, again be changed into just being changed into negative again from negative from being just changed into negative in gradient direction, such gradient
When direction change is more than 2 times, learning rate is changed according to formula (1)：
Δηt=ρ Δs ηt-1-θ[▽C(ηt-1)]T (1)
ΔηtFor the changing value of learning rate after the t times iteration, Δ ηt-1Represent the changing value of the t-1 times iterative learning rate, t
Represent iterations.Wherein learning rate η ∈ (0,1), initial value takes 0.9.ρ ∈ (0,1) are factor of momentum, determine what learning rate changed
Degree, value is bigger, and learning rate change is smaller, takes 0.8.θ=0.2 is the result that gradient direction change frequency removes 10, is constant, ladder
Degree direction change then needs for 2 times to carry out learning rate adjustment.▽C(ηt-1) it is image array changing value, it is worth to be adjusted with the t-1 times iteration
The result that image array subtracts each other after whole, wherein C represent image array, and T represents transposition.
It so can both avoid convergence point from shaking problem, and also solve the problems, such as that convergence was slow.Convolution kernel initial value is chosen
It is improper easily to produce locally optimal solution.Use micro- batch gradient descent algorithm herein, every time 50 Sample Refreshment gradients of selection and
Sample.It can so avoid disposably initializing all convolution kernels, problem be refined, Decomposition iteration.
In the preferred embodiment of the present invention, the equipment that the embodiment of the present invention uses is 4 core Intel Core 2.2GHz
I5-5200U CPU and tall and handsome reach Geforce GT 750MGPU.Software environment uses Matlab and JAVA programming realizations, program
Bottom has used CUDA concurrent operation framework.Matlab is mainly used in locating pretreatment image, and image preprocessing is mainly comprising number
According to standardization and image dimension-reduction treatment.The present invention 1000 monitoring images of selection include 4 altogether as training sample, the data set
Individual classification is normal, colour cast, blank screen respectively and blocks respectively and has 250 pictures.100 pictures are therefrom randomly selected as test
Sample, each type respectively select 25.Each data sample of data set be 53*53 RGB image, the value model of pixel value
It is 0 to 255 to enclose, and image data samples save as an one-dimensional vector.
The CNN frameworks that the embodiment of the present invention is used are as shown in figure 4, the parameter setting such as specific convolution kernel size refers to table 1.It is defeated
Enter the RGB figures that picture is defined as 53*53；The image (normal, colour cast, blank screen, blocking) of corresponding 4 types of output.The present invention
The CNN that embodiment is used with reference to LeNet.Parameter setting based on LeNet, the present embodiment is according to real image size by CNN
Parameter setting is adjusted to as shown in table 1.The training accuracy rate stabilization that experiment finally obtains identifies in 82.3%, specifically each classification
Accuracy rate be shown in Table 2, wherein for blank screen recognition efficiency be 100%, be 68% He for colour cast and normal picture discrimination
76.9%, the discrimination for shielded image is 84.4%.
Table 1
Table 2
The present invention is incorporated with CUDA parallelization computing framework, parallelization operation of the present invention while neutral net is realized
Framework realizes that java can not directly invoke CUDA, it is necessary to call CUDA kernel program by JNI using java herein.Under
Introduce the process that JNI calls CUDA in face:
(1) the java classes of the abstract method with native statements are write, it is not necessary to realize, such is mainly the meter of convolution
Calculate the runnable interface that parallelization is provided.
(2) ExtactLib file generateds .h header file is compiled using javah orders, to import in c program.Tool
Body order is javah-jni ExtactLib.
(3) C nation method is realized, writes CUDA kernel programs, imports the ExtactLib.h of generation in a program
File, and c program is stored in using .cu as in the file of suffix name.And the .cu files of generation are compiled into dynamic link library
LibGPU.dll supplies java routine calls.
By the addition of parallelization computing framework, training effectiveness and forecasting efficiency are greatly improved.Compare parallel and string
Capable implementation, efficiency improve 1.1-1.2 times, as shown in Figure 5.
It is of the invention to have done comparative analysis with following classic algorithm in the classification degree of accuracy and in terms of the training time：
(1) statistic histogram：The colour cast and blank screen of image can be by counting color histogram analysis result directly perceived.
(2) BP neural network：Classical three-layer neural network.
(3)GoogLeNet：Google proposes 22 layers of deep neural network structure within 2014.
(4)VGGnet：The network structure that Oxford in 2014 team establishes on ILSVRC2014.
Each method contrast realizes that effect is as shown in table 3：
Table 3
The weights shared structure of convolutional neural networks (CNN) reduces the complexity of network model, reduces the number of weights
Amount.What the advantage showed when the input of network is multidimensional image becomes apparent, and allows input of the image directly as network,
Avoid feature extraction complicated in tional identification algorithm and data reconstruction processes.And CAE is reconstructed using important local feature
Initial data, and all local features of input data share weight matrix, the extraction to characteristics of image has huge advantage.This
Text is that the local shape factor to monitoring image is realized based on improved convolution autocoder.
The present invention for possess a large amount of cameras at present and can not fast positioning damage plant issue, proposition uses depth
The method of habit solves video monitoring equipment fault detect.The method for deep learning proposes convolutional neural networks combination simultaneously
Improved convolution self-encoding encoder is monitored image classification.Parallelization computing frame is proposed in convolutional neural networks training process
Structure, improve the operational efficiency of algorithm.The accuracy rate that 500 times are finally trained in 1000 Image Iteratives is 82.3%.Test one
The time of image is 1.252s.
It should be appreciated that for those of ordinary skills, can according to the above description be improved or converted,
And all these modifications and variations should all belong to the protection domain of appended claims of the present invention.
Claims (6)
1. a kind of video monitoring equipment failure automatic detection method, it is characterised in that step 1, convolution self-encoding encoder use different
The feature of the network number of plies and concealed nodes number extraction monitoring image, in pond, layer carries out image characteristics extraction, and iterative convolution is self-editing
Code device is restrained until the accuracy rate of disaggregated model；Step 2, the monitoring image that convolutional neural networks extract convolution self-encoding encoder
Foundation of the feature as image classification, so as to realize supervised learning；Step 3, convolutional neural networks use parallelization computing frame
Structure carries out network model training, and after network model training terminates, the model is applied to the assorting process of test image；Step
Four, judge video monitoring equipment damaged condition according to classification results.
2. video monitoring equipment failure automatic detection method according to claim 1, it is characterised in that the step 2
In, during the batch convolution of convolutional neural networks and pond, using different size of convolution core and pond region to convolution
The image classification accuracy rate of neutral net is analyzed, and the scope of model accuracy rate is 79.8%-82.3%.
3. video monitoring equipment failure automatic detection method according to claim 1 or 2, it is characterised in that convolution is self-editing
Code device is made up of successively input layer, convolutional layer, pond layer, anti-pond layer, warp lamination, output layer, and input layer uses 6 8*8
Convolution kernel carry out convolution algorithm obtain convolutional layer；Then using convolutional layer as input, operated and reduced using 2*2 average pondization
Image array；Then average anti-pondization that image array after diminution is carried out to 2*2 operates, partial reduction image array；Finally
Original matrix is gone back using matrix zero padding operation.
4. video monitoring equipment failure automatic detection method according to claim 1 or 2, it is characterised in that in step 3
In the network model training, the feature of the video monitoring image automatically extracted based on convolution self-encoding encoder, add for training sample
Tag.
5. video monitoring equipment failure automatic detection method according to claim 1 or 2, it is characterised in that convolutional Neural
Network by input layer, the first convolutional layer, the first pond layer, the second convolutional layer, the second pond layer, full articulamentum, output layer successively
Composition.Input layer carries out convolution algorithm with 6 8*8 convolution kernel and obtains convolutional layer, then using convolutional layer as input, utilizes
2*2 average pondization operation downscaled images matrix, so circulation 3 times, finally by image array be converted into 120 yuan it is one-dimensional to
Amount, one-dimensional vector is input in grader and realizes classification.
6. video monitoring equipment failure automatic detection method according to claim 4, it is characterised in that in backpropagation
Cheng Zhong, the training method that convolutional neural networks carry out network model training uses class gradient descent method, i.e., by minimization error
Algorithm propagates adjustment weights and threshold value, selects momentum method to change learning rate, and in the consistent place of gradient direction, learning rate selection exists
(0.65,0.95] between, it is changed into just being changed into negative again from negative again from being just changed into bearing in gradient direction, such gradient direction change exceedes
At 2 times, learning rate is changed according to formula (1)：
Δηt=ρ Δs ηt-1-θ[▽C(ηt-1)]T (1)
ΔηtFor the changing value of learning rate after the t times iteration, Δ ηt-1The changing value of the t-1 times iterative learning rate is represented, t represents to change
Generation number；ρ ∈ (0,1) are factor of momentum, and θ=0.2 is the result that gradient direction change frequency removes 10, are constant, ▽ C (ηt-1)
For image array changing value, it is worth the result to subtract each other with image array after the t-1 times iteration adjustment, wherein C represents image array,
T represents transposition.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710629187.5A CN107506695A (en) | 2017-07-28 | 2017-07-28 | Video monitoring equipment failure automatic detection method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710629187.5A CN107506695A (en) | 2017-07-28 | 2017-07-28 | Video monitoring equipment failure automatic detection method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107506695A true CN107506695A (en) | 2017-12-22 |
Family
ID=60690287
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710629187.5A Pending CN107506695A (en) | 2017-07-28 | 2017-07-28 | Video monitoring equipment failure automatic detection method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107506695A (en) |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108182452A (en) * | 2017-12-29 | 2018-06-19 | 哈尔滨工业大学（威海） | Aero-engine fault detection method and system based on grouping convolution self-encoding encoder |
| CN108844735A (en) * | 2018-06-22 | 2018-11-20 | 上海电力学院 | Epicyclic gearbox fault detection method based on convolution coder and Min formula distance |
| CN109063854A (en) * | 2018-08-23 | 2018-12-21 | 河南中裕广恒科技股份有限公司 | Intelligent O&M cloud platform system and its control method |
| CN109087259A (en) * | 2018-08-01 | 2018-12-25 | 中国石油大学(北京) | Pre stack data denoising method and system based on convolution self-encoding encoder |
| CN109413411A (en) * | 2018-09-06 | 2019-03-01 | 腾讯数码（天津）有限公司 | A kind of blank screen recognition methods, device and the server of monitoring circuit |
| CN109816625A (en) * | 2018-11-27 | 2019-05-28 | 广东电网有限责任公司 | A kind of video quality score implementation method |
| EP3522079A1 (en) * | 2018-02-01 | 2019-08-07 | Siemens Healthcare Limited | Data encoding and classification |
| CN110210378A (en) * | 2019-05-30 | 2019-09-06 | 中国电子科技集团公司第三十八研究所 | A kind of embedded video method for analyzing image and device based on edge calculations |
| CN110415220A (en) * | 2019-07-09 | 2019-11-05 | 国电大渡河瀑布沟发电有限公司 | A kind of device intelligence method for inspecting of large hydropower station |
| CN110458838A (en) * | 2019-08-23 | 2019-11-15 | 讯飞智元信息科技有限公司 | A kind of detection method of fault type, device, storage medium and equipment |
| CN110472695A (en) * | 2019-08-22 | 2019-11-19 | 李世鹏 | A kind of the unusual service condition detection and classification method of industrial processes |
| CN111079631A (en) * | 2019-12-12 | 2020-04-28 | 哈尔滨市科佳通用机电股份有限公司 | Method and system for identifying falling fault of hook lifting rod of railway wagon |
| CN111225202A (en) * | 2018-11-27 | 2020-06-02 | 杭州海康威视数字技术股份有限公司 | Picture fault diagnosis method, device and system |
| CN111404274A (en) * | 2020-04-29 | 2020-07-10 | 平顶山天安煤业股份有限公司 | Online monitoring and early warning system for displacement of power transmission system |
| CN112595918A (en) * | 2020-12-26 | 2021-04-02 | 广东电网有限责任公司广州供电局 | Low-voltage meter reading fault detection method and device |
| CN113194297A (en) * | 2021-04-30 | 2021-07-30 | 重庆市科学技术研究院 | Intelligent monitoring system and method |
| WO2021238734A1 (en) * | 2020-05-29 | 2021-12-02 | 华为技术有限公司 | Method for training neural network, and related device |
| CN115331155A (en) * | 2022-10-14 | 2022-11-11 | 智慧齐鲁(山东)大数据科技有限公司 | Mass video monitoring point location graph state detection method and system |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104268594A (en) * | 2014-09-24 | 2015-01-07 | 中安消技术有限公司 | Method and device for detecting video abnormal events |
| CN106709511A (en) * | 2016-12-08 | 2017-05-24 | 华中师范大学 | Urban rail transit panoramic monitoring video fault detection method based on depth learning |
| CN106778740A (en) * | 2016-12-06 | 2017-05-31 | 北京航空航天大学 | A kind of TFDS non-faulting image detecting methods based on deep learning |
-
2017
- 2017-07-28 CN CN201710629187.5A patent/CN107506695A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104268594A (en) * | 2014-09-24 | 2015-01-07 | 中安消技术有限公司 | Method and device for detecting video abnormal events |
| CN106778740A (en) * | 2016-12-06 | 2017-05-31 | 北京航空航天大学 | A kind of TFDS non-faulting image detecting methods based on deep learning |
| CN106709511A (en) * | 2016-12-08 | 2017-05-24 | 华中师范大学 | Urban rail transit panoramic monitoring video fault detection method based on depth learning |
Cited By (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108182452A (en) * | 2017-12-29 | 2018-06-19 | 哈尔滨工业大学（威海） | Aero-engine fault detection method and system based on grouping convolution self-encoding encoder |
| US10977558B2 (en) | 2018-02-01 | 2021-04-13 | Siemens Healthcare Gmbh | Data encoding and classification |
| EP3522079A1 (en) * | 2018-02-01 | 2019-08-07 | Siemens Healthcare Limited | Data encoding and classification |
| CN108844735A (en) * | 2018-06-22 | 2018-11-20 | 上海电力学院 | Epicyclic gearbox fault detection method based on convolution coder and Min formula distance |
| CN109087259A (en) * | 2018-08-01 | 2018-12-25 | 中国石油大学(北京) | Pre stack data denoising method and system based on convolution self-encoding encoder |
| CN109063854A (en) * | 2018-08-23 | 2018-12-21 | 河南中裕广恒科技股份有限公司 | Intelligent O&M cloud platform system and its control method |
| CN109413411A (en) * | 2018-09-06 | 2019-03-01 | 腾讯数码（天津）有限公司 | A kind of blank screen recognition methods, device and the server of monitoring circuit |
| CN111225202A (en) * | 2018-11-27 | 2020-06-02 | 杭州海康威视数字技术股份有限公司 | Picture fault diagnosis method, device and system |
| CN109816625A (en) * | 2018-11-27 | 2019-05-28 | 广东电网有限责任公司 | A kind of video quality score implementation method |
| CN111225202B (en) * | 2018-11-27 | 2022-02-11 | 杭州海康威视数字技术股份有限公司 | Picture fault diagnosis method, device and system |
| CN110210378B (en) * | 2019-05-30 | 2023-04-07 | 中国电子科技集团公司第三十八研究所 | Embedded video image analysis method and device based on edge calculation |
| CN110210378A (en) * | 2019-05-30 | 2019-09-06 | 中国电子科技集团公司第三十八研究所 | A kind of embedded video method for analyzing image and device based on edge calculations |
| CN110415220A (en) * | 2019-07-09 | 2019-11-05 | 国电大渡河瀑布沟发电有限公司 | A kind of device intelligence method for inspecting of large hydropower station |
| CN110472695B (en) * | 2019-08-22 | 2022-10-18 | 李世鹏 | Abnormal working condition detection and classification method in industrial production process |
| CN110472695A (en) * | 2019-08-22 | 2019-11-19 | 李世鹏 | A kind of the unusual service condition detection and classification method of industrial processes |
| CN110458838A (en) * | 2019-08-23 | 2019-11-15 | 讯飞智元信息科技有限公司 | A kind of detection method of fault type, device, storage medium and equipment |
| CN111079631A (en) * | 2019-12-12 | 2020-04-28 | 哈尔滨市科佳通用机电股份有限公司 | Method and system for identifying falling fault of hook lifting rod of railway wagon |
| CN111404274A (en) * | 2020-04-29 | 2020-07-10 | 平顶山天安煤业股份有限公司 | Online monitoring and early warning system for displacement of power transmission system |
| WO2021238734A1 (en) * | 2020-05-29 | 2021-12-02 | 华为技术有限公司 | Method for training neural network, and related device |
| CN113743570A (en) * | 2020-05-29 | 2021-12-03 | 华为技术有限公司 | Neural network training method and related equipment |
| CN112595918A (en) * | 2020-12-26 | 2021-04-02 | 广东电网有限责任公司广州供电局 | Low-voltage meter reading fault detection method and device |
| CN113194297A (en) * | 2021-04-30 | 2021-07-30 | 重庆市科学技术研究院 | Intelligent monitoring system and method |
| CN115331155A (en) * | 2022-10-14 | 2022-11-11 | 智慧齐鲁(山东)大数据科技有限公司 | Mass video monitoring point location graph state detection method and system |
| CN115331155B (en) * | 2022-10-14 | 2023-02-03 | 智慧齐鲁(山东)大数据科技有限公司 | Mass video monitoring point location graph state detection method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107506695A (en) | Video monitoring equipment failure automatic detection method | |
| Chen et al. | Global context-aware progressive aggregation network for salient object detection | |
| Wang et al. | Detect globally, refine locally: A novel approach to saliency detection | |
| CN109492612A (en) | Fall detection method and its falling detection device based on skeleton point | |
| CN108961245A (en) | Picture quality classification method based on binary channels depth parallel-convolution network | |
| CN106570474A (en) | Micro expression recognition method based on 3D convolution neural network | |
| CN111539290B (en) | Video motion recognition method and device, electronic equipment and storage medium | |
| US20220180624A1 (en) | Method and device for automatic identification of labels of an image | |
| CN111860171A (en) | Method and system for detecting irregular-shaped target in large-scale remote sensing image | |
| CN103761537B (en) | Image classification method based on low-rank optimization feature dictionary model | |
| CN107247952B (en) | Deep supervision-based visual saliency detection method for cyclic convolution neural network | |
| CN111242127A (en) | Vehicle detection method with granularity level multi-scale characteristics based on asymmetric convolution | |
| CN111738344A (en) | Rapid target detection method based on multi-scale fusion | |
| CN113569881A (en) | Self-adaptive semantic segmentation method based on chain residual error and attention mechanism | |
| CN113468996A (en) | Camouflage object detection method based on edge refinement | |
| CN114882278A (en) | Tire pattern classification method and device based on attention mechanism and transfer learning | |
| Bhise et al. | Plant disease detection using machine learning | |
| CN111401209B (en) | Action recognition method based on deep learning | |
| Shi et al. | SANet: Smoothed attention network for single stage face detector | |
| Ren et al. | Student behavior detection based on YOLOv4-Bi | |
| CN115578624A (en) | Agricultural disease and pest model construction method, detection method and device | |
| CN108764233A (en) | A kind of scene character recognition method based on continuous convolution activation | |
| Chen et al. | Saliency detection via topological feature modulated deep learning | |
| CN114332491A (en) | Saliency target detection algorithm based on feature reconstruction | |
| CN113505640A (en) | Small-scale pedestrian detection method based on multi-scale feature fusion |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20171222 |
|
| RJ01 | Rejection of invention patent application after publication |