CN107742093A - A kind of infrared image power equipment component real-time detection method, server and system - Google Patents

A kind of infrared image power equipment component real-time detection method, server and system Download PDF

Info

Publication number
CN107742093A
CN107742093A CN201710780528.9A CN201710780528A CN107742093A CN 107742093 A CN107742093 A CN 107742093A CN 201710780528 A CN201710780528 A CN 201710780528A CN 107742093 A CN107742093 A CN 107742093A
Authority
CN
China
Prior art keywords
prediction block
power equipment
equipment component
infrared image
prediction
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
Application number
CN201710780528.9A
Other languages
Chinese (zh)
Other versions
CN107742093B (en
Inventor
林颖
秦佳峰
辜超
郭志红
李程启
杨祎
白德盟
张皓
李娜
朱梅
徐冉
张围围
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
State Grid Corp of China SGCC
Electric Power Research Institute of State Grid Shandong Electric Power Co Ltd
Original Assignee
State Grid Corp of China SGCC
Electric Power Research Institute of State Grid Shandong Electric Power Co Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by State Grid Corp of China SGCC, Electric Power Research Institute of State Grid Shandong Electric Power Co Ltd filed Critical State Grid Corp of China SGCC
Priority to CN201710780528.9A priority Critical patent/CN107742093B/en
Publication of CN107742093A publication Critical patent/CN107742093A/en
Application granted granted Critical
Publication of CN107742093B publication Critical patent/CN107742093B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/07—Target detection

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

The invention discloses a kind of infrared image power equipment component real-time detection method, server and system, wherein the method comprising the steps of (1)：Obtain the infrared image comprising known power equipment component and form sample set, every width infrared image has indicated target frame and had component-level label wherein in sample set, and target frame is the image-region containing single known power equipment component；Step (2)：The neutral net of the power equipment component detection based on YOLO target detection frameworks is built, will be inputted using the infrared image in sample set and its corresponding component-level label to the neutral net built and it is trained；Step (3)：The neutral net detected using the power equipment component after training is handled the infrared image to be measured with unknown component level label, the testing result of output power part of appliance.

Description

A kind of infrared image power equipment component real-time detection method, server and system
Technical field
The invention belongs to electric device maintenance field, more particularly to a kind of infrared image power equipment component side of detection in real time Method, server and system.
Background technology
Power equipment is the elementary cell of operation of power networks, and status of electric power is carried out effectively, accurately to detect and assess, It is the premise of Diagnostic Examination And Repair of Electric Power Facilities and Life Cycle Management, and the important evidence of intelligent scheduling operation, can be power network Safe, reliable, efficient operation provides strong technical support.
In order to carry out fault diagnosis to power equipment, it is necessary first to which the power equipment in image is detected and positioned. In particular it is required that all parts on power equipment are carried out accurately to position and identify.It is traditional based on computer vision Infrared image power equipment component detection technique still in the feature using engineer, is not only needed under special scenes Using the parameter for adjusting many models, and when the background of infrared image is relatively complicated, traditional method can not Gratifying result is provided.
The content of the invention
In order to solve the deficiencies in the prior art, the first object of the present invention provides a kind of infrared image power equipment component Real-time detection method, the degree of accuracy that this method positions to power equipment component is high, is capable of answering for fast power part of appliance detection Use scene.
A kind of infrared image power equipment component real-time detection method of the present invention, including：
Step (1)：Obtain the infrared image comprising known power equipment component and form sample set, every width wherein in sample set Infrared image has indicated target frame and has been respectively provided with component-level label, and target frame is the figure containing single known power equipment component As region；
Step (2)：The neutral net of the power equipment component detection based on YOLO target detection frameworks is built, sample will be utilized The infrared image of this concentration and its corresponding component-level label input to the neutral net built and it are trained；
Step (3)：The neutral net detected using the power equipment component after training is to unknown component level label Infrared image to be measured is handled, the testing result of output power part of appliance.
Further, during the step (2) is trained to the neutral net built, multiple dimensioned spy is passed through Sign extracts the characteristic pattern of fusion Analysis On Multi-scale Features, prediction block is established in characteristic pattern, then handled by multi-task learning So that prediction block is close to target frame.
In step (2), infrared image is subjected to multiple dimensioned processing, wherein, yardstick refers to the change in image size, obtains Obtain a series of characteristic pattern of different scales；
Take the characteristic pattern of any lower level to carry out reorganization and make it that image size is original a quarter and depth For original 4 times, the characteristic pattern of the lower level after reorganization and the characteristic pattern of higher level are attached in the depth direction, obtained Final characteristic pattern after to fusion, this feature figure is again as the defeated of multi-task learning step after a convolution operation Enter.
In specific implementation, Multi resolution feature extraction include traditional deep neural network in convolution, activation, pond and The computing of batch standardization, specifically gradually carries out image down by original by infrared image I, and depth increases while image down Add, so as to obtain the characteristic pattern of a series of different scale.In order to obtain merging the characteristic pattern of multi-scale information, this method will be low It is original a quarter that the higher characteristic pattern restructuring of the resolution ratio of layer, which turns into length and width resolution ratio, and depth is original 4 times of characteristic pattern, Then this feature figure characteristic pattern consistent with high-rise resolution ratio is attached in the depth direction, the feature after being merged Figure.
Wherein, the image length and width resolution ratio of the characteristic pattern of lower level is larger, and the characteristic pattern length and width resolution ratio of higher level is smaller.
Further, the process being trained in the step (2) to the neutral net built, in addition to：
Infrared image is divided into the grid of default size, has at each in the grid of target frame and generates several at random Prediction block, each prediction block have box label；
In each grid with target frame, find and the maximum prediction block of rate is overlapped between target frame as the grid Actual prediction frame；
Use and be iterated computing training by training object of the box label of actual prediction frame with momentum SGD algorithms so that Actual prediction frame in each grid moves closer to target frame, completes training.
Wherein, the ratio that overlapping area of the rate between prediction block and target frame accounts for prediction block and the target frame gross area is overlapped Example.
Further, in the step (3) using the power equipment component detection after training neutral net to The process that the infrared image to be measured of unknown component level label is handled includes：
The net that unknown testing image is divided into default size is obtained in the output end of power equipment component detection neutral net Lattice and each grid obtain the result of respective prediction block；
Non-maxima suppression is carried out to all prediction blocks, prediction block is selected as final prediction knot according to confidence level Fruit.
Further, non-maxima suppression is carried out to all prediction blocks, prediction block is selected as final according to confidence level The process of prediction result be：
Firstly, for all prediction blocks for belonging to same power equipment component classification, if the friendship of any two prediction block Folded rate is more than non-maxima suppression and overlaps rate threshold value, then the confidence level of the less prediction block of confidence level is arranged into 0, confidence level compared with Big prediction block remains；
Then, the prediction block remained is screened with confidence threshold value, excludes confidence level in the box label of prediction block Less than the prediction block of confidence threshold value, confidence level is more than or equal to the prediction block of confidence threshold value in the box label of retention forecasting frame.
The second object of the present invention there is provided a kind of real-time detection service device of infrared image power equipment component.
A kind of real-time detection service device of infrared image power equipment component of the present invention, including：
Sample set builds module, and it is used to obtain the infrared image composition sample set comprising known power equipment component, its Every width infrared image has indicated target frame and has been respectively provided with component-level label in middle sample set, and target frame is containing single known electric The image-region of power part of appliance；
Neural metwork training module, it is used to build the god of the power equipment component detection based on YOLO target detection frameworks Through network, will be inputted using the infrared image in sample set and its corresponding component-level label to the neutral net that has built and right It is trained；
Part detection module, it is used for the neutral net using the power equipment component detection after training to unknown portion The infrared image to be measured of part level label is handled, the testing result of output power part of appliance.
Further, in the neural metwork training module, Multi resolution feature extraction to fusion Analysis On Multi-scale Features is passed through Characteristic pattern, establish prediction block in characteristic pattern, then handle prediction block close to target frame by multi-task learning.
In the module, infrared image is subjected to multiple dimensioned processing, wherein, yardstick refers to the change in image size, obtains Obtain a series of characteristic pattern of different scales；
Take the characteristic pattern of any lower level to carry out reorganization and make it that image size is original a quarter and depth For original 4 times, the characteristic pattern of the lower level after reorganization and the characteristic pattern of higher level are attached in the depth direction, obtained Final characteristic pattern after to fusion, this feature figure is again as the defeated of multi-task learning step after a convolution operation Enter.
In specific implementation, Multi resolution feature extraction include traditional deep neural network in convolution, activation, pond and The computing of batch standardization, specifically gradually carries out image down by original by infrared image I, and depth increases while image down Add, so as to obtain the characteristic pattern of a series of different scale.In order to obtain merging the characteristic pattern of multi-scale information, this method will be low It is original a quarter that the higher characteristic pattern restructuring of the resolution ratio of layer, which turns into length and width resolution ratio, and depth is original 4 times of characteristic pattern, Then this feature figure characteristic pattern consistent with high-rise resolution ratio is attached in the depth direction, the feature after being merged Figure.
Further, the neural metwork training module, is additionally operable to：
Infrared image is divided into the grid of default size, has at each in the grid of target frame and generates several at random Prediction block, each prediction block have box label；
In each grid with target frame, find and the maximum prediction block of rate is overlapped between target frame as the grid Actual prediction frame；
Use and be iterated computing training by training object of the box label of actual prediction frame with momentum SGD algorithms so that Actual prediction frame in each grid moves closer to target frame, completes training.
Wherein, the ratio that overlapping area of the rate between prediction block and target frame accounts for prediction block and the target frame gross area is overlapped Example.
Wherein, the image length and width resolution ratio of the characteristic pattern of lower level is larger, and the characteristic pattern length and width resolution ratio of higher level is smaller.
Further, the part detection module, is additionally operable to：
The net that unknown testing image is divided into default size is obtained in the output end of power equipment component detection neutral net Lattice and each grid obtain the result of respective prediction block；
Non-maxima suppression is carried out to all prediction blocks, prediction block is selected as final prediction knot according to confidence level Fruit.
Further, the part detection module, is additionally operable to：
Firstly, for all prediction blocks for belonging to same power equipment component classification, if the friendship of any two prediction block Folded rate is more than non-maxima suppression and overlaps rate threshold value, then the confidence level of the less prediction block of confidence level is arranged into 0, confidence level compared with Big prediction block remains；
Then, the prediction block remained is screened with confidence threshold value, excludes confidence level in the box label of prediction block Less than the prediction block of confidence threshold value, confidence level is more than or equal to the prediction block of confidence threshold value in the box label of retention forecasting frame.
The third object of the present invention there is provided a kind of infrared image power equipment component real-time detecting system.
A kind of infrared image power equipment component real-time detecting system of the present invention, including detection service device and client, The detection service device is configured as：
Obtain the infrared image comprising known power equipment component and form sample set, every width infrared image wherein in sample set Target frame is indicated and has been respectively provided with component-level label, target frame is the image-region containing single known power equipment component；
The neutral net of the power equipment component detection based on YOLO target detection frameworks is built, by using in sample set Infrared image and its corresponding component-level label input to the neutral net built and it are trained；
The neutral net detected using the power equipment component after training is to the to be measured infrared of unknown component level label Image is handled, the testing result of output power part of appliance.
Further, the detection service device is additionally configured to：During being trained to the neutral net built, Characteristic pattern by Multi resolution feature extraction to fusion Analysis On Multi-scale Features, establishes prediction block, then pass through multitask in characteristic pattern Study carries out processing and causes prediction block close to target frame.
Further, the detection service device is additionally configured to：
Infrared image is divided into the grid of default size, has at each in the grid of target frame and generates several at random Prediction block, each prediction block have box label；
In each grid with target frame, find and the maximum prediction block of rate is overlapped between target frame as the grid Actual prediction frame；
Use and be iterated computing training by training object of the box label of actual prediction frame with momentum SGD algorithms so that Actual prediction frame in each grid moves closer to target frame, completes training.
Further, the detection service device is additionally configured to：In the output end of power equipment component detection neutral net Obtain the result that unknown testing image is divided into the grid of default size and each grid obtains respective prediction block；
Non-maxima suppression is carried out to all prediction blocks, prediction block is selected as final prediction knot according to confidence level Fruit.
Further, the detection service device is additionally configured to：Non-maxima suppression, root are being carried out to all prediction blocks During selecting prediction block as final prediction result according to confidence level, firstly, for belonging to same power equipment component class Other all prediction blocks, if the overlapping rate of any two prediction block is more than non-maxima suppression and overlaps rate threshold value, by confidence The confidence level for spending less prediction block is arranged to 0, and the larger prediction block of confidence level remains；
Then, the prediction block remained is screened with confidence threshold value, excludes confidence level in the box label of prediction block Less than the prediction block of confidence threshold value, confidence level is more than or equal to the prediction block of confidence threshold value in the box label of retention forecasting frame.
Compared with prior art, the beneficial effects of the invention are as follows：
The inventive method utilizes YOLO target detection frameworks, is trained on the infrared image for largely having marked part, Fully study obtains the parameter of network, and the design method of the full convolutional neural networks of use enables the test speed of model to exist GPU is upper per second more than 20 frames, is adapted to high accuracy, the application scenarios of fast power part of appliance detection.
Brief description of the drawings
The Figure of description for forming the part of the application is used for providing further understanding of the present application, and the application's shows Meaning property embodiment and its illustrate be used for explain the application, do not form the improper restriction to the application.
Fig. 1 is the infrared image power equipment component real-time detection method flow chart of the present invention.
Fig. 2 is the neutral net schematic diagram of the structure of the present invention.
Fig. 3 (a) is the power equipment component testing result schematic diagram of the embodiment of the present invention one.
Fig. 3 (b) is the power equipment component testing result schematic diagram of the embodiment of the present invention two.
Fig. 3 (c) is the power equipment component testing result schematic diagram of the embodiment of the present invention three.
Fig. 3 (d) is the power equipment component testing result schematic diagram of the embodiment of the present invention four.
Fig. 4 is the real-time detection service device structural representation of infrared image power equipment component of the present invention.
Embodiment
It is noted that described further below is all exemplary, it is intended to provides further instruction to the application.It is unless another Indicate, all technologies used herein and scientific terminology are with usual with the application person of an ordinary skill in the technical field The identical meanings of understanding.
It should be noted that term used herein above is merely to describe embodiment, and be not intended to restricted root According to the illustrative embodiments of the application.As used herein, unless the context clearly indicates otherwise, otherwise singulative It is also intended to include plural form, additionally, it should be understood that, when in this manual using term "comprising" and/or " bag Include " when, it indicates existing characteristics, step, operation, device, component and/or combinations thereof.
Due to deep learning because its superior learning ability and ability to express, achieved in extensive object detection field Breakthrough progress.In order to provide the data required for training deep learning model, this method have collected nearly 8000 electricity first The Infrared Thermogram of power equipment, and the other mark of component-level has been carried out to them.Power equipment component detection in infrared image It is probably inclined in infrared image that difference compared to extensive target detection, which is mainly reflected in power equipment, and big absolutely at present Partial target Detection task is directed to positive target detection.
Fig. 1 is the infrared image power equipment component real-time detection method flow chart of the present invention.
As shown in figure 1, a kind of infrared image power equipment component real-time detection method of the present invention, including：
Step (1)：Obtain the infrared image comprising known power equipment component and form sample set, every width wherein in sample set Infrared image has indicated target frame and has been respectively provided with component-level label, and target frame is the figure containing single known power equipment component As region.
Every width infrared image I has indicated target frame, and target frame is the image district containing single known power equipment component Domain, every width infrared image I are respectively provided with component-level label, and component-level label is [ci,xi,yi,θi,wi,hi], wherein i represents target The sequence number of frame, ciThe classification that target inframe includes power equipment component is represented, shares the power equipment component of C classification；xi, yiThe x coordinate and y-coordinate of target frame central point, θ are represented respectivelyi,wi,hiInclination angle, width and the height of target frame are represented respectively Degree；X coordinate and y-coordinate refer respectively to the transverse and longitudinal coordinate of image.Inclination angle be target frame longitudinal edge and image abscissa direction it Between angle.
In specific implementation, power equipment component is divided into insulator, seal closure-CT, seal closure-PT, flange, grading ring, arc extinguishing The class of room six.Wherein, power equipment is included in infrared image I, is subdivided into part under power equipment so that different in image Power equipment component may belong to different power equipments, may belong to identical power equipment.
Step (2)：The neutral net of the power equipment component detection based on YOLO target detection frameworks is built, such as Fig. 2 institutes Show, will be inputted using the infrared image in sample set and its corresponding component-level label to the neutral net built and it is entered Row training.
During the step (2) is trained to the neutral net built, arrived by Multi resolution feature extraction The characteristic pattern of Analysis On Multi-scale Features is merged, prediction block is established in characteristic pattern, then carries out by multi-task learning handling prediction Frame is close to target frame.
In step (2), infrared image is subjected to multiple dimensioned processing, wherein, yardstick refers to the change in image size, obtains Obtain a series of characteristic pattern of different scales；
Take the characteristic pattern of any lower level to carry out reorganization and make it that image size is original a quarter and depth For original 4 times, the characteristic pattern of the lower level after reorganization and the characteristic pattern of higher level are attached in the depth direction, obtained Final characteristic pattern after to fusion, this feature figure is again as the defeated of multi-task learning step after a convolution operation Enter.
In specific implementation, Multi resolution feature extraction include traditional deep neural network in convolution, activation, pond and The computing of batch standardization, specifically gradually carries out image down by original by infrared image I, and depth increases while image down Add, so as to obtain the characteristic pattern of a series of different scale.In order to obtain merging the characteristic pattern of multi-scale information, this method will be low It is original a quarter that the higher characteristic pattern restructuring of the resolution ratio of layer, which turns into length and width resolution ratio, and depth is original 4 times of characteristic pattern, Then this feature figure characteristic pattern consistent with high-rise resolution ratio is attached in the depth direction, the feature after being merged Figure.
Wherein, the image length and width resolution ratio of the characteristic pattern of lower level is larger, and the characteristic pattern length and width resolution ratio of higher level is smaller.
In the process that the step (2) is trained to the neutral net built, in addition to：
Infrared image is divided into the grid of S × S sizes, has in the grid of target frame at random that (B is generation B at each Integer more than or equal to 1) individual prediction block, each prediction block has box label；Wherein, do not generated without the grid of target frame Prediction block, B prediction block allow whether to judge whether have in grid in grid with overlapping with the central point of target frame Target frame；
Each prediction block has box label [s, p, tx,ty,tθ,tw,th], wherein s represents that electric power be present in prediction block sets For the confidence level of part, the probability of power equipment component generic in the case of power equipment component in p expression prediction blocks be present Distribution, tθRepresent the inclination angle of prediction block, tx,tyThe x of prediction block central point, y-coordinate, t are represented respectivelyw,thTable is divided to represent prediction The width and height of frame；Wherein, confidence level s and probability distribution p initial value are randomly generated, and initial value is not equal to Zero.
In each grid with target frame, find and the maximum prediction block of rate is overlapped between target frame as the grid Actual prediction frame；
Use and be iterated computing training by training object of the box label of actual prediction frame with momentum SGD algorithms so that Actual prediction frame in each grid moves closer to target frame, completes training.
Wherein, the ratio that overlapping area of the rate between prediction block and target frame accounts for prediction block and the target frame gross area is overlapped Example.
Positioning loss function, Classification Loss function is specifically included in the training process of the neutral net of step (2) and is inclined Oblique angle consistency constraint loss function：
L=Lloc+Lcls+Lort
Wherein, L represents all kinds of loss function sums, LlocRepresent positioning loss function, LclsPresentation class loss function, LortRepresent inclination angle uniformity loss function；
Positioning loss function is expressed as：
Wherein,Represent whether j-th of prediction block in i-th of grid is closest to the indicator function of prediction block, when i-th J-th of prediction block is closest to indicator function during prediction block in individual gridFor 1, when j-th of prediction block be not in i-th of grid It is closest to indicator function during prediction blockFor 0；
Represent the confidence level true value of power equipment component in prediction block be present,Represent target frame position and angle parameter Actual value,λnoobjRepresent the loss function weight of the grid in the absence of power equipment component, λlocRepresent The loss function weight of location tasks；
Represent whether j-th of prediction block be non-closest to prediction block in i-th of grid, when j-th pre- in i-th of grid Survey frame and be not closest to indicator function during prediction blockFor 1, when j-th of prediction block is closest to prediction block in i-th of grid When indicator functionFor 0.
sijThe forecast confidence of j-th of prediction block in i-th of grid is represented,Represent j-th of prediction block in i-th of grid Corresponding true confidence level.When j-th of prediction block is closest to prediction block in i-th of grid,For 1, otherwiseFor 0.
tijThe predicted position and angle parameter of j-th of prediction block in i-th of grid are represented,Represent jth in i-th of grid The position of real goal frame and angle parameter corresponding to individual prediction block.It is only closest when j-th of prediction block in i-th of grid During prediction block, it just has corresponding real goal frame.L2 norm calculations are sought in expression.
Classification Loss function representation is：
Wherein, λclsThe loss function weight of presentation class task；
Wherein, pijRepresent that j-th of prediction block in i-th of grid belongs to the probability of the prediction of each power equipment component classification Distribution,Represent that j-th of prediction block belongs to each real probability distribution of power equipment component classification in i-th of grid.Only When j-th of prediction block is closest to prediction block in i-th of grid, this partial loss function can be just calculated.
Inclination angle consistency constraint loss function is expressed as：
Wherein, g represents g-th of equipment group, it is assumed that image I mono- shares G equipment group.Represent the of i-th of grid The indicator function of relation between corresponding g-th of the equipment group of j prediction block, when j-th of prediction block of i-th of grid be closest to it is pre- Survey frame and belong to g-th of equipment group ΩgWhen,It is otherwise 0 for 1；Represent the average value at g-th of equipment group inclination angle；Represent the inclination angle of the prediction of j-th of prediction block in i-th of grid.
The present invention allows all parts belonged on same power equipment corresponding by inclination angle consistency constraint loss function The inclination angle of prediction block be consistent：All parts on same power equipment are found first so that any two portion The inclination angle of part prediction block approaches, and the inclination angle of the both parts prediction block line of centres and the average value at inclination angle connect Closely.So that each equipment part corresponding to one group of inclination angle close to prediction block of G equipment forms on image.
This method employs the target detection framework based on YOLO, is carried in training by adding inclination angle consistency constraint The stationkeeping ability of high model.Propagated forward during test directly by the image of Unknown Label by a neutral net, process are non- Maximum suppresses to can obtain the testing result of power equipment component.20 frames speed per second can be reached more than during test on GPU Degree.
In specific implementation, momentum is arranged to 0.9, altogether iteration 90000 times, and preceding 30000 learning rates are 0.01, after 60000 learning rates are the parameter that the neutral net of power equipment component detection is preserved after 0.001. training terminates.
Step (3)：The neutral net detected using the power equipment component after training is to unknown component level label Infrared image to be measured is handled, the testing result of output power part of appliance.
Using the neutral net of the power equipment component detection after training to unknown component level in the step (3) The process that the infrared image to be measured of label is handled includes：
The net that unknown testing image is divided into default size is obtained in the output end of power equipment component detection neutral net Lattice and each grid obtain the result of respective prediction block；
Non-maxima suppression is carried out to all prediction blocks, prediction block is selected as final prediction knot according to confidence level Fruit.
Non-maxima suppression is carried out to all prediction blocks, prediction block is selected as final prediction result according to confidence level Process be：
Firstly, for all prediction blocks for belonging to same power equipment component classification, if the friendship of any two prediction block Folded rate is more than non-maxima suppression and overlaps rate threshold value, then the confidence level of the less prediction block of confidence level is arranged into 0, confidence level compared with Big prediction block remains；
Then, the prediction block remained is screened with confidence threshold value, excludes confidence level in the box label of prediction block Less than the prediction block of confidence threshold value, confidence level is more than or equal to the prediction block of confidence threshold value in the box label of retention forecasting frame.
In specific implementation, the input for the neutral net that unknown images are detected as power equipment component, all nets are obtained The prediction block of lattice prediction, using non-maxima suppression algorithm, when the overlapping rate for being predicted as two mutually similar prediction blocks is more than 0.4 When, the confidence level of the less prediction block of confidence level is arranged to 0, the larger prediction block of confidence level retains.Finally, confidence level is selected Prediction block more than 0.2 is as final prediction result.Some power equipment component testing results of Fig. 2 displaying embodiments.
The present embodiment is finally tested on the infrared image power equipment component data set of collection, such as Fig. 3 (a)-Fig. 3 (d) shown in, a shared insulator, seal closure-CT, seal closure-PT, flange, grading ring, arc-chutes totally 6 kinds of power equipment components Type.Wherein, 0 is insulator；1 is seal closure-CT；2 be seal closure-PT；3 be grading ring；4 be flange；5 be arc-chutes.
Randomly select data set 60% is trained, and remaining 40% is tested.Commented using the standard of target detection Sentence criterion AP and mAP to be evaluated and tested, table 1 gives AP and mAP value of this method on test set, and wherein mAP value is each The AP values of classification are averaged.AP and mAP values are bigger, illustrate that performance is better.
It can be seen that, the mAP values of this method have reached 92.4, and the AP values of wherein particular category have exceeded 95 from upper table. It can show that this method protrudes significant technique effect from test result.
The inventive method utilizes YOLO target detection frameworks, is trained on the infrared image for largely having marked part, Fully study obtains the parameter of network, and the standard that this method positions to power equipment component is improved using inclination angle consistency constraint Exactness.This method use full convolutional neural networks design method enable model test speed on GPU more than 20 Frame is per second, is adapted to high accuracy, the application scenarios of fast power part of appliance detection.
Fig. 4 is the real-time detection service device structural representation of infrared image power equipment component of the present invention.
As shown in figure 4, a kind of real-time detection service device of infrared image power equipment component of the present invention, including：
(1) sample set structure module, it is used to obtain the infrared image composition sample set comprising known power equipment component, Every width infrared image has indicated target frame and has been respectively provided with component-level label wherein in sample set, and target frame is containing single known The image-region of power equipment component.
(2) neural metwork training module, it is used to build the power equipment component detection based on YOLO target detection frameworks Neutral net, will be inputted using the infrared image in sample set and its corresponding component-level label to the neutral net built And it is trained；
In the neural metwork training module, the characteristic pattern by Multi resolution feature extraction to fusion Analysis On Multi-scale Features, Prediction block is established in characteristic pattern, then handle prediction block close to target frame by multi-task learning.
In the module, infrared image is subjected to multiple dimensioned processing, wherein, yardstick refers to the change in image size, obtains Obtain a series of characteristic pattern of different scales；
Take the characteristic pattern of any lower level to carry out reorganization and make it that image size is original a quarter and depth For original 4 times, the characteristic pattern of the lower level after reorganization and the characteristic pattern of higher level are attached in the depth direction, obtained Final characteristic pattern after to fusion, this feature figure is again as the defeated of multi-task learning step after a convolution operation Enter.
In specific implementation, Multi resolution feature extraction include traditional deep neural network in convolution, activation, pond and The computing of batch standardization, specifically gradually carries out image down by original by infrared image I, and depth increases while image down Add, so as to obtain the characteristic pattern of a series of different scale.In order to obtain merging the characteristic pattern of multi-scale information, this method will be low It is original a quarter that the higher characteristic pattern restructuring of the resolution ratio of layer, which turns into length and width resolution ratio, and depth is original 4 times of characteristic pattern, Then this feature figure characteristic pattern consistent with high-rise resolution ratio is attached in the depth direction, the feature after being merged Figure.
The neural metwork training module, is additionally operable to：
Infrared image is divided into the grid of default size, has at each in the grid of target frame and generates several at random Prediction block, each prediction block have box label；
In each grid with target frame, find and the maximum prediction block of rate is overlapped between target frame as the grid Actual prediction frame；
Use and be iterated computing training by training object of the box label of actual prediction frame with momentum SGD algorithms so that Actual prediction frame in each grid moves closer to target frame, completes training.
Wherein, the ratio that overlapping area of the rate between prediction block and target frame accounts for prediction block and the target frame gross area is overlapped Example.
Wherein, the image length and width resolution ratio of the characteristic pattern of lower level is larger, and the characteristic pattern length and width resolution ratio of higher level is smaller.
(3) part detection module, it is used for the neutral net using the power equipment component detection after training to not Know that the infrared image to be measured of component-level label is handled, the testing result of output power part of appliance.
The part detection module, is additionally operable to：
The net that unknown testing image is divided into default size is obtained in the output end of power equipment component detection neutral net Lattice and each grid obtain the result of respective prediction block；
Non-maxima suppression is carried out to all prediction blocks, prediction block is selected as final prediction knot according to confidence level Fruit.
The part detection module, is additionally operable to：
Firstly, for all prediction blocks for belonging to same power equipment component classification, if the friendship of any two prediction block Folded rate is more than non-maxima suppression and overlaps rate threshold value, then the confidence level of the less prediction block of confidence level is arranged into 0, confidence level compared with Big prediction block remains；
Then, the prediction block remained is screened with confidence threshold value, excludes confidence level in the box label of prediction block Less than the prediction block of confidence threshold value, confidence level is more than or equal to the prediction block of confidence threshold value in the box label of retention forecasting frame.
The server by utilizing YOLO target detection frameworks of the present invention, are carried out on the infrared image for largely having marked part Training, fully study obtain the parameter of network, and the design method for the full convolutional neural networks that the system uses causes the survey of model Trying speed can be per second more than 20 frames on GPU, is adapted to high accuracy, the application scenarios of fast power part of appliance detection.
Present invention also offers a kind of infrared image power equipment component real-time detecting system.
A kind of infrared image power equipment component real-time detecting system of the present invention, including detection service device and client, The detection service device is configured as：
Obtain the infrared image comprising known power equipment component and form sample set, every width infrared image wherein in sample set Target frame is indicated and has been respectively provided with component-level label, target frame is the image-region containing single known power equipment component；
The neutral net of the power equipment component detection based on YOLO target detection frameworks is built, by using in sample set Infrared image and its corresponding component-level label input to the neutral net built and it are trained；
The neutral net detected using the power equipment component after training is to the to be measured infrared of unknown component level label Image is handled, the testing result of output power part of appliance.
Further, the detection service device is additionally configured to：During being trained to the neutral net built, Characteristic pattern by Multi resolution feature extraction to fusion Analysis On Multi-scale Features, establishes prediction block, then pass through multitask in characteristic pattern Study carries out processing and causes prediction block close to target frame.
Further, the detection service device is additionally configured to：
Infrared image is divided into the grid of default size, has at each in the grid of target frame and generates several at random Prediction block, each prediction block have box label；
In each grid with target frame, find and the maximum prediction block of rate is overlapped between target frame as the grid Actual prediction frame；
Use and be iterated computing training by training object of the box label of actual prediction frame with momentum SGD algorithms so that Actual prediction frame in each grid moves closer to target frame, completes training.
Further, the detection service device is additionally configured to：In the output end of power equipment component detection neutral net Obtain the result that unknown testing image is divided into the grid of default size and each grid obtains respective prediction block；
Non-maxima suppression is carried out to all prediction blocks, prediction block is selected as final prediction knot according to confidence level Fruit.
Further, the detection service device is additionally configured to：Non-maxima suppression, root are being carried out to all prediction blocks During selecting prediction block as final prediction result according to confidence level, firstly, for belonging to same power equipment component class Other all prediction blocks, if the overlapping rate of any two prediction block is more than non-maxima suppression and overlaps rate threshold value, by confidence The confidence level for spending less prediction block is arranged to 0, and the larger prediction block of confidence level remains；
Then, the prediction block remained is screened with confidence threshold value, excludes confidence level in the box label of prediction block Less than the prediction block of confidence threshold value, confidence level is more than or equal to the prediction block of confidence threshold value in the box label of retention forecasting frame.
The system of the present invention utilizes YOLO target detection frameworks, is instructed on the infrared image for largely having marked part Practice, fully study obtains the parameter of network, and the design method for the full convolutional neural networks that the system uses causes the test of model Speed can be per second more than 20 frames on GPU, is adapted to high accuracy, the application scenarios of fast power part of appliance detection.
Although above-mentioned the embodiment of the present invention is described with reference to accompanying drawing, model not is protected to the present invention The limitation enclosed, one of ordinary skill in the art should be understood that on the basis of technical scheme those skilled in the art are not Need to pay various modifications or deformation that creative work can make still within protection scope of the present invention.

Claims (15)

  1. A kind of 1. infrared image power equipment component real-time detection method, it is characterised in that including：
    Step (1)：Obtain the infrared image comprising known power equipment component and form sample set, it is every infrared wherein in sample set Image has indicated target frame and has been respectively provided with component-level label, and target frame is the image district containing single known power equipment component Domain；
    Step (2)：The neutral net of the power equipment component detection based on YOLO target detection frameworks is built, sample set will be utilized In infrared image and its corresponding component-level label input to the neutral net built and it be trained；
    Step (3)：The neutral net detected using the power equipment component after training is to the to be measured of unknown component level label Infrared image is handled, the testing result of output power part of appliance.
  2. 2. a kind of infrared image power equipment component real-time detection method as claimed in claim 1, it is characterised in that described During step (2) is trained to the neutral net built, pass through Multi resolution feature extraction to fusion Analysis On Multi-scale Features Characteristic pattern, establish prediction block in characteristic pattern, then handle prediction block close to target frame by multi-task learning.
  3. 3. a kind of infrared image power equipment component real-time detection method as claimed in claim 2, it is characterised in that described The process that step (2) is trained to the neutral net built, in addition to：
    Infrared image is divided into the grid of default size, has at each and generates several predictions in the grid of target frame at random Frame, each prediction block have box label；
    In each grid with target frame, the reality that the maximum prediction block of rate is overlapped between target frame as the grid is found Border prediction block；
    Use and be iterated computing training by training object of the box label of actual prediction frame with momentum SGD algorithms so that be each Actual prediction frame in grid moves closer to target frame, completes training.
  4. 4. a kind of infrared image power equipment component real-time detection method as claimed in claim 1, it is characterised in that described Using the neutral net of the power equipment component detection after training to the to be measured infrared of unknown component level label in step (3) The process that image is handled includes：
    The output end of power equipment component detection neutral net obtain unknown testing image be divided into the grid of default size with And each grid obtains the result of respective prediction block；
    Non-maxima suppression is carried out to all prediction blocks, prediction block is selected as final prediction result according to confidence level.
  5. 5. a kind of infrared image power equipment component real-time detection method as claimed in claim 4, it is characterised in that to all Prediction block carry out non-maxima suppression, according to confidence level select prediction block as final prediction result process for：
    Firstly, for all prediction blocks for belonging to same power equipment component classification, if the overlapping rate of any two prediction block Rate threshold value is overlapped more than non-maxima suppression, then the confidence level of the less prediction block of confidence level is arranged to 0, confidence level is larger Prediction block remains；
    Then, the prediction block remained is screened with confidence threshold value, is excluded confidence level in the box label of prediction block and is less than The prediction block of confidence threshold value, confidence level is more than or equal to the prediction block of confidence threshold value in the box label of retention forecasting frame.
  6. A kind of 6. real-time detection service device of infrared image power equipment component, it is characterised in that including：
    Sample set builds module, and it is used to obtain the infrared image composition sample set comprising known power equipment component, wherein sample The every width infrared image of this concentration has indicated target frame and has been respectively provided with component-level label, and target frame is set containing single known electric power The image-region of standby part；
    Neural metwork training module, it is used to build the nerve net of the power equipment component detection based on YOLO target detection frameworks Network, it will be inputted using the infrared image in sample set and its corresponding component-level label to the neutral net built and it entered Row training；
    Part detection module, it is used for the neutral net using the power equipment component detection after training to unknown component level The infrared image to be measured of label is handled, the testing result of output power part of appliance.
  7. 7. a kind of real-time detection service device of infrared image power equipment component as claimed in claim 6, it is characterised in that in institute State in neural metwork training module, the characteristic pattern by Multi resolution feature extraction to fusion Analysis On Multi-scale Features, built in characteristic pattern Vertical prediction block, then handle prediction block close to target frame by multi-task learning.
  8. 8. a kind of real-time detection service device of infrared image power equipment component as claimed in claim 7, it is characterised in that described Neural metwork training module, is additionally operable to：
    Infrared image is divided into the grid of default size, has at each and generates several predictions in the grid of target frame at random Frame, each prediction block have box label；
    In each grid with target frame, the reality that the maximum prediction block of rate is overlapped between target frame as the grid is found Border prediction block；
    Use and be iterated computing training by training object of the box label of actual prediction frame with momentum SGD algorithms so that be each Actual prediction frame in grid moves closer to target frame, completes training.
  9. 9. a kind of real-time detection service device of infrared image power equipment component as claimed in claim 6, it is characterised in that described Part detection module, is additionally operable to：
    The output end of power equipment component detection neutral net obtain unknown testing image be divided into the grid of default size with And each grid obtains the result of respective prediction block；Non-maxima suppression is carried out to all prediction blocks, selected according to confidence level Prediction block is selected as final prediction result.
  10. A kind of 10. real-time detection service device of infrared image power equipment component as claimed in claim 9, it is characterised in that institute Part detection module is stated, is additionally operable to：
    Firstly, for all prediction blocks for belonging to same power equipment component classification, if the overlapping rate of any two prediction block Rate threshold value is overlapped more than non-maxima suppression, then the confidence level of the less prediction block of confidence level is arranged to 0, confidence level is larger Prediction block remains；
    Then, the prediction block remained is screened with confidence threshold value, is excluded confidence level in the box label of prediction block and is less than The prediction block of confidence threshold value, confidence level is more than or equal to the prediction block of confidence threshold value in the box label of retention forecasting frame.
  11. 11. a kind of infrared image power equipment component real-time detecting system, it is characterised in that including detection service device and client End, the detection service device are configured as：
    Obtain the infrared image comprising known power equipment component and form sample set, every width infrared image has been wherein in sample set Indicate target frame and be respectively provided with component-level label, target frame is the image-region containing single known power equipment component；
    The neutral net of the power equipment component detection based on YOLO target detection frameworks is built, it is infrared in sample set by utilizing Image and its corresponding component-level label input to the neutral net built and it are trained；
    The neutral net detected using the power equipment component after training is to the infrared image to be measured with unknown component level label Handled, the testing result of output power part of appliance.
  12. 12. a kind of infrared image power equipment component real-time detecting system as claimed in claim 11, it is characterised in that described Detection service device is additionally configured to：During being trained to the neutral net built, arrived by Multi resolution feature extraction The characteristic pattern of Analysis On Multi-scale Features is merged, prediction block is established in characteristic pattern, then carries out by multi-task learning handling prediction Frame is close to target frame.
  13. 13. a kind of infrared image power equipment component real-time detecting system as claimed in claim 12, it is characterised in that described Detection service device is additionally configured to：
    Infrared image is divided into the grid of default size, has at each and generates several predictions in the grid of target frame at random Frame, each prediction block have box label；
    In each grid with target frame, the reality that the maximum prediction block of rate is overlapped between target frame as the grid is found Border prediction block；
    Use and be iterated computing training by training object of the box label of actual prediction frame with momentum SGD algorithms so that be each Actual prediction frame in grid moves closer to target frame, completes training.
  14. 14. a kind of infrared image power equipment component real-time detecting system as claimed in claim 11, it is characterised in that described Detection service device is additionally configured to：Unknown testing image is obtained in the output end of power equipment component detection neutral net to be divided into The grid and each grid of default size obtain the result of respective prediction block；
    Non-maxima suppression is carried out to all prediction blocks, prediction block is selected as final prediction result according to confidence level.
  15. 15. a kind of infrared image power equipment component real-time detecting system as claimed in claim 14, it is characterised in that described Detection service device is additionally configured to：Non-maxima suppression is being carried out to all prediction blocks, is selecting prediction block to make according to confidence level During for final prediction result, firstly, for all prediction blocks for belonging to same power equipment component classification, if appointing The overlapping rate for two prediction blocks of anticipating is more than non-maxima suppression and overlaps rate threshold value, then by the confidence level of the less prediction block of confidence level 0 is arranged to, the larger prediction block of confidence level remains；
    Then, the prediction block remained is screened with confidence threshold value, is excluded confidence level in the box label of prediction block and is less than The prediction block of confidence threshold value, confidence level is more than or equal to the prediction block of confidence threshold value in the box label of retention forecasting frame.
CN201710780528.9A 2017-09-01 2017-09-01 Real-time detection method, server and system for infrared image power equipment components Active CN107742093B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201710780528.9A CN107742093B (en) 2017-09-01 2017-09-01 Real-time detection method, server and system for infrared image power equipment components

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201710780528.9A CN107742093B (en) 2017-09-01 2017-09-01 Real-time detection method, server and system for infrared image power equipment components

Publications (2)

Publication Number Publication Date
CN107742093A true CN107742093A (en) 2018-02-27
CN107742093B CN107742093B (en) 2020-05-05

Family

ID=61235173

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201710780528.9A Active CN107742093B (en) 2017-09-01 2017-09-01 Real-time detection method, server and system for infrared image power equipment components

Country Status (1)

Country Link
CN (1) CN107742093B (en)

Cited By (21)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108288055A (en) * 2018-03-14 2018-07-17 台州智必安科技有限责任公司 Block of bow collector of electric locomotive based on depth network and placement test and arc method for measuring
CN108416979A (en) * 2018-03-08 2018-08-17 辽宁工程技术大学 A kind of intelligence the elderly's tumbling alarm system
CN108597053A (en) * 2018-04-25 2018-09-28 北京御航智能科技有限公司 Shaft tower and channel targets identification based on image data and neural network and defect diagnostic method
CN108968811A (en) * 2018-06-20 2018-12-11 四川斐讯信息技术有限公司 A kind of object identification method and system of sweeping robot
CN109165611A (en) * 2018-08-31 2019-01-08 广州市派客朴食信息科技有限责任公司 A kind of dishes identification intelligent settlement method based on machine vision and neural network
CN109255320A (en) * 2018-09-03 2019-01-22 电子科技大学 A kind of improved non-maxima suppression method
CN109344850A (en) * 2018-07-27 2019-02-15 华南理工大学 A kind of water meter automatic testing method based on YOLO
CN109409197A (en) * 2018-08-31 2019-03-01 广州市派客朴食信息科技有限责任公司 A kind of dishes detection method neural network based
CN109740662A (en) * 2018-12-28 2019-05-10 成都思晗科技股份有限公司 Image object detection method based on YOLO frame
CN109784336A (en) * 2019-01-30 2019-05-21 云南电网有限责任公司电力科学研究院 A kind of infrared image fault point recognition methods based on YOLO algorithm of target detection
CN109784348A (en) * 2018-12-17 2019-05-21 中国科学院深圳先进技术研究院 A kind of infrared power equipment identification and inline diagnosis method and its system
CN109783560A (en) * 2019-01-16 2019-05-21 江苏圣通电力新能源科技有限公司 The detection method of status of electric power based on deep learning multiple features fusion
CN109815800A (en) * 2018-12-17 2019-05-28 广东电网有限责任公司 Object detection method and system based on regression algorithm
CN110618424A (en) * 2019-09-27 2019-12-27 中科九度(北京)空间信息技术有限责任公司 Remote high-voltage line discovery method based on multi-sensor fusion
CN110738238A (en) * 2019-09-18 2020-01-31 平安科技（深圳）有限公司 certificate information classification positioning method and device
CN110992307A (en) * 2019-11-04 2020-04-10 华北电力大学（保定） Insulator positioning and identifying method and device based on YOLO
CN111666958A (en) * 2019-03-05 2020-09-15 中科院微电子研究所昆山分所 Method, device, equipment and medium for detecting equipment state based on image recognition
CN111784767A (en) * 2020-06-08 2020-10-16 珠海格力电器股份有限公司 Method and device for determining target position
CN111815581A (en) * 2020-06-27 2020-10-23 国网上海市电力公司 Power isolating switch and automatic part identification method and device thereof
CN111815567A (en) * 2020-06-15 2020-10-23 国网上海市电力公司 Automatic labeling method and device for high-recognition-rate power equipment
CN112380952A (en) * 2020-11-10 2021-02-19 广西大学 Power equipment infrared image real-time detection and identification method based on artificial intelligence

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106022237A (en) * 2016-05-13 2016-10-12 电子科技大学 Pedestrian detection method based on end-to-end convolutional neural network
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
CN106778472A (en) * 2016-11-17 2017-05-31 成都通甲优博科技有限责任公司 The common invader object detection and recognition method in transmission of electricity corridor based on deep learning
CN107038448A (en) * 2017-03-01 2017-08-11 中国科学院自动化研究所 Target detection model building method

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106022237A (en) * 2016-05-13 2016-10-12 电子科技大学 Pedestrian detection method based on end-to-end convolutional neural network
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
CN106778472A (en) * 2016-11-17 2017-05-31 成都通甲优博科技有限责任公司 The common invader object detection and recognition method in transmission of electricity corridor based on deep learning
CN107038448A (en) * 2017-03-01 2017-08-11 中国科学院自动化研究所 Target detection model building method

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
BAOGUANG SHI ET AL.: "Detecting Oriented Text in Natural Images by Linking Segments", 《ARXIV》 *
JOSEPH REDMON ET AL.: "YOLO9000:Better, Faster, Stronger", 《2017 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 *
JOSEPH REDMON ET AL.: "You Only Look Once:Unified, Real-Time Object Detection", 《2016 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 *
MOHD SHAWAL JADIN ET AL.: "Recent progress in diagnosing the reliability of electrical equipment by using infrared thermography", 《ELSEVIER》 *
S. HE AND R. W. LAU: "Oriented Object Proposals", 《 PROC. INT. CONF.COMPUT. VIS.》 *

Cited By (28)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108416979A (en) * 2018-03-08 2018-08-17 辽宁工程技术大学 A kind of intelligence the elderly's tumbling alarm system
CN108288055B (en) * 2018-03-14 2020-09-22 台州智必安科技有限责任公司 Electric locomotive pantograph and electric arc detection method based on deep network and classification test
CN108288055A (en) * 2018-03-14 2018-07-17 台州智必安科技有限责任公司 Block of bow collector of electric locomotive based on depth network and placement test and arc method for measuring
CN108597053A (en) * 2018-04-25 2018-09-28 北京御航智能科技有限公司 Shaft tower and channel targets identification based on image data and neural network and defect diagnostic method
CN108968811A (en) * 2018-06-20 2018-12-11 四川斐讯信息技术有限公司 A kind of object identification method and system of sweeping robot
CN109344850A (en) * 2018-07-27 2019-02-15 华南理工大学 A kind of water meter automatic testing method based on YOLO
CN109165611A (en) * 2018-08-31 2019-01-08 广州市派客朴食信息科技有限责任公司 A kind of dishes identification intelligent settlement method based on machine vision and neural network
CN109409197A (en) * 2018-08-31 2019-03-01 广州市派客朴食信息科技有限责任公司 A kind of dishes detection method neural network based
CN109165611B (en) * 2018-08-31 2020-03-24 广州市派客朴食信息科技有限责任公司 Intelligent dish identification settlement method based on machine vision and neural network
CN109255320A (en) * 2018-09-03 2019-01-22 电子科技大学 A kind of improved non-maxima suppression method
CN109255320B (en) * 2018-09-03 2020-09-25 电子科技大学 Improved non-maximum suppression method
CN109815800A (en) * 2018-12-17 2019-05-28 广东电网有限责任公司 Object detection method and system based on regression algorithm
CN109784348B (en) * 2018-12-17 2021-06-18 中国科学院深圳先进技术研究院 Infrared power equipment identification and online diagnosis method and system
CN109784348A (en) * 2018-12-17 2019-05-21 中国科学院深圳先进技术研究院 A kind of infrared power equipment identification and inline diagnosis method and its system
CN109740662A (en) * 2018-12-28 2019-05-10 成都思晗科技股份有限公司 Image object detection method based on YOLO frame
CN109783560A (en) * 2019-01-16 2019-05-21 江苏圣通电力新能源科技有限公司 The detection method of status of electric power based on deep learning multiple features fusion
CN109784336A (en) * 2019-01-30 2019-05-21 云南电网有限责任公司电力科学研究院 A kind of infrared image fault point recognition methods based on YOLO algorithm of target detection
CN111666958A (en) * 2019-03-05 2020-09-15 中科院微电子研究所昆山分所 Method, device, equipment and medium for detecting equipment state based on image recognition
CN110738238A (en) * 2019-09-18 2020-01-31 平安科技（深圳）有限公司 certificate information classification positioning method and device
CN110738238B (en) * 2019-09-18 2023-05-26 平安科技（深圳）有限公司 Classification positioning method and device for certificate information
CN110618424A (en) * 2019-09-27 2019-12-27 中科九度(北京)空间信息技术有限责任公司 Remote high-voltage line discovery method based on multi-sensor fusion
CN110618424B (en) * 2019-09-27 2021-09-21 中科九度(北京)空间信息技术有限责任公司 Remote high-voltage line discovery method based on multi-sensor fusion
CN110992307A (en) * 2019-11-04 2020-04-10 华北电力大学（保定） Insulator positioning and identifying method and device based on YOLO
CN111784767A (en) * 2020-06-08 2020-10-16 珠海格力电器股份有限公司 Method and device for determining target position
CN111815567A (en) * 2020-06-15 2020-10-23 国网上海市电力公司 Automatic labeling method and device for high-recognition-rate power equipment
CN111815581A (en) * 2020-06-27 2020-10-23 国网上海市电力公司 Power isolating switch and automatic part identification method and device thereof
CN112380952A (en) * 2020-11-10 2021-02-19 广西大学 Power equipment infrared image real-time detection and identification method based on artificial intelligence
CN112380952B (en) * 2020-11-10 2022-10-11 广西大学 Power equipment infrared image real-time detection and identification method based on artificial intelligence

Also Published As

Publication number Publication date
CN107742093B (en) 2020-05-05

Similar Documents

Publication Publication Date Title
CN107742093A (en) A kind of infrared image power equipment component real-time detection method, server and system
CN107563412A (en) A kind of infrared image power equipment real-time detection method based on deep learning
CN106780612B (en) Object detecting method and device in a kind of image
CN106356757A (en) Method for inspecting electric power lines by aid of unmanned aerial vehicle on basis of human vision characteristics
CN103839065B (en) Extraction method for dynamic crowd gathering characteristics
CN108038846A (en) Transmission line equipment image defect detection method and system based on multilayer convolutional neural networks
CN110188720A (en) A kind of object detection method and system based on convolutional neural networks
CN109784333A (en) Based on an objective detection method and system for cloud bar power channel characteristics
CN109615611A (en) A kind of insulator self-destruction defect inspection method based on inspection image
CN106651886A (en) Cloud image segmentation method based on superpixel clustering optimization CNN
CN108052881A (en) The method and apparatus of multiclass entity object in a kind of real-time detection construction site image
CN109446925A (en) A kind of electric device maintenance algorithm based on convolutional neural networks
CN110378297A (en) A kind of Remote Sensing Target detection method based on deep learning
CN107392247A (en) Atural object safe distance real-time detection method below a kind of power line
CN110516677A (en) A kind of neural network recognization model, target identification method and system
CN109559310A (en) Power transmission and transformation inspection image quality evaluating method and system based on conspicuousness detection
CN115731164A (en) Insulator defect detection method based on improved YOLOv7
CN107843913B (en) The 3D display method and apparatus of radiation field
CN109712127A (en) A kind of electric transmission line fault detection method for patrolling video flowing for machine
CN112528913A (en) Grit particulate matter particle size detection analytic system based on image
CN109272487A (en) The quantity statistics method of crowd in a kind of public domain based on video
CN107480826A (en) The application of powerline ice-covering early warning three dimension system based on GIS
CN109636764A (en) A kind of image style transfer method based on deep learning and conspicuousness detection
CN109543617A (en) The detection method of intelligent vehicle movement traffic information based on YOLO target detection technique
CN115169855A (en) Unsafe state detection method based on digital twin workshop mixed data set

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
CB02 Change of applicant information

Address after: Ji'nan City, Shandong Province Wang Yue Road 250003 No. 2000

Applicant after: Electric Power Research Institute of State Grid Shandong Electric Power Company

Applicant after: State Grid Co., Ltd.

Address before: Ji'nan City, Shandong Province Wang Yue Road 250003 No. 2000

Applicant before: Electric Power Research Institute of State Grid Shandong Electric Power Company

Applicant before: State Grid Corporation

CB02 Change of applicant information
CB03 Change of inventor or designer information

Inventor after: Lin Ying

Inventor after: Li Na

Inventor after: Zhu Mei

Inventor after: Xu Ran

Inventor after: Zhang Weiwei

Inventor after: Qin Jiafeng

Inventor after: Gu Chao

Inventor after: Qian Yuanliang

Inventor after: Guo Zhihong

Inventor after: Li Chengqi

Inventor after: Yang Dai

Inventor after: Bai Demeng

Inventor after: Zhang Hao

Inventor before: Lin Ying

Inventor before: Zhu Mei

Inventor before: Xu Ran

Inventor before: Zhang Weiwei

Inventor before: Qin Jiafeng

Inventor before: Gu Chao

Inventor before: Guo Zhihong

Inventor before: Li Chengqi

Inventor before: Yang Dai

Inventor before: Bai Demeng

Inventor before: Zhang Hao

Inventor before: Li Na

CB03 Change of inventor or designer information
GR01 Patent grant
GR01 Patent grant