CN109886947A - The high-tension bus-bar defect inspection method of convolutional neural networks based on region - Google Patents
The high-tension bus-bar defect inspection method of convolutional neural networks based on region Download PDFInfo
- Publication number
- CN109886947A CN109886947A CN201910122569.8A CN201910122569A CN109886947A CN 109886947 A CN109886947 A CN 109886947A CN 201910122569 A CN201910122569 A CN 201910122569A CN 109886947 A CN109886947 A CN 109886947A
- Authority
- CN
- China
- Prior art keywords
- region
- tension bus
- bar
- proposal
- defect
- 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
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a kind of high-tension bus-bar defect inspection methods of convolutional neural networks based on region.Obtain high-tension bus-bar image deflects sample set;Convolutional neural networks are input to, the defects of high-tension bus-bar image feature is extracted to obtain characteristic pattern;Characteristic pattern is input in the region proposal network for having additional full articulamentum, network proposing offers region is proposed by region and judges to propose region;The proposal region for belonging to prospect and characteristic pattern are inputted and carry out defect classification in full articulamentum;Correcting process proposes that the boundary in region, optimization boundary obtain more accurate proposal region;Step is repeated to region proposal network iteration until completing to train, obtains the high-tension bus-bar defects detection model progress actual defects detection for training parameter.The present invention carries out identifying and positioning for electric wire defect, recognition speed with higher and high accuracy rate using Faster R-CNN algorithm and has preferable robustness.
Description
Technical field
The present invention relates to machine vision and depth learning technology field, specially the height of the convolutional neural networks based on region
Piezoelectric wire defect inspection method.
Background technique
The stable operation of ultra-high-tension power transmission line is to ensure the important link of national economic development.With the development of artificial intelligence
With it is universal, intelligent patrol detection mode also gradually replaces the artificial inspection operation for carrying out transmission line of electricity, for example, unmanned plane inspection and
Robot inspection.Inspection device by image transmitting by the inspection picture transmission of shooting into earth station, for patrol officer observe,
But the large nuber of images generated makes patrol officer too plenty for the eye to take it all in, so there has been image procossing to detect the defect of high-tension bus-bar.
There are mainly three types of the defects of high-tension bus-bar: stranded, corrosion and foreign matter.Inspection device mainly carries out figure to high-tension bus-bar
As acquisition, obtained image is complicated with image background, electric wire is low with the contrast of background, it is a series of to there is largely interference etc.
Problem.Earliest image detection algorithm is concentrated mainly on shape (edge or profile) and textural characteristics for target and is known
It indescribably takes, that is, utilizes morphological feature, also there are a small number of Processing Algorithms to extract the RGB color domain of target using thermal imaging to sentence
Disconnected defect situation, that is, utilize color science.The above method can reach final testing goal to a certain extent, but can only office
Limit in even a certain environment, and among practical application, is influenced, phase among specific a certain classification by environment
The influence and algorithm limitation itself, the shadow for handling recognition speed for the objective factors such as machine image quality, equipment moving be fuzzy
It rings, causes the validity and reliability to target detection all not ideal enough, and do not have scalability.
Become hot spot followed by the method identification transmission line part of machine learning.Traditional power components identification process
Be often used classical machine learning algorithm, such as support vector machines (SVM), random forest or adaboost, in conjunction with gradient, color or
The shallow-layers such as texture feature detects power components, but existing classical machine learning algorithm is to the effect of electric wire defect recognition
Rate and accuracy rate are all relatively low, it is difficult to make full use of the information of inspection image.
Summary of the invention
In view of the above-mentioned problems, the present invention is provided using the convolutional neural networks algorithm (Faster R-CNN) based on region
A kind of high efficiency, the high-tension bus-bar defect image detection technique of high accuracy.Utilize the high pressure of actual acquisition and making in laboratory
Electric wire defect image constructs sample set, carries out learning training using deep learning algorithm and establishes electric wire defects detection model, and answers
For in actual scene.
To achieve the above object, the technical solution adopted by the present invention is that including the following steps:
(1) the high-tension bus-bar image for acquiring various different defect classifications, by various different high-tension bus-bar defect classifications and
Have been marked with the high-tension bus-bar image composition high-tension bus-bar image deflects sample set of high-tension bus-bar defect classification and defective locations;
The defect classification includes three kinds of situations of stranded corrosion and foreign matter.
Defective locations are the image-region where a defect, and image-region edge pixel point is boundary.
(2) each panel height piezoelectricity line image in high-tension bus-bar image deflects sample set is input to convolutional neural networks
(CNN), the defects of high-tension bus-bar image feature is extracted to obtain characteristic pattern (Feature Map), characteristic pattern is subsequent
RPN network internal and full articulamentum it is shared;
Convolutional neural networks (CNN) use ZFnet neural network model.
(3) characteristic pattern is input in region proposal network (RPN) for having additional full articulamentum (FC), net is proposed by region
Network proposing offers region (region proposal) simultaneously judges to propose that region belongs to prospect or background;Prospect will be belonged to again
Propose that region and characteristic pattern input together and carries out defect classification in full articulamentum (FC);It is mentioned followed by logistic regression correcting process
The boundary in region is discussed, constantly optimizes boundary and obtains more accurate proposal region;
(4) step (3) are repeated, network (RPN), which is iterated update, is proposed to region, learning rate is varied multiple times until maximum
The number of iterations completes training, obtains the parameter in trained region proposal network (RPN) as high-tension bus-bar defects detection mould
Type;
(5) defects detection is carried out to actual high-tension bus-bar image to be measured using high-tension bus-bar defects detection model.
In the step (1), defect classification includes three kinds of situations of stranded corrosion and foreign matter.
Since the defects of actual environment is not easy to be arrived by eye-observation, just simulation electric wire defect makes sample collection in an experiment
To increase sample size, reduced under experimental situation as possible with the background subtraction of the defect image under actual environment away from therefore can will
Actual photographed and the sample of production mix, as trained sample set.
The ZFnet neural network model is specially 8 layers of sequentially connected network structure, and 8 layers are 5 convolutional layers and 3
A full articulamentum, 5 convolutional layers and 3 full articulamentums are sequentially connected composition.
The step (3) specifically:
Region proposes that network (RPN) includes classification layer, full articulamentum and logistic regression layer, and full articulamentum setting is connected to point
Between class layer and logistic regression layer;
Characteristic pattern is input to classification layer proposing offers region first, classification layer uses normalized function (softmax), point
Class layer, which judges to obtain simultaneously, proposes that region belongs to prospect or background, that is, judges whether the defect of detection image mesohigh electric wire deposits
?;
Then the proposal region for belonging to prospect and characteristic pattern are input in full articulamentum to the proposal region for belonging to prospect
Classify, i.e., classifies to the defect of high-tension bus-bar;
Finally sorted result is input in logistic regression layer, logistic regression layer uses linear regression, returns in logic
Return mark anchor point bounding box, anchor point bounding box in layer to be determined by four anchor points, predicts to obtain electric wire defect using anchor point bounding box
Predicted boundary frame, make propose region constantly returns the predicted boundary frame after optimize, with cross entropy cost function L calculating in advance
The error amount between bounding box and real border frame is surveyed, then error amount is carried out using stochastic gradient descent optimizer (SGD)
Backpropagation calculates, and is constantly changing network parameter values λ and is returned using coordinate parametersization optimization bounding box, so that error amount is protected
It holds within threshold value, using finally obtained proposal region as more accurate proposal area results.
Boundary is the figure where the defects of each panel height piezoelectricity line image of high-tension bus-bar image deflects sample set position
As zone boundary.
The cross entropy cost function L specifically:
In formula, anchor point bounding box, there are three types of area and three kinds of frame ratios altogether, and predict electric wire defective bit as first time
The reference bounding box set, in specific implementation, three kinds of areas are 128,256,512, and three kinds of ratios are: { 1:2 }, { 1:1 }, { 2:1 },
The width of bounding box is higher than bounding box.I is the ordinal number of anchor point bounding box, and pi is prediction probability of the anchor point i as target;It is
The true tag value of anchor point i, if anchor point is prospect, true tag valueIt is 1, if anchor point is background, true tag
ValueIt is 0;tiIt is the vector for indicating 4 parametrization coordinates of predicted boundary frame,Be in the proposal region for belonging to prospect
The vector of the relevant real border frame of anchor point;λ indicates network parameter values;Real border frame is that the electric wire of the determination marked lacks
Fall into frame.NclsIndicate that the quantity of total anchor point takes 128, N in trainingregIndicate that the size of characteristic pattern Feature Map is being instructed
2400 are taken when practicing;Presentation class loss function is the logarithm of (defect is prospect or background) in two classifications
Loss calculates are as follows:
It indicates to return loss function, calculate are as follows:
Wherein, R is smooth loss function, is calculated are as follows:
Wherein, r indicates the difference of prediction probability and true tag value.The output that the corresponding classification of λ and frame return separately includes
{piAnd { ti, { piIndicate set of the anchor point i as the prediction probability of target, { tiIndicate anchor point i predicted boundary frame ginseng
The vector set of numberization coordinate, this two pass through { NclsAnd { NregBe normalized, { NclsAnd { NregRespectively indicate parameter
NclsAnd NregVector value, and balance weight is added --- network parameter values λ.Before expression only belongs to
The defect anchor point of scapeReturning loss can just be activated, if it is background defect anchor pointIts value is then 0.
The step (5) specifically: acquire high-tension bus-bar image to be detected using inspection device, which is inputted
The classification of detection output electric wire defect and position into high-tension bus-bar defects detection model.
In the step (3), returned using coordinate parametersization optimization bounding box specifically:
Wherein, the centre coordinate and its width and height of the bounding box of x, y, w and h expression electric wire defect, variable x, xaAnd x*Respectively
Indicate predicted boundary frame, anchor point bounding box and real border frame;txIndicate the vector of the central point lateral coordinates of predicted boundary frame;Indicate and belong to prospect proposal region anchor point real border frame central point lateral coordinates vector;tyIndicate prediction
The vector of the central point lateral coordinates of bounding box;It indicates and belongs in the real border frame for the anchor point for proposing region of prospect
The vector of heart point longitudinal coordinate;twIndicate the wide vector of the central point of predicted boundary frame;Indicate and belong to the proposal area of prospect
The wide vector of the central point point of the real border frame of the anchor point in domain;thIndicate the high vector of the central point of predicted boundary frame;Table
Show and belong to the high vector of the central point point of the real border frame of the anchor point in the proposal region of prospect.
The beneficial effects of the present invention are:
The pond layer of Faster R-CNN is integrated into RPN network and full articulamentum is added by the present invention.In new RPN net
In network, the characteristic pattern for directly generating the prospect combination CNN judged is inputted in full articulamentum, carries out point of electric wire defect
Class simplifies trained step in the case where guaranteeing training quality.
Show to be feasible using Faster R-CNN detection high-tension bus-bar defect after simplification in the actual operation process,
And simplify algorithm compared to former algorithm using this, it is significantly shortened on the time of detection, carries out the inspection of high-tension bus-bar defect
Survey can achieve higher recognition speed and high accuracy rate and have preferable robustness.
The present invention can help patrol officer in the process for carrying out polling transmission line operation using inspection device, with most fast
Velocity measuring go out high-tension bus-bar defect, the effect of the real-time detection reached has been significantly increased routing inspection efficiency, has shortened figure
As the time of processing.
Detailed description of the invention
Fig. 1 is that the present invention is based on the resume module processes of the high-tension bus-bar defect inspection method of the convolutional neural networks in region
Figure;
Fig. 2 is the specific flow chart of Faster R-CNN of the present invention training;
Fig. 3 is the processing result exemplary diagram of high-tension bus-bar defect in practical applications of the invention.
Specific embodiment
Since the bibliography of the convolutional neural networks algorithm based on region is nearly all English, in order to avoid because of translation
And the problem of causing ambiguity, Chinese English english abbreviation control is indicated hereby.
The invention will be further described with embodiment with reference to the accompanying drawing.
The library Tensorflow SLIM that the present invention is based on open source to the training of network and the detection of test sample carries out.
Tensorflow is a clear and efficient deep learning frame, and readable, terseness and performance are all very outstanding, and straight
It connects and is integrated with convolutional network nervous layer.Due to the characteristic of depth convolutional network itself, accelerate operation that can greatly shorten calculation with GPU
The method training time.Tensorflow also provides corresponding interface.Complete training process will be illustrated in conjunction with Fig. 2.
In embodiment, the environment of deep learning: central processing unit is for i7-5960X, video card GeForce to the present invention
GTX 1080Ti (video memory GDDR5 11G).
The embodiment of the present invention and its implementation process are:
(1) for three kinds of situations such as stranded, corrosion and foreign matters, high-tension bus-bar defect sample collection is made, by being manually electric wire
Defect labels, and marks electric wire defect classification and position.Since the defects of actual environment is not easy to be arrived by eye-observation, just in reality
It tests in room to simulate electric wire defect and make sample and collect to increase sample size, reduced under laboratory environment and under actual environment as possible
The background subtraction of defect image is away from therefore can mixing actual photographed and the sample of production, as trained sample set;
Photo size used in the embodiment of the present invention is 1920*1080.
(2) the high-tension bus-bar image that each width in sample set has marked is input to convolutional neural networks (CNN) network,
The electric wire defect characteristic figure (Feature Map) in image is extracted using ZFnet network model, this feature figure will be by
The sharing feature of subsequent RPN network and full articulamentum (FC);
(3) characteristic pattern extracted with previous step initialization RPN network formation zone is proposed that classification layer passes through normalization letter
(softmax) is counted to judge that region proposes to belong to prospect or background, i.e., judgement detection target electric wire defect whether there is, later
Prospect and characteristic pattern are input in full articulamentum, foreground classification is carried out, i.e., classifies to electric wire defect;Utilize logistic regression
Make region propose constantly to return, finally obtain bounding box, with the cross entropy cost function L bounding box being calculated and actual side
Then the error amount of boundary's frame carries out backpropagation calculating to error amount using stochastic gradient descent optimizer (SGD), constantly changes
Become network parameter values λ and is returned using coordinate parametersization optimization bounding box, so that error amount is maintained within defined threshold value, it will
Network parameter values λ record.By above-mentioned calculating, so that it is obtained more accurate region and propose;
(4) step (3) are repeated and update is iterated to RPN network, learning rate is varied multiple times until maximum number of iterations is complete
At training, acquisition trains later parameter, finally exports a high-tension bus-bar defects detection model;
During specific implementation, training iteration 200000 times, 0~50000 learning rate position 0.0002,50000~
It is 0.000002 that 75000 learning rates, which are 0.00002,75000~200000 learning rates, obtains network through repeatedly optimizing and revising
Parameter lambda=10 (hyper parameter is rule of thumb set), frame of the frame that detected at this time closest to real marking, detection effect
Most preferably.After training, the parameter of deep neural network is saved.
(5) high-tension bus-bar image is acquired using inspection device, is input to high-tension bus-bar defect for the image as detection collection
In detection model, after testing, classification and the position of electric wire defect are marked, as shown in Figure 3.
In embodiment, what inspection device took is video to the present invention, the video processing speed in Faster R-CNN
It is faster than image procossing.For conventional Faster R-CNN algorithm, improved Faster R-CNN algorithm is in accuracy of identification
On, it is consistent with former algorithm;In recognition speed, fast 30%, it can achieve 15FPS.
Claims (6)
1. a kind of high-tension bus-bar defect inspection method of the convolutional neural networks algorithm based on region, it is characterized in that: including as follows
Step:
(1) by various different high-tension bus-bar defect classifications and have been marked with the high pressure of high-tension bus-bar defect classification and defective locations
Electric wire image forms high-tension bus-bar image deflects sample set;
(2) each panel height piezoelectricity line image in high-tension bus-bar image deflects sample set is input to convolutional neural networks (CNN),
The defects of high-tension bus-bar image feature is extracted to obtain characteristic pattern (Feature Map);
(3) characteristic pattern is input in region proposal network (RPN) for having additional full articulamentum (FC), proposes that network is raw by region
At proposal region (region proposal) and judge to propose that region belongs to prospect or background;The proposal of prospect will be belonged to again
Region and characteristic pattern input together carries out defect classification in full articulamentum (FC);Then correcting process proposes the boundary in region, no
Optimize boundary disconnectedly and obtains more accurate proposal region;
(4) step (3) are repeated, network (RPN), which is iterated update, is proposed to region, learning rate is varied multiple times until greatest iteration
Number completes training, obtains the parameter in trained region proposal network (RPN) as high-tension bus-bar defects detection model;
(5) defects detection is carried out to actual high-tension bus-bar image to be measured using high-tension bus-bar defects detection model.
2. the high-tension bus-bar defect inspection method of the convolutional neural networks algorithm according to claim 1 based on region,
Be characterized in that: in the step (1), defect classification includes three kinds of situations of stranded corrosion and foreign matter.
3. the high-tension bus-bar defect inspection method of the convolutional neural networks algorithm according to claim 1 based on region,
Be characterized in that: the ZFnet neural network model is specially 8 layers of sequentially connected network structure, and 8 layers are 5 convolutional layers and 3
A full articulamentum, 5 convolutional layers and 3 full articulamentums are sequentially connected composition.
4. the high-tension bus-bar defect inspection method of the convolutional neural networks algorithm according to claim 1 based on region,
It is characterized in that: the step (3) specifically: propose that network (RPN) includes classification layer, full articulamentum and logistic regression in region
Layer, full articulamentum setting are connected between classification layer and logistic regression layer;Characteristic pattern is input to classification layer proposing offers first
Region, classification layer use normalized function (softmax), and classification layer, which judges to obtain simultaneously, proposes that region belongs to prospect and still carries on the back
Scape;Then the proposal region for belonging to prospect and characteristic pattern are input in full articulamentum and the proposal region for belonging to prospect is divided
Class classifies to the defect of high-tension bus-bar;Finally sorted result is input in logistic regression layer, logistic regression layer
Using linear regression, anchor point bounding box is marked in logistic regression layer, anchor point bounding box is determined by four anchor points, utilizes anchor point side
Frame prediction in boundary's obtains the predicted boundary frame of electric wire defect, makes to propose that region constantly returns the predicted boundary frame after being optimized, use
Cross entropy cost function L calculates the error amount between predicted boundary frame and real border frame, then excellent using stochastic gradient descent
Change device (SGD) and backpropagation calculating is carried out to error amount, be constantly changing network parameter values λ and optimize side using coordinate parametersization
Boundary's frame returns, so that error amount is maintained within threshold value, using finally obtained proposal region as more accurate proposal region
As a result.
5. the high-tension bus-bar defect inspection method of the convolutional neural networks algorithm according to claim 4 based on region,
It is characterized in that:
The cross entropy cost function L specifically:
In formula, i is the ordinal number of anchor point bounding box, piIt is prediction probability of the anchor point i as target;It is the true tag of anchor point i
Value;tiIt is the vector for indicating 4 parametrization coordinates of predicted boundary frame,The anchor point phase proposed in region for being and belonging to prospect
The vector of the real border frame of pass;λ indicates network parameter values;NclsIndicate that the quantity of total anchor point takes 128, N in trainingregTable
Show that the size of characteristic pattern Feature Map takes 2400 in training;Presentation class loss function calculates are as follows:
It indicates to return loss function, calculate are as follows:
Wherein, R is smooth loss function, is calculated are as follows:
Wherein, r indicates the difference of prediction probability and true tag value.
6. the high-tension bus-bar defect inspection method of the convolutional neural networks algorithm according to claim 1 based on region,
It is characterized in that: the step (5) specifically: high-tension bus-bar image to be detected is acquired using inspection device, the image is defeated
Enter into high-tension bus-bar defects detection model classification and the position of detection output electric wire defect.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910122569.8A CN109886947A (en) | 2019-02-19 | 2019-02-19 | The high-tension bus-bar defect inspection method of convolutional neural networks based on region |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910122569.8A CN109886947A (en) | 2019-02-19 | 2019-02-19 | The high-tension bus-bar defect inspection method of convolutional neural networks based on region |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109886947A true CN109886947A (en) | 2019-06-14 |
Family
ID=66928516
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910122569.8A Pending CN109886947A (en) | 2019-02-19 | 2019-02-19 | The high-tension bus-bar defect inspection method of convolutional neural networks based on region |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109886947A (en) |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110232687A (en) * | 2019-06-21 | 2019-09-13 | 华北电力大学(保定) | The detection method of cotter bolt defect in a kind of electric inspection process image |
CN110555842A (en) * | 2019-09-10 | 2019-12-10 | 太原科技大学 | Silicon wafer image defect detection method based on anchor point set optimization |
CN111340787A (en) * | 2020-02-28 | 2020-06-26 | 广东电网有限责任公司 | Method and device for detecting and identifying defects of conducting wire of power transmission line and computer equipment |
CN111402248A (en) * | 2020-03-23 | 2020-07-10 | 华南理工大学 | Transmission line lead defect detection method based on machine vision |
CN111798447A (en) * | 2020-07-18 | 2020-10-20 | 太原理工大学 | Deep learning plasticized material defect detection method based on fast RCNN |
CN111860439A (en) * | 2020-07-31 | 2020-10-30 | 广东电网有限责任公司 | Unmanned aerial vehicle inspection image defect detection method, system and equipment |
CN112016449A (en) * | 2020-08-27 | 2020-12-01 | 西华大学 | Vehicle identification and detection method based on deep learning |
CN112288694A (en) * | 2020-10-19 | 2021-01-29 | 武汉大学 | Mask region convolution neural network-based power transformation equipment defect identification method |
CN113112446A (en) * | 2020-03-05 | 2021-07-13 | 成都理工大学 | Tunnel surrounding rock level intelligent judgment method based on residual convolutional neural network |
CN114972225A (en) * | 2022-05-16 | 2022-08-30 | 上海可明科技有限公司 | Two-stage photovoltaic panel defect detection method based on deep learning |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN201421435Y (en) * | 2009-04-14 | 2010-03-10 | 山东电力研究院 | Video detector of power transmission line inspection robot |
CN107274451A (en) * | 2017-05-17 | 2017-10-20 | 北京工业大学 | Isolator detecting method and device based on shared convolutional neural networks |
US20180158189A1 (en) * | 2016-12-07 | 2018-06-07 | Samsung Electronics Co., Ltd. | System and method for a deep learning machine for object detection |
-
2019
- 2019-02-19 CN CN201910122569.8A patent/CN109886947A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN201421435Y (en) * | 2009-04-14 | 2010-03-10 | 山东电力研究院 | Video detector of power transmission line inspection robot |
US20180158189A1 (en) * | 2016-12-07 | 2018-06-07 | Samsung Electronics Co., Ltd. | System and method for a deep learning machine for object detection |
CN107274451A (en) * | 2017-05-17 | 2017-10-20 | 北京工业大学 | Isolator detecting method and device based on shared convolutional neural networks |
Non-Patent Citations (1)
Title |
---|
WEISHAN ZHANG等: "RCNN-based foreign object detection for securing power transmission lines", 《PROCEDIA COMPUTER SCIENCE》 * |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110232687A (en) * | 2019-06-21 | 2019-09-13 | 华北电力大学(保定) | The detection method of cotter bolt defect in a kind of electric inspection process image |
CN110232687B (en) * | 2019-06-21 | 2022-04-12 | 华北电力大学(保定) | Method for detecting defects of bolts with pins in power inspection image |
CN110555842A (en) * | 2019-09-10 | 2019-12-10 | 太原科技大学 | Silicon wafer image defect detection method based on anchor point set optimization |
CN111340787A (en) * | 2020-02-28 | 2020-06-26 | 广东电网有限责任公司 | Method and device for detecting and identifying defects of conducting wire of power transmission line and computer equipment |
CN111340787B (en) * | 2020-02-28 | 2023-09-29 | 广东电网有限责任公司 | Method and device for detecting and identifying wire defects of power transmission line and computer equipment |
CN113112446A (en) * | 2020-03-05 | 2021-07-13 | 成都理工大学 | Tunnel surrounding rock level intelligent judgment method based on residual convolutional neural network |
CN111402248A (en) * | 2020-03-23 | 2020-07-10 | 华南理工大学 | Transmission line lead defect detection method based on machine vision |
CN111402248B (en) * | 2020-03-23 | 2023-05-09 | 华南理工大学 | Power transmission line wire defect detection method based on machine vision |
CN111798447B (en) * | 2020-07-18 | 2023-03-10 | 太原理工大学 | Deep learning plasticized material defect detection method based on fast RCNN |
CN111798447A (en) * | 2020-07-18 | 2020-10-20 | 太原理工大学 | Deep learning plasticized material defect detection method based on fast RCNN |
CN111860439A (en) * | 2020-07-31 | 2020-10-30 | 广东电网有限责任公司 | Unmanned aerial vehicle inspection image defect detection method, system and equipment |
CN112016449A (en) * | 2020-08-27 | 2020-12-01 | 西华大学 | Vehicle identification and detection method based on deep learning |
CN112288694A (en) * | 2020-10-19 | 2021-01-29 | 武汉大学 | Mask region convolution neural network-based power transformation equipment defect identification method |
CN112288694B (en) * | 2020-10-19 | 2022-10-04 | 武汉大学 | Method for identifying defects of power transformation equipment based on mask region convolution neural network |
CN114972225A (en) * | 2022-05-16 | 2022-08-30 | 上海可明科技有限公司 | Two-stage photovoltaic panel defect detection method based on deep learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109886947A (en) | The high-tension bus-bar defect inspection method of convolutional neural networks based on region | |
CN109948425B (en) | Pedestrian searching method and device for structure-aware self-attention and online instance aggregation matching | |
CN113160192B (en) | Visual sense-based snow pressing vehicle appearance defect detection method and device under complex background | |
CN107092870B (en) | A kind of high resolution image Semantic features extraction method | |
CN108021947B (en) | A kind of layering extreme learning machine target identification method of view-based access control model | |
CN109635875A (en) | A kind of end-to-end network interface detection method based on deep learning | |
CN108830188A (en) | Vehicle checking method based on deep learning | |
CN110222792A (en) | A kind of label defects detection algorithm based on twin network | |
CN107016413B (en) | A kind of online stage division of tobacco leaf based on deep learning algorithm | |
CN107423760A (en) | Based on pre-segmentation and the deep learning object detection method returned | |
CN114299380A (en) | Remote sensing image semantic segmentation model training method and device for contrast consistency learning | |
CN104992223A (en) | Dense population estimation method based on deep learning | |
CN109284779A (en) | Object detecting method based on the full convolutional network of depth | |
CN111209832B (en) | Auxiliary obstacle avoidance training method, equipment and medium for substation inspection robot | |
CN109214308A (en) | A kind of traffic abnormity image identification method based on focal loss function | |
CN108537222A (en) | A kind of image-recognizing method and system for electric instrument | |
CN108492298A (en) | Based on the multispectral image change detecting method for generating confrontation network | |
CN108961308A (en) | A kind of residual error depth characteristic method for tracking target of drift detection | |
CN113435282B (en) | Unmanned aerial vehicle image ear recognition method based on deep learning | |
CN109886356A (en) | A kind of target tracking method based on three branch's neural networks | |
CN110472572A (en) | The quick identification and classification method of naval target under a kind of complex environment | |
CN106650823A (en) | Probability extreme learning machine integration-based foam nickel surface defect classification method | |
CN114049477A (en) | Fish passing fishway system and dynamic identification and tracking method for fish quantity and fish type | |
CN109543585A (en) | Underwater optics object detection and recognition method based on convolutional neural networks | |
CN112347930B (en) | High-resolution image scene classification method based on self-learning semi-supervised deep neural network |
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 | ||
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190614 |