CN110378408A - Power equipment image-recognizing method and device based on transfer learning and neural network - Google Patents
Power equipment image-recognizing method and device based on transfer learning and neural network Download PDFInfo
- Publication number
- CN110378408A CN110378408A CN201910633986.9A CN201910633986A CN110378408A CN 110378408 A CN110378408 A CN 110378408A CN 201910633986 A CN201910633986 A CN 201910633986A CN 110378408 A CN110378408 A CN 110378408A
- Authority
- CN
- China
- Prior art keywords
- power equipment
- image
- model
- target power
- training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
Abstract
The embodiment of the present application discloses a kind of power equipment image-recognizing method and device based on transfer learning and neural network, this method comprises: obtaining the simulated image data collection of target power equipment as the first training sample, simulated image data collection includes the analog image and corresponding tag along sort of target power equipment, and the model of each type of target power equipment is multiple；The actual image data collection of target power equipment is obtained as the second training sample, actual image data collection includes the real image and corresponding tag along sort of target power equipment；Using the convolutional neural networks model that the training of the first training sample constructs in advance, first object convolutional neural networks model is obtained；The second target convolution neural network model is obtained using the second training sample training first object convolutional neural networks model；The image of power equipment to be identified is input to the second target convolution neural network model to identify.Improve the accuracy of power equipment image recognition.
Description
Technical field
The invention relates to the image processing techniques of field of electrical equipment, more particularly to it is a kind of based on transfer learning and
The power equipment image-recognizing method and device of neural network.
Background technique
With the propulsion that the development of China's electric utility and smart grid are built, power train is unified line production unit and is just used
The methods of a variety of monitoring means, such as unmanned plane Image Acquisition, video monitoring, infrared thermal imaging, which are adopted, assists staff's completion defeated
The inspection of electric line.These methods are mentioned by the way that the equipment image information such as transformer, electric power line steel tower is passed to patrol officer
Its high working efficiency, also ensures its operation safety.
But these image datas are only passed to operation maintenance personnel by current monitoring means, it is still desirable to by artificial
Image is re-started classification and identification otherwise by Zhen, is then further diagnosed again.And during this, to image
The energy that simple classification wastes a large amount of technical staff is carried out, labor efficiency is also reduced.
Consequent is automated intelligent to power equipment progress image recognition, for example, being realized using deep learning, so
And in deep learning frame, the accuracy rate that computer completes image recognition tasks depends not only upon the property of deep learning model
Can, also depend on the quality and quantity of training sample.Using polling transmission line as application background under computer picture know
In other task, the quality and quantity of training sample is all difficult to ensure.In this way, will lead to power equipment image recognition inaccuracy.
Summary of the invention
This application provides a kind of power equipment image-recognizing method and device based on transfer learning and neural network, with
It solves to be not allowed in power equipment image recognition processes in the prior art since sample size and quality not can guarantee bring identification
True problem.
The present invention adopts the following technical scheme:
In a first aspect, the embodiment of the present application is provided and a kind of is known based on transfer learning and the power equipment image of neural network
Other method, this method comprises:
The simulated image data collection of target power equipment is obtained as the first training sample, wherein the analog image number
Include the analog image and corresponding tag along sort of target power equipment according to collection, the type of the target power equipment be it is multiple,
The model of each type of target power equipment is multiple；
The actual image data collection of the target power equipment is obtained as the second training sample, wherein the practical figure
As data set includes the real image and corresponding tag along sort of target power equipment；
Using the convolutional neural networks model that first training sample training constructs in advance, first object convolution mind is obtained
Through network model；
The second target convolutional Neural is obtained using the second training sample training first object convolutional neural networks model
Network model；
The image of power equipment to be identified is input to the second target convolution neural network model to identify,
In, the type of the power equipment to be identified belongs to the one or more of the target power equipment.
Second aspect, the embodiment of the present application provide a kind of based on transfer learning and the knowledge of the power equipment image of neural network
Other device, the device include:
First sample obtains module, for obtaining the simulated image data collection of target power equipment as the first training sample
This, wherein the simulated image data collection includes the analog image and corresponding tag along sort of target power equipment, the target
The type of power equipment be it is multiple, the model of each type of target power equipment is multiple；
Second sample acquisition module, for obtaining the actual image data collection of the target power equipment as the second training
Sample, wherein the actual image data collection includes the real image and corresponding tag along sort of target power equipment；
First training module trains the convolutional neural networks model constructed in advance for application first training sample,
Obtain first object convolutional neural networks model；
Second training module is obtained for application the second training sample training first object convolutional neural networks model
Second target convolution neural network model；
Picture recognition module, for the image of power equipment to be identified to be input to the second target convolutional neural networks
Model is identified, wherein the type of the power equipment to be identified belongs to the one or more of the target power equipment.
It has the advantages that in the technical solution adopted by the present invention: being matched by analog image and real image, both
Consider a large amount of training samples in analog image includes a variety of situations, then passes through transfer learning again, it is contemplated that actual
On the other hand environmental characteristic under application environment improves processing accuracy in this way, on the one hand improving processing speed, effectively solve
Certainly two days electric power equipment drawings are as identifying inaccurate problem since sample size and quality not can guarantee bring in identification process.
Detailed description of the invention
By reading a detailed description of non-restrictive embodiments in the light of the attached drawings below, the application's is other
Feature, objects and advantages will become more apparent upon:
Fig. 1 is a kind of power equipment image recognition side based on transfer learning and neural network provided by the embodiments of the present application
Method flow chart；
Fig. 2 is a kind of analog image for the transformer being applicable in the embodiment of the present application；
Fig. 3 is a kind of analog image for the insulator being applicable in the embodiment of the present application；
Fig. 4 is a kind of analog image for the shaft tower being applicable in the embodiment of the present application；
Fig. 5 is a kind of analog image for the disconnecting switch being applicable in the embodiment of the present application；
Fig. 6 is a kind of real image for the transformer being applicable in the embodiment of the present application；
Fig. 7 is a kind of real image for the insulator being applicable in the embodiment of the present application；
Fig. 8 is a kind of real image for the shaft tower being applicable in the embodiment of the present application；
Fig. 9 is a kind of real image for the disconnecting switch being applicable in the embodiment of the present application；
Figure 10 is the training process mould of the VGG-16 under a kind of power equipment based on simulation being applicable in the embodiment of the present application
The schematic diagram of the quasi- classification accuracy to test set；
Figure 11 is a kind of feature vector change procedure of the VGG-16 applicable in the embodiment of the present application when identifying image
Schematic diagram；
Figure 12 is feature vector change procedure of another VGG-16 applicable in the embodiment of the present application when identifying image
Schematic diagram；
Figure 13 is a kind of schematic diagram for the process by transfer learning training VGG-16 being applicable in the embodiment of the present application；
Figure 14 is a kind of T-SNE image for the simulation power equipment image being applicable in the embodiment of the present application；
Figure 15 is a kind of T-NSE image for the practical power equipment image being applicable in the embodiment of the present application；
Figure 16 is a kind of schematic diagram of the training for the VGG-16+Softmax being applicable in the embodiment of the present application；
Figure 17 is a kind of VGG-16 model structure being applicable in the embodiment of the present application；
Figure 18 is a kind of power equipment image recognition based on transfer learning and neural network provided by the embodiments of the present application
The structural schematic diagram of device.
Specific embodiment
It is specifically real to the application with reference to the accompanying drawing in order to keep the purposes, technical schemes and advantages of the application clearer
Example is applied to be described in further detail.It is understood that specific embodiment described herein is used only for explaining the application,
Rather than the restriction to the application.It also should be noted that illustrating only for ease of description, in attached drawing related to the application
Part rather than full content.It should be mentioned that some exemplary realities before exemplary embodiment is discussed in greater detail
It applies example and is described as the processing or method described as flow chart.Although operations (or step) are described as sequence by flow chart
Processing, but many of these operations can be implemented concurrently, concomitantly or simultaneously.In addition, the sequence of operations
It can be rearranged.The processing can be terminated when its operations are completed, be not included in attached drawing it is also possible to have
Additional step.The processing can correspond to method, function, regulation, subroutine, subprogram etc..
Fig. 1 gives provided by the embodiments of the present application a kind of based on transfer learning and the knowledge of the power equipment image of neural network
The flow chart of other method, the power equipment image-recognizing method provided in this embodiment based on transfer learning and neural network can be with
By executing based on the power equipment pattern recognition device of transfer learning and neural network, transfer learning and neural network should be based on
Power equipment pattern recognition device can be realized by way of hardware and/or software.With reference to Fig. 1, this method specifically be can wrap
It includes:
S101, the simulated image data collection of target power equipment is obtained as the first training sample, wherein the simulation drawing
As data set includes the analog image and corresponding tag along sort of target power equipment, the type of the target power equipment is more
A, the model of each type of target power equipment is multiple.
Wherein, convolutional neural networks (Convolutional Neural Networks, CNN) are a kind of comprising convolution meter
The feedforward neural network for calculating and having depth structure, is one of representative algorithm of deep learning.Convolutional neural networks have characterization
Learning ability can carry out translation invariant classification to input information by its hierarchical structure, therefore also referred to as " translation invariant is artificial
Neural network ".The training result of CNN is largely related with the quality of training set and quantity, in utilization CNN model to electric power
When the image of equipment carries out classification based training, since there is no the data set disclosed in one based on power equipment, therefore this Shen
Please in need simulate establish a simulation power equipment data set, that is, simulated image data collection.
Specifically, obtaining the simulated image data collection of target power equipment as the first training sample, wherein analog image
Data set includes the analog image and corresponding tag along sort of target power equipment, and the type of the target power equipment is more
A, the model of each type of target power equipment is multiple.In a specific example, the type of target power equipment can
To be transformer, insulator, shaft tower and disconnecting switch.Corresponding tag along sort can be 1,2,3 and 4.In actual application
Cheng Zhong, each target power equipment can also include different models, the analog image of the target power equipment got in this way
Data set is also different.Fig. 2 shows a kind of analog image of transformer, Fig. 3 shows a kind of analog image of insulator, Fig. 4
A kind of analog image of shaft tower is shown, Fig. 5 shows a kind of analog image of disconnecting switch.
S102, the actual image data collection of the target power equipment is obtained as the second training sample, wherein the reality
Border image data set includes the real image and corresponding tag along sort of target power equipment.
The information such as color, background and noise due to simulation power equipment data set and real image have differences, and CNN is only
By learning to simulation power equipment image, the accuracy rate directly applied to when identifying to real image is lower.And
Currently without disclosed power equipment data set, therefore, small, high quality a practical power equipment number is pre-established here
According to collection, the CNN model by simulation power equipment data set training is trained again by way of transfer learning, to reach
Allow CNN that can accurately identify that practical electric power is set under the premise of not needing and establishing extensive practical power equipment image data set
Standby image.
In practical applications, power equipment image is often obtained by specific equipment such as unmanned plane, infrared thermal imaging camera shooting etc.,
These particular devices, which obtain image, often has its exclusive feature.For example, unmanned plane obtain it is mostly be to take a crane shot at a distance, and it is infrared
Imaging indicates the heat radiation characteristic of object by strong color contrast.If directly will be according to the power equipment collected, arranged
The CNN that image trains is applied in the identification to such image, it tends to be difficult to obtain preferable effect.An and line production unit root
According to specific power equipment image establish a high quality, large sample data set specialized training needed for CNN model this usually
It is more difficult.Therefore, the embodiment of the present application proposes by way of introducing transfer learning, will be by simulation power equipment data set
Trained CNN model extraction comes out, and is instructed by the specific power equipment image data set of a small sample, high quality to it
Practice, realize CNN to the specific power equipment image study of base's production unit and identification, and then applies ring actual
Reach higher accuracy of identification under border.
Specifically, obtain the actual image data collection of target power equipment that pre-establishes as the second training sample,
In, target power equipment is consistent with the classification for the power equipment that simulated image data is concentrated and model.In a specific example
In, the type of target power equipment can be transformer, insulator, shaft tower and disconnecting switch.Corresponding tag along sort can be
1,2,3 and 4.Illustratively, Fig. 6 shows a kind of real image of transformer, and Fig. 7 shows a kind of practical figure of insulator
Picture, Fig. 8 show a kind of real image of shaft tower, and Fig. 9 shows a kind of real image of disconnecting switch.
S103, the convolutional neural networks model constructed in advance using first training sample training, obtain first object
Convolutional neural networks model.
Specifically, being identified simultaneously to four kinds of common power equipments as target power equipment in the embodiment of the present application
Classification is tested.When test, 1500 width of at random concentrating the simulated image data of target power equipment as training set,
500 width are as test set, and experiment can be used the processor of Core i7-6700k, 16GB memory, GTX1070ti video card,
The convolutional neural networks prototype network that training constructs in advance under Caffe frame.Optionally, the convolutional Neural net constructed in advance
Network model is VGG-16 model；VGG-16 model includes convolutional layer, pond layer, full articulamentum and pond layer.
VGG-16 model is improved in the embodiment of the present application, since in a specific example, target power is set
Standby classification quantity is 4, it is therefore proposed that replacing the last layer of VGG-16 to join by 128 full articulamentum using a parameter
Number be 1000 full articulamentum, to improve training speed, and in transfer learning reduce classifier classification difficulty.
Illustratively, the convolutional neural networks model constructed in advance using first training sample training, obtains first
Target convolution neural network model, comprising: network parameter is determined using stochastic gradient descent method；According to the network parameter, answer
The convolutional neural networks model constructed in advance with first training sample training, obtains first object convolutional neural networks mould
Type.
Specifically, when training CNN identification simulation power equipment image, using stochastic gradient descent method training CNN, by more
The optimal network parameter of secondary trial and error experimental selection, final CNN have been up to 92.6% in the identification to simulation electric power image
Accuracy rate.Figure 10 shows a kind of classification simulated based on the training process for simulating the VGG-16 under power equipment to test set
The schematic diagram of accuracy rate.In Figure 10, curve represents VGG16+Softmax training process, can be with according to the training process of Figure 10
It is very fast to obtain perception model convergence rate, accuracy rate just fluctuates in 89%~92.6% section after iteration 3000 times, this is because
Although the feature for simulating power equipment image is more simple, by addition different scenes picture as background, image is increased
Diversity, prevent parts of images is from being correctly validated, but this more meets practical application scene, in order to further appreciate that CNN
CNN model when choosing the 6000th iteration to each processing result image, in the embodiment of the present application sets 500 simulation electric power
Standby image is classified, and the results are shown in Table 1 for output.
1 VGG-16 of table is to simulation power equipment image recognition result
It can learn that trained VGG-16 model is relatively accurate according to table 1, since the training process of CNN is close to one
Black box uses the T-SNE tool pair in manifold learning to study influence of the CNN model to original input signal feature
It the high dimensional data dimensionality reduction of different levels and is visualized during model training, passes through this visual means and understand VGG-16 pairs
The treatment process of simulated image data collection.Figure 11 shows a kind of feature vector change procedure of VGG-16 when identifying image
Schematic diagram；Figure 12 shows the schematic diagram of feature vector change procedure of another VGG-16 when identifying image.
The CNN mould choosing 400 width images and completing using training is concentrated from the simulated image data of target power equipment at random
Type (first object convolutional neural networks model) is classified, and the last layer for exporting every kind of classification connects image entirely, is recycled
T-SNE tool carries out dimensionality reduction, as a result as is illustrated by figs. 11 and 12.The input of original image has had certain as can be seen from FIG. 12
Characteristic of division, but classify accuracy it is inadequate, it is still mixed in together by much classifying；It is shown in Figure 12 final complete
Classification situation after 1 × 64 vector dimensionality reduction of articulamentum, it can be found that just clearly, most similar samples are poly- for characteristic of division at this time
It gathers together, only some unrecognized samples are scattered in different classifications region.
It, can according to the visualization classification of the full articulamentum output of input picture and the last layer of the T-SNE to VGG-16 model
The extraction to characteristics of image can be completed to find out VGG-16 after overfitting, so that the characteristic of division contained in the picture is clear
It indicates in full articulamentum output vector, completes the classification task to image.
S104, the second target convolution is obtained using the second training sample training first object convolutional neural networks model
Neural network model.
Figure 13 shows a kind of schematic diagram of process by transfer learning training VGG-16, passes through mode as shown in figure 13
CNN is trained, it is preferable an accuracy of identification can be obtained under the premise of a small amount of, high quality power equipment image data set
CNN model.
Since the feature of practical power equipment image is frequently more various, Figure 14 shows a kind of simulation power equipment image
T-SNE image, Figure 15 shows a kind of T-NSE image of practical power equipment image, if Figure 15 is that the practical electric power in part is set
Standby image by T-SNE treated visual image, opposite Figure 14, classification difference is less obvious, and CNN learns its classification
Feature may be more difficult.
Therefore, when being trained by way of transfer learning to practical power equipment using VGG-16, the application is real
Applying example selection SVM classifier replaces Softmax classifier to connect output vector entirely to the last layer of VGG-16 and classify, into
And training speed is improved, obtain higher training precision.Figure 16 shows a kind of schematic diagram of the training of VGG-16+Softmax,
As shown in figure 16.The wherein identification process of 161 expression VGG-16+SVM, 162 indicate the identification process of VGG-16+Ssftmax,
Middle VGG-16+SVM is to the accuracy for having reached 93.5% in practical power equipment image recognition, compared to VGG-16 in mould electric power
Training process in equipment, faster, precision is higher for CNN convergence rate.This is because VGG-16 model have passed through based on simulation electricity
The training of power equipment image, has had certain recognition capability, and by transfer learning, it can be on practical power equipment data set
More rapid convergence.Moreover, the test set of practical power equipment is smaller, the diversity of the practical power equipment image of collection is relatively
It is few, therefore VGG-16 reaches higher accuracy of identification on practical electric power data collection.
S105, the image of power equipment to be identified is input to the second target convolution neural network model knows
Not, wherein the type of the power equipment to be identified belongs to the one or more of the target power equipment.
It is identified specifically, the image of power equipment to be identified is input to the second target convolution neural network model,
With the classification of determination power equipment to be identified, illustratively, the type of target power equipment belongs to one in target power equipment
Kind is a variety of.
It has the advantages that in the technical solution adopted by the present invention: being matched by analog image and real image, both
Consider a large amount of training samples in analog image includes a variety of situations, then passes through transfer learning again, it is contemplated that actual
On the other hand environmental characteristic under application environment improves processing accuracy in this way, on the one hand improving processing speed, effectively solve
Certainly two days electric power equipment drawings are as identifying inaccurate problem since sample size and quality not can guarantee bring in identification process.
Based on the above technical solution, the simulated image data collection of target power equipment is obtained as the first training sample
This, can be accomplished in that the 3D model for establishing target power equipment；Adjustment rendering background and/or the 3D model
Angle is to obtain target power equipment image；Export described image and corresponding tag along sort, described image and corresponding label
Simulated image data collection is formed as the first training sample.
Illustratively, in the embodiment of the present application, it is with four kinds of transformer, insulator, shaft tower and disconnecting switch distinct devices
Example, the Multiple Type 3D mould of four kinds of transformer, insulator, shaft tower and disconnecting switch distinct devices is established first with CAD software
Then type is obtained the power equipment image of a variety of patterns by modes such as replacement rendering background, adjustment model angles, finally exported
It renders image and its tag along sort establishes simulation power equipment data set, by analog image and the composition simulation of corresponding tag along sort
Image data set is as the first training sample.In the embodiment of the present application, by the power equipment 3D model of multiple models from difference
The rendering image of angle, different background is acquired, unified to select resolution ratio for 224 × 224.Altogether obtain transformer, insulator,
Totally 2000 width images are used as simulation power equipment image data set for shaft tower and each 500 width of disconnecting switch.
Based on the above technical solution, the actual image data collection of the target power equipment is obtained as the second instruction
Practice sample, can specifically be accomplished in that and shoot the target power equipment in the reality of different time, different location
Image；The real image and corresponding tag along sort group are combined into actual image data collection as the second training sample.
Specifically, different time can be the different moments in morning, noon or dusk or 24 hours, differently
Point can be different the different regions etc. under climatic environment.
In a specific embodiment, power equipment image to be identified is input to the second target convolution nerve net
Network model is identified, comprising: the image of the power equipment to be identified is input to the second target convolutional neural networks
Model；Identify the corresponding label respectively of the power equipment image to be identified；It is determined according to the tag along sort described wait know
Other power equipment.
Wherein, during identifying power equipment image to be identified, the image of power equipment to be identified is input to
Two target convolution neural network models identify that power equipment image correspondence to be identified divides tag along sort, such as 2 first, then may be used
To determine power equipment to be identified as insulator according to tag along sort.
In conclusion carrying out 3D modeling in the embodiment of the present application to insulator, shaft tower, transformer power equipment, passing through wash with watercolours
Dye generates the data set for containing a large amount of, high quality power equipment pictures greatly；Then using the data set to convolutional neural networks
It is trained, obtains preliminary disaggregated model；Finally, being allowed by the CNN of initial training using the method for transfer learning to high-quality
The practical power equipment data set of amount small sample is learnt, and power equipment identification model is obtained, this avoids virtual sample and
Interference between actual sample also reduces the learning difficulty of CNN.It finally, can by testing the CNN model of acquisition
Fast and accurately to classify to power equipment image.
Specifically, for there presently does not exist one for CNN study open altogether, the power equipment picture number of high quality
The case where according to collection, the embodiment of the present application carry out 3D modeling to multiple common power equipments, are rendered by changing background and angle
Obtain the simulation power equipment image data set under more scenes.VGG-16 model outstanding in current CNN is recycled to carry out it
Study makes CNN concentrate primary learning to " knowledge " to power equipment image recognition from simulation power equipment image data.Most
Afterwards, the practical power equipment number of low volume data will be applied to by way of transfer learning by the VGG-16 model of initial training
According in the study of collection, and achieve higher accuracy of identification.In addition, refreshing based on transfer learning and convolution in the embodiment of the present application
Not only solve the less problem of current power equipment image data set through the power equipment image-recognizing method under network, also for
One line production unit completes the specific equipment image recognition classification task of our unit using CNN and provides a kind of thinking.
In order to be easier to understand the technical solution of the application, below to the convolutional Neural net applied in the embodiment of the present application
Network is illustrated.
Deep learning has the cerebral nervous system of abundant hierarchical structure by simulation, establishes the layering mould similar to human brain
Type structure carries out feature extraction step by step to input data, so that forming abstract high-level characteristic indicates.Why it is known as depth
Study is to have broken artificial intelligence field up to number because deep learning network can learn the substantive characteristics to mass data
10 years situations for failing to have substantive breakthroughs have started an innovation sexual revolution in artificial intelligence field.
In deep learning algorithm, CNN is first multilayered structure learning algorithm truly, is closest at present
The intelligence learning method of human brain, it reduces number of parameters using spatial correlation, and the essence of CNN, which is that study is multiple, to be extracted
The characteristic filter device of input data feature carries out layer-by-layer convolution by these characteristic filter devices and input data and pondization operates,
And then the feature being hidden in data is extracted step by step.
Due in the embodiment of the present application, needing the means by transfer learning that will integrate using virtual electric power device data as base
The CNN model of plinth training is applied in the classification to practical power equipment image, therefore the embodiment of the present application selects VGG-16
Model of the model as convolutional neural networks.
(1) convolutional layer (convolution layer)
The input of each neuron is connected with upper one layer of local receptor field region in convolutional layer, by convolution kernel to this
Local receptor field region carries out feature extraction, while every layer of convolutional layer is reduced using one group of identical convolution kernel (weight is shared)
The parameter of the required calculating of neural network model, substantially increases the efficiency of e-learning.The input of convolutional layer is
Output are as follows:
Wherein, RiIndicate the set of input data signal；xi l-1It is the activation value of the ith feature vector of (l-1) layer；
Wij lIt is the convolution kernel of l layer j-th of feature vector and (l-1) layer ith feature vector；bj lBe l layers j-th of feature to
The bias of amount；uj lIt is the weighted sum of l layers of j-th of feature vector, the f () in formula (2) indicates activation primitive, general to select
The functions such as ReLU, tanh, Sigmoid are taken, activation primitive is confirmed as ReLU in VGG-16, and " * " is convolution symbol.
(2) pond layer (pooling layer)
The output of this layer is adjusted in the layer of pond using pond function.The output of pond layer:
In formula, xj lFor l layers of feature vector, bj lIt is the bias term of down-sampling layer.Pool () indicates pond function.
Pond function replaces the output of next layer network in the position using the general characteristic of the adjacent output of a certain spatial position, leads to
Frequently with method have mean value pond, maximum pond and random pool etc., pond method is confirmed as maximum pond in VGG-16
Change.
(3) full articulamentum summarizes all data characteristicses by convolution, pond process as the defeated of full articulamentum
Enter, the output of Quan Lian stratum can be by being calculated as follows:
xl=f (ul) (4)
ul=ωlxl-1+bl (5)
In formula, ulIt is that characteristic pattern x is exported by preceding layerl-1It is weighted and obtains after biasing.ωlIndicate full articulamentum
Weight coefficient, blFor the bias term of full articulamentum l.
(5) output layer
Output layer is the final result carried out after Classification and Identification to the feature vector that CNN model extraction goes out.According to practical point
Suitable classification method may be selected in generic task, and common classification method has multinomial logistic regression, Softmax classifier, supports
Vector machine etc..Classifier is Softmax in original VGG-16.Figure 17 shows a kind of VGG-16 model structures.
On the basis of the above embodiments, Figure 18 is provided by the embodiments of the present application a kind of based on transfer learning and nerve net
The structural schematic diagram of the power equipment pattern recognition device of network.It is provided in this embodiment to be based on transfer learning and mind with reference to Figure 18
Power equipment pattern recognition device through network specifically includes: first sample obtains module 1801, the second sample acquisition module
1802, the first training module 1803, the second training module 1804 and picture recognition module 1805.
Wherein, first sample obtains module 1801, for obtaining the simulated image data collection of target power equipment as the
One training sample, wherein the simulated image data collection includes the analog image and corresponding tag along sort of target power equipment,
The type of the target power equipment be it is multiple, the model of each type of target power equipment is multiple；Second sample acquisition
Module 1802, for obtaining the actual image data collection of the target power equipment as the second training sample, wherein the reality
Border image data set includes the real image and corresponding tag along sort of target power equipment；First training module 1803, is used for
Using the convolutional neural networks model that first training sample training constructs in advance, first object convolutional neural networks mould is obtained
Type；Second training module 1804 is obtained for application the second training sample training first object convolutional neural networks model
Second target convolution neural network model；Picture recognition module 1805, for the image of power equipment to be identified to be input to institute
It states the second target convolution neural network model to be identified, wherein the type of the power equipment to be identified belongs to the target
Power equipment it is one or more.
It has the advantages that in the technical solution adopted by the present invention: being matched by analog image and real image, both
Consider a large amount of training samples in analog image includes a variety of situations, then passes through transfer learning again, it is contemplated that actual
On the other hand environmental characteristic under application environment improves processing accuracy in this way, on the one hand improving processing speed, effectively solve
Certainly two days electric power equipment drawings are as identifying inaccurate problem since sample size and quality not can guarantee bring in identification process.
Further, first sample obtains module 1801 and is specifically used for:
Establish the 3D model of target power equipment；
Adjustment rendering background and/or the 3D model angle are to obtain target power equipment image；
Described image and corresponding tag along sort are exported, described image and corresponding label composition simulated image data collection are made
For the first training sample.
Further, the second sample acquisition module 1802 is specifically used for:
The target power equipment is shot in the real image of different time, different location；
The real image and corresponding tag along sort group are combined into actual image data collection as the second training sample.
Further, identification module 1805 is specifically used for:
The image of the power equipment to be identified is input to the second target convolution neural network model；
Identify the corresponding label respectively of the power equipment image to be identified；
The power equipment to be identified is determined according to the tag along sort.
Further, the convolutional neural networks model constructed in advance is VGG-16 model.
Further, the VGG-16 model includes convolutional layer, pond layer, full articulamentum and pond layer.
Further, the first training module 1803 is specifically used for:
Network parameter is determined using stochastic gradient descent method；
According to the network parameter, the convolutional neural networks model constructed in advance is trained using first training sample,
Obtain first object convolutional neural networks model.
Power equipment pattern recognition device provided by the embodiments of the present application based on transfer learning and neural network can be used
In executing the power equipment image-recognizing method provided by the above embodiment based on transfer learning and neural network, have corresponding
Function and beneficial effect.
Note that above are only the preferred embodiment and institute's application technology principle of the application.It will be appreciated by those skilled in the art that
The application is not limited to specific embodiment described here, be able to carry out for a person skilled in the art it is various it is apparent variation,
The protection scope readjusted and substituted without departing from the application.Therefore, although being carried out by above embodiments to the application
It is described in further detail, but the application is not limited only to above embodiments, in the case where not departing from the application design, also
It may include more other equivalent embodiments, and scope of the present application is determined by the scope of the appended claims.
Claims (10)
1. a kind of power equipment image-recognizing method based on transfer learning and neural network characterized by comprising
The simulated image data collection of target power equipment is obtained as the first training sample, wherein the simulated image data collection
Analog image and corresponding tag along sort including target power equipment, the type of the target power equipment be it is multiple, every kind
The model of the target power equipment of type is multiple；
The actual image data collection of the target power equipment is obtained as the second training sample, wherein the real image number
It include the real image and corresponding tag along sort of target power equipment according to collection；
Using the convolutional neural networks model that first training sample training constructs in advance, first object convolutional Neural net is obtained
Network model；
The second target convolutional neural networks are obtained using the second training sample training first object convolutional neural networks model
Model；
The image of power equipment to be identified is input to the second target convolution neural network model to identify, wherein institute
The type for stating power equipment to be identified belongs to the one or more of the target power equipment.
2. the method according to claim 1, wherein obtaining the simulated image data collection conduct of target power equipment
First training sample, comprising:
Establish the 3D model of target power equipment；
Adjustment rendering background and/or the 3D model angle are to obtain target power equipment image；
Export described image and corresponding tag along sort, described image and corresponding label composition simulated image data collection are as the
One training sample.
3. the method according to claim 1, wherein obtaining the actual image data collection of the target power equipment
As the second training sample, comprising:
The target power equipment is shot in the real image of different time, different location；
The real image and corresponding tag along sort group are combined into actual image data collection as the second training sample.
4. the method according to claim 1, wherein power equipment image to be identified is input to second mesh
Mark convolutional neural networks model is identified, comprising:
The image of the power equipment to be identified is input to the second target convolution neural network model；
Identify the corresponding label respectively of the power equipment image to be identified；
The power equipment to be identified is determined according to the tag along sort.
5. the method according to claim 1, wherein the convolutional neural networks model constructed in advance is VGG-
16 models.
6. according to the method described in claim 5, it is characterized in that, the VGG-16 model includes convolutional layer, pond layer, Quan Lian
Connect layer and pond layer.
7. the method according to claim 1, wherein the volume constructed in advance using first training sample training
Product neural network model, obtains first object convolutional neural networks model, comprising:
Network parameter is determined using stochastic gradient descent method；
It is obtained according to the network parameter using the convolutional neural networks model that first training sample training constructs in advance
First object convolutional neural networks model.
8. a kind of power equipment pattern recognition device based on transfer learning and neural network characterized by comprising
First sample obtains module, for obtaining the simulated image data collection of target power equipment as the first training sample,
In, the simulated image data collection includes the analog image and corresponding tag along sort of target power equipment, the target power
The type of equipment be it is multiple, the model of each type of target power equipment is multiple；
Second sample acquisition module, for obtaining the actual image data collection of the target power equipment as the second training sample
This, wherein the actual image data collection includes the real image and corresponding tag along sort of target power equipment；
First training module is obtained for the convolutional neural networks model that application first training sample training constructs in advance
First object convolutional neural networks model；
Second training module obtains second for application the second training sample training first object convolutional neural networks model
Target convolution neural network model；
Picture recognition module, for the image of power equipment to be identified to be input to the second target convolution neural network model
It is identified, wherein the type of the power equipment to be identified belongs to the one or more of the target power equipment.
9. device according to claim 8, which is characterized in that first sample obtains module and is specifically used for:
Establish the 3D model of target power equipment；
Adjustment rendering background and/or the 3D model angle are to obtain target power equipment image；
Export described image and corresponding tag along sort, described image and corresponding label composition simulated image data collection are as the
One training sample.
10. device according to claim 8, which is characterized in that the second sample acquisition module is specifically used for:
The target power equipment is shot in the real image of different time, different location；
The real image and corresponding tag along sort group are combined into actual image data collection as the second training sample.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910633986.9A CN110378408A (en) | 2019-07-12 | 2019-07-12 | Power equipment image-recognizing method and device based on transfer learning and neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910633986.9A CN110378408A (en) | 2019-07-12 | 2019-07-12 | Power equipment image-recognizing method and device based on transfer learning and neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110378408A true CN110378408A (en) | 2019-10-25 |
Family
ID=68253085
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910633986.9A Pending CN110378408A (en) | 2019-07-12 | 2019-07-12 | Power equipment image-recognizing method and device based on transfer learning and neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110378408A (en) |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111046932A (en) * | 2019-12-03 | 2020-04-21 | 内蒙古拜欧牧业科技有限公司 | Model training method, meat identification method, device, terminal and storage medium |
| CN111062310A (en) * | 2019-12-13 | 2020-04-24 | 哈尔滨工程大学 | Few-sample unmanned aerial vehicle image identification method based on virtual sample generation |
| CN111062448A (en) * | 2019-12-24 | 2020-04-24 | 北京知道创宇信息技术股份有限公司 | Equipment type recognition model training method, equipment type recognition method and device |
| CN111199213A (en) * | 2020-01-03 | 2020-05-26 | 云南电网有限责任公司电力科学研究院 | Equipment defect detection method and device for transformer substation |
| CN111209864A (en) * | 2020-01-07 | 2020-05-29 | 上海交通大学 | Target identification method for power equipment |
| CN111382782A (en) * | 2020-02-23 | 2020-07-07 | 华为技术有限公司 | Method and device for training classifier |
| CN111429215A (en) * | 2020-03-18 | 2020-07-17 | 北京互金新融科技有限公司 | Data processing method and device |
| CN111626357A (en) * | 2020-05-27 | 2020-09-04 | 北京微智信业科技有限公司 | Image identification method based on neural network model |
| CN112070134A (en) * | 2020-08-28 | 2020-12-11 | 广东电网有限责任公司 | Power equipment image classification method and device, power equipment and storage medium |
| CN112183303A (en) * | 2020-09-24 | 2021-01-05 | 南方电网数字电网研究院有限公司 | Transformer equipment image classification method and device, computer equipment and medium |
| CN112990121A (en) * | 2021-04-25 | 2021-06-18 | 中国人民解放军国防科技大学 | Target detection method and device, electronic equipment and storage medium |
| CN113033337A (en) * | 2021-03-08 | 2021-06-25 | 深圳市安软科技股份有限公司 | TensorRT-based pedestrian re-identification method and device |
| CN113449750A (en) * | 2020-03-26 | 2021-09-28 | 顺丰科技有限公司 | Model training method, using method, related device and storage medium |
| CN113673583A (en) * | 2021-07-30 | 2021-11-19 | 浙江大华技术股份有限公司 | Image recognition method, recognition network training method and related device |
| CN113936165A (en) * | 2021-09-07 | 2022-01-14 | 上海商涌科技有限公司 | CT image processing method, terminal and computer storage medium |
| CN114267040A (en) * | 2022-01-11 | 2022-04-01 | 云南电网有限责任公司电力科学研究院 | Image recognition sample training method based on three-dimensional model |
| CN114386524A (en) * | 2022-01-17 | 2022-04-22 | 深圳市城图科技有限公司 | Power equipment identification method for dynamic self-adaptive graph layering simulation learning |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20070104767A (en) * | 2006-04-24 | 2007-10-29 | 주식회사 로템 | The railway car operation disappointment is shift was possible simulator, system |
| CN107909095A (en) * | 2017-11-07 | 2018-04-13 | 江苏大学 | A kind of image-recognizing method based on deep learning |
| 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 |
-
2019
- 2019-07-12 CN CN201910633986.9A patent/CN110378408A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20070104767A (en) * | 2006-04-24 | 2007-10-29 | 주식회사 로템 | The railway car operation disappointment is shift was possible simulator, system |
| CN107909095A (en) * | 2017-11-07 | 2018-04-13 | 江苏大学 | A kind of image-recognizing method based on deep learning |
| 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 |
Cited By (25)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111046932A (en) * | 2019-12-03 | 2020-04-21 | 内蒙古拜欧牧业科技有限公司 | Model training method, meat identification method, device, terminal and storage medium |
| CN111062310A (en) * | 2019-12-13 | 2020-04-24 | 哈尔滨工程大学 | Few-sample unmanned aerial vehicle image identification method based on virtual sample generation |
| CN111062310B (en) * | 2019-12-13 | 2022-07-29 | 哈尔滨工程大学 | Few-sample unmanned aerial vehicle image identification method based on virtual sample generation |
| CN111062448A (en) * | 2019-12-24 | 2020-04-24 | 北京知道创宇信息技术股份有限公司 | Equipment type recognition model training method, equipment type recognition method and device |
| CN111062448B (en) * | 2019-12-24 | 2024-05-28 | 北京知道创宇信息技术股份有限公司 | Equipment type recognition model training method, equipment type recognition method and device |
| CN111199213A (en) * | 2020-01-03 | 2020-05-26 | 云南电网有限责任公司电力科学研究院 | Equipment defect detection method and device for transformer substation |
| CN111199213B (en) * | 2020-01-03 | 2023-09-26 | 云南电网有限责任公司电力科学研究院 | Method and device for detecting defects of equipment for transformer substation |
| CN111209864A (en) * | 2020-01-07 | 2020-05-29 | 上海交通大学 | Target identification method for power equipment |
| CN111209864B (en) * | 2020-01-07 | 2023-05-26 | 上海交通大学 | Power equipment target identification method |
| WO2021164254A1 (en) * | 2020-02-23 | 2021-08-26 | 华为技术有限公司 | Method and apparatus for training classifier |
| CN111382782A (en) * | 2020-02-23 | 2020-07-07 | 华为技术有限公司 | Method and device for training classifier |
| CN111382782B (en) * | 2020-02-23 | 2024-04-26 | 华为技术有限公司 | Method and device for training classifier |
| CN111429215A (en) * | 2020-03-18 | 2020-07-17 | 北京互金新融科技有限公司 | Data processing method and device |
| CN111429215B (en) * | 2020-03-18 | 2023-10-31 | 北京互金新融科技有限公司 | Data processing method and device |
| CN113449750A (en) * | 2020-03-26 | 2021-09-28 | 顺丰科技有限公司 | Model training method, using method, related device and storage medium |
| CN111626357A (en) * | 2020-05-27 | 2020-09-04 | 北京微智信业科技有限公司 | Image identification method based on neural network model |
| CN112070134A (en) * | 2020-08-28 | 2020-12-11 | 广东电网有限责任公司 | Power equipment image classification method and device, power equipment and storage medium |
| CN112183303A (en) * | 2020-09-24 | 2021-01-05 | 南方电网数字电网研究院有限公司 | Transformer equipment image classification method and device, computer equipment and medium |
| CN113033337A (en) * | 2021-03-08 | 2021-06-25 | 深圳市安软科技股份有限公司 | TensorRT-based pedestrian re-identification method and device |
| CN112990121A (en) * | 2021-04-25 | 2021-06-18 | 中国人民解放军国防科技大学 | Target detection method and device, electronic equipment and storage medium |
| CN113673583A (en) * | 2021-07-30 | 2021-11-19 | 浙江大华技术股份有限公司 | Image recognition method, recognition network training method and related device |
| CN113936165A (en) * | 2021-09-07 | 2022-01-14 | 上海商涌科技有限公司 | CT image processing method, terminal and computer storage medium |
| CN113936165B (en) * | 2021-09-07 | 2024-06-07 | 上海商涌科技有限公司 | CT image processing method, terminal and computer storage medium |
| CN114267040A (en) * | 2022-01-11 | 2022-04-01 | 云南电网有限责任公司电力科学研究院 | Image recognition sample training method based on three-dimensional model |
| CN114386524A (en) * | 2022-01-17 | 2022-04-22 | 深圳市城图科技有限公司 | Power equipment identification method for dynamic self-adaptive graph layering simulation learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110378408A (en) | Power equipment image-recognizing method and device based on transfer learning and neural network | |
| CN108416266B (en) | Method for rapidly identifying video behaviors by extracting moving object through optical flow | |
| CN108021947B (en) | A kind of layering extreme learning machine target identification method of view-based access control model | |
| CN104978580B (en) | A kind of insulator recognition methods for unmanned plane inspection transmission line of electricity | |
| CN106503687A (en) | The monitor video system for identifying figures of fusion face multi-angle feature and its method | |
| CN108090403A (en) | Face dynamic identification method and system based on 3D convolutional neural network | |
| CN107506722A (en) | One kind is based on depth sparse convolution neutral net face emotion identification method | |
| CN106504233A (en) | Image electric power widget recognition methodss and system are patrolled and examined based on the unmanned plane of Faster R CNN | |
| CN105825511A (en) | Image background definition detection method based on deep learning | |
| CN107451609A (en) | Lung neoplasm image identification system based on depth convolutional neural networks | |
| CN109583419A (en) | A kind of emotional prediction system based on depth convolutional network | |
| CN104298974A (en) | Human body behavior recognition method based on depth video sequence | |
| CN108615228A (en) | Facial image complementing method based on hybrid neural networks | |
| CN113870160B (en) | Point cloud data processing method based on transformer neural network | |
| CN114092697A (en) | Building facade semantic segmentation method with attention fused with global and local depth features | |
| CN107247952A (en) | The vision significance detection method for the cyclic convolution neutral net supervised based on deep layer | |
| CN108986091A (en) | Casting defect image detecting method based on depth Hash network | |
| Jayaraman et al. | Shapecodes: self-supervised feature learning by lifting views to viewgrids | |
| Liu et al. | Feature extraction and classification of hyperspectral image based on 3D-convolution neural network | |
| Zhang et al. | Synthetic IR image refinement using adversarial learning with bidirectional mappings | |
| CN114187506B (en) | Remote sensing image scene classification method of viewpoint-aware dynamic routing capsule network | |
| CN113361520B (en) | Transmission line equipment defect detection method based on sample offset network | |
| Ling et al. | A facial expression recognition system for smart learning based on YOLO and vision transformer | |
| Du | The computer vision simulation of athlete’s wrong actions recognition model based on artificial intelligence | |
| Xu et al. | Transformer image recognition system based on deep learning |
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: 20191025 |