CN114049305A - Distribution line pin defect detection method based on improved ALI and fast-RCNN - Google Patents

Distribution line pin defect detection method based on improved ALI and fast-RCNN Download PDF

Info

Publication number
CN114049305A
CN114049305A CN202111203745.4A CN202111203745A CN114049305A CN 114049305 A CN114049305 A CN 114049305A CN 202111203745 A CN202111203745 A CN 202111203745A CN 114049305 A CN114049305 A CN 114049305A
Authority
CN
China
Prior art keywords
network
training
rcnn
ali
fast
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202111203745.4A
Other languages
Chinese (zh)
Other versions
CN114049305B (en
Inventor
张磊
张家瑞
叶靖
薛田良
李振华
黄悦华
张涛
程江洲
熊致知
胡仕林
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
China Three Gorges University CTGU
Original Assignee
China Three Gorges University CTGU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by China Three Gorges University CTGU filed Critical China Three Gorges University CTGU
Priority to CN202111203745.4A priority Critical patent/CN114049305B/en
Publication of CN114049305A publication Critical patent/CN114049305A/en
Application granted granted Critical
Publication of CN114049305B publication Critical patent/CN114049305B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0004—Industrial image inspection
    • G06T7/0008—Industrial image inspection checking presence/absence
    • 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
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/088—Non-supervised learning, e.g. competitive learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/04—Inference or reasoning models
    • G06N5/041—Abduction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10004—Still image; Photographic image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20172—Image enhancement details
    • G06T2207/20192—Edge enhancement; Edge preservation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30108—Industrial image inspection
    • G06T2207/30164—Workpiece; Machine component

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Quality & Reliability (AREA)
  • Image Analysis (AREA)

Abstract

The distribution line pin defect detection method based on the improved ALI and the fast-RCNN comprises the following aspects: and collecting pin defect images of the distribution lines, manually marking the pin defect images on data, and constructing a training sample set. The method comprises the steps of building a network structure, wherein the first is an antagonistic learning inference network, and a basic structure consists of an inference network, a generation network and a judger. The second is the fast-RCNN network. And (4) according to the obtained training sample, carrying out detection model training, and finishing training after training for a specified number of steps. And inputting the image to be detected into a trained counterstudy inference network, outputting a high-quality reconstructed pin image, and finally completing defect identification through a trained fast-RCNN network. The method can enhance detail information such as local texture, edges and the like of the distribution line pin image to improve the image quality, and extract accurate characteristics by combining a target detection algorithm to realize intelligent detection of pin defects.

Description

Distribution line pin defect detection method based on improved ALI and fast-RCNN
Technical Field
The invention relates to the technical field of distribution line equipment image recognition, in particular to a distribution line pin defect detection method based on improved ALI and fast-RCNN.
Background
Along with the increasing development of unmanned aerial vehicle technology, unmanned aerial vehicle patrols and examines and replace artifical the patrolling and examining gradually and become main electric power defect and patrol and examine the mode. The pin has the effect of preventing the interlocking part from moving in the circuit, and the pin can fall off and the like under severe weather environment and mechanical vibration, so that the normal operation of the circuit is seriously influenced. Because the pin is in large quantity, small, the personnel of patrolling and examining need confirm the pin condition in the unmanned aerial vehicle patrols and examines the image repeatedly, and the work load that not only brings is huge, and rate of missing reporting and erroneous judgement are also higher. In order to solve the existing problems, an intelligent distribution line pin defect detection technology needs to be researched.
At present, the defect detection of the overhead distribution line is mainly divided into a detection method based on artificial feature design and a detection method based on a deep learning model. The former method focuses on the design of manual characteristics, and the manual design of characteristics is not only complicated, but also has the problem of low robustness, and can not realize end-to-end defect detection. Therefore, defect detection based on a deep learning algorithm is proposed, most of the methods adopt a convolutional neural network to extract and learn image sample characteristics, and a network structure is improved, so that hardware defects are identified.
Although the method can accurately identify the target object, the method still has the defects that after the convolution processing is carried out on the image data, the dominant information characteristics of the image have rich semantic information, the implicit information cannot be well represented by the characteristics, and further, part of pins are not easy to be identified by a detection model, so that the accuracy rate of detecting the pin defects by adopting a deep learning algorithm is low.
The inference model ALI based on antagonistic learning provides a method for obtaining depth characterization through unsupervised learning, which has previously been shown to have been able to generate images of sufficient fidelity, with significant success in image synthesis, video detection, image semantic interpretation, and image super-resolution restoration. The main advantages are as follows: (1) a Markov chain is not needed for sampling, and inference is not needed in the learning process, so that the complex probability calculation problem is avoided; (2) assuming that the training result of the judger is better, the generated network can effectively learn the distribution of the training samples.
Disclosure of Invention
The invention provides a distribution line pin defect detection method based on improved ALI and fast-RCNN, which provides an improved ALI model, wherein original distribution line pin image data to be detected is used as a training sample, ALI is used as a training algorithm to train the model, and pin detection image data is used for generating a reconstructed pin image data set through the trained improved ALI model; and finally, detecting whether the image has pin defects or not through a trained fast-RCNN model. The method can enhance detail information such as local texture, edges and the like of the distribution line pin image to improve the image quality, and extract accurate characteristics by combining a target detection algorithm to realize intelligent detection of pin defects.
The technical scheme adopted by the invention is as follows:
the distribution line pin defect detection method based on the improved ALI and the fast-RCNN comprises the following steps:
step 1: acquiring pin defect images of a distribution line, manually labeling a pin defect image data set, and constructing a training sample set;
step 2: building a network structure model:
the first one is an adversarial inference learning ALI network model, the basic structure of which comprises an inference network, a generation network and a judger;
the second one is a fast-RCNN network model, and the basic structure of the model comprises a feature extraction network CNN, a target detection network RCNN and a region suggestion network RPN;
and step 3: performing countermeasure inference learning ALI network model training according to the obtained training sample;
and 4, step 4: inputting the image to be detected into a trained confrontation inference learning ALI network model, outputting a reconstructed pin image, and finally completing defect identification through a trained fast-RCNN network model.
In the step 1, original images of pin defect images of the distribution line are acquired by unmanned aerial vehicle inspection, and the states of the pins are divided into normal, falling and rusty states; manually marking a pin defect image data set, wherein the marking format is VOC, dividing original image data into a normal nail _ good type and a defect (nail _ bad), marking the normal nail (nail _ good) when the pin is complete, and marking the defect (nail _ bad) when the other conditions are all the defects; and (4) cutting the original image data to construct a corresponding training sample set.
In the step 2, in the countermeasure inference learning ALI network model,
generating a network to generate a pseudo sample according with the distribution of the original sample;
the judger identifies the pseudo samples generated by the generation network and the samples extracted in the training set;
the inference network, the generation network, and the judger structures are all full convolution neural networks.
The counterinference learning ALI network model firstly trains the distribution line pin image data, and determines the weights of the inference network, the generation network and the internal convolution layer neurons of the judger.
In the step 3, training of the countermeasure inference learning ALI network model is performed, the inference network adopts batch type training, in order to prevent the training model from generating an overfitting phenomenon, penalty parameters are set for the loss function, and the training process is completed, specifically as follows:
the inference network adopts batch type training, a batch block with a fixed size in a sample T is randomly selected as input x in each training, and then processed image data x 'is obtained as input of the inference network through a processing process C (x' | x);
an additional penalty parameter is set in the optimization objective function to achieve the purpose of sparsity limitation, namely p is realized by minimizing the penalty parameteri' an effect close to p, so the penalty parameter can be represented by:
Figure BDA0003305978700000031
wherein: s is the number of hidden neurons in the convolutional layer in the hidden layer, and i is the neurons in the convolutional layer in the hidden layer in sequence, so the loss function after setting sparsity limit is as follows:
Figure BDA0003305978700000032
wherein: beta is a weight parameter that controls the sparsity penalty parameter.
And executing competitive learning by the countermeasure inference learning ALI network model and the fast-RCNN network model, and continuously updating parameters. And repeating the specified times to finish the model training.
And training the Faster-RCNN by using the reconstructed pin image data set to obtain a pin detection model.
In the step 4, 80% of the image data is selected as a training set, 20% of the image data is used as a verification set, and the epoch is trained for 100 times. Inputting the verification image data into the trained model and calculating the normal distribution of the verification image data, and finally comparing the reconstructed pin image data through the trained fast-RCNN network to judge the defect identification condition of the distribution line pin image data.
The invention discloses a distribution line pin defect detection method based on improved ALI and fast-RCNN, which has the following technical effects:
1) the invention applies an antagonistic learning inference mechanism, combines unsupervised model training and supervised classification and regression tasks, does not need explicit expression to generate distribution, does not have complicated variation lower limit, effectively avoids the complicated Markov chain sampling and inference process in the traditional generation model, and avoids the complicated probability calculation problem.
2) The training process of the invention not only retains the characteristics of the original image data, but also ensures that the hidden variable obtains useful characteristic representation, thereby enhancing the characteristic information of local texture, edge and the like of the pin image to be detected in the image, greatly reducing the training complexity and simultaneously improving the detection precision.
Drawings
FIG. 1 is a flow chart of defect detection according to the present invention.
FIG. 2 is a flow chart of an improved ALI model structure of the present invention.
FIG. 3 is a P-R plot of the tail _ bad (pin defect) of the present invention.
FIG. 4 is a P-R graph of nail good of the present invention.
FIG. 5(a) is a first diagram illustrating a part of the detection effects of the present invention;
fig. 5(b) is a partial detection effect diagram of the present invention.
Detailed Description
As shown in fig. 1, the method for detecting the pin defect of the power distribution line based on improved Antagonistic Learning Inference (ALI) and fast-RCNN is provided with an improved antagonistic learning inference ALI model, original image data of the power distribution line pin to be detected is used as a training sample, ALI is used as a training algorithm to train the model, and the pin detection image data is used for generating a reconstructed pin image data set through a trained improved antagonistic inference learning ALI network model; and finally, detecting whether the image has pin defects through a trained fast-RCNN model, wherein the method comprises the following steps:
step 1, data acquisition:
the original image is acquired by unmanned aerial vehicle inspection, the size of the image is 5000 x 3500, the proportion of the pin in the whole image is lower than 1%, and the damage condition of the pin of the distribution line is mainly collected in the data set. The state of the pin is divided into normal, falling, corrosion and the like; firstly, manually marking a data set, marking the data set with a VOC (volatile organic compound), dividing original image data into a normal (nail _ good) type and a defect (nail _ bad), marking the data set with the normal (nail _ good) type under the condition that the nail is normal, and marking the data set with the defect (nail _ bad) under the other conditions. The original image data is clipped to an image of 500 × 375 size, thereby constituting a training sample set.
Step 2, building a network structure:
1) improving ALI model:
in the ALI model, the training process of the inference network is to copy input to output, although the characteristics of original pin image data are extracted, the effective characteristic extraction can not be obtained by determining hidden variables. In order to improve the training efficiency of deducing and generating the network, the invention considers that the processing process C (x' | x) is introduced preferentially, the processed pin image data is obtained by utilizing the original pin image data through the process and is used as the input of the ALI model, then some constraint conditions are added to the hidden layer of the ALI model, so that the model preferentially learns the key features in the input pin image data under the constraint conditions, and finally the model extracts the features capable of better expressing the sample. The basic structure of the improved ALI is composed of an inference network, a generation network and a judger, wherein the generation network needs to generate a pseudo sample which accords with the distribution of an original sample, and the judger tries to identify the pseudo sample generated by the generator and the sample extracted from a training set. The inference network, the generation network, and the judger structures are all full convolution neural networks. The ALI model firstly trains the distribution line pin image data and determines the weights of the internal convolution layer neurons of the inference network, the generation network and the discriminator. And then, carrying out high-quality reconstruction on the distribution line pin image data to be detected by using the trained generation network to obtain reconstructed pin image data. The ALI model structure flow is shown in FIG. 2.
2) fast-RCNN model:
the fast-RCNN model is an integrated network capable of effectively realizing input image object detection, and the fast-RCNN model mainly comprises 3 parts: a feature extraction network CNN, a target detection network RCNN and a region suggestion network RPN.
Step 3, according to the obtained training sample, carrying out detection model training:
the embodiment of the invention takes the fine inspection acquisition image of the 10kV distribution line unmanned aerial vehicle of a certain-grade city power supply company as a training sample.
(a) The method comprises the following steps The inference network adopts batch type training, a batch block with a fixed size in a sample T is randomly selected as input x in each training, and then processed image data x 'is obtained through a processing process C (x' | x) and is used as input of the inference network.
(b) The method comprises the following steps The infer network and generate network processes are represented by the following equations:
z'＝ω1x'+b1
zh＝f(z')
x”＝ω2z+b2
xt＝g(x”)
wherein z ishTo infer the output of the network, xtTo generate a network output; total parameter λ ═ { ω ═ ω1,b1,ω2,b2}; f and g are sigmod activation functions, indicating activation when the neuron output approaches 1 and inhibition when the output is 0.
z' is the neuronal output, ω1To infer network weight parameters, x' is the inferred network input, b1To infer network bias.
x' isNeuronal output, ω2To generate a network weight parameter, z to generate a network input, b2To generate the network bias. The goal of inferring and generating networks is to minimize the input x and generate the network output xtAverage reconstruction error in between.
In order to prevent the over-fitting phenomenon generated by the training model, a penalty term is set for a loss function, and the loss function is as follows:
Figure BDA0003305978700000051
σ is a penalty factor.
The average activation degree of convolutional layer neurons i in the hidden layer is:
Figure BDA0003305978700000052
wherein:
Figure BDA0003305978700000053
represents the degree of activation of the convolutional layer neuron i of the hidden layer,
Figure BDA0003305978700000054
extrapolating the network for the input x and generating the activation of the convolutional layer neuron i in the network hidden layer.
m is the number of samples and j is the number of neurons.
pi′＝p
Wherein p is a sparse parameter.
The above formula shows that the average activation of the neurons i in the convolutional layer of the inference network and the generation network hidden layer is close to pi'. An additional penalty parameter is set in the optimization objective function to achieve the purpose of sparsity limitation, namely p is realized by minimizing the penalty parameteri' an effect close to p, so the penalty parameter can be represented by:
Figure BDA0003305978700000061
wherein: s is the number of hidden neurons in the convolution layer in the hidden layer, i is the neurons in the convolution layer in the hidden layer in sequence,
KL divergence KL (pp | | | p)i') to measure p and pi' similarity between them.
Therefore, the penalty function after setting the sparsity limit is:
Figure BDA0003305978700000062
wherein: β is the weight parameter that controls the sparsity penalty parameter, and L (ω, b) is the loss function.
The generation network takes random sampling from p (z) obeying normal distribution as input, and x is obtained by the generation networkh。
X, xhAnd xtAn input judger for generating a network output xhAnd xtIt is recognized from the real sample x and the production network is to fool the judger as much as possible. The two networks perform competitive learning, constantly updating parameters. The penalty function for the discriminator is:
L＝log(T(x))+log(1-T(xh))+log(1-T(xt))
T(x)、T(xh) Are output from the judger.
And repeating the specified times to finish the model training.
And 4, inputting the image to be detected into a trained counterstudy inference network, outputting a high-quality reconstructed pin image, and finally completing defect identification through the trained fast-RCNN network.
80% of image data is selected from the data as a training set, and 20% of image data is selected as a verification set. In the detection experiment, the accuracy, the recall rate, the AP value and the MAP value are selected as evaluation indexes.
1): the accuracy rate represents the proportion of correct detection in the detected faults; the recall rate indicates the proportion of all faults detected; wherein: n is a radical ofTPJudging the correct number, N, of selected faulty pinsFPTo determine the number of errors, NFNNumber of failed pins not detected:
Figure BDA0003305978700000071
Figure BDA0003305978700000072
2): and sequencing the detected target boundary frames from high to low according to the confidence level, drawing a P-R curve by taking the recall rate as an abscissa and the accuracy rate as an ordinate, wherein the area of the curve in the range of the coordinate axes is the AP value. The MAP is an average of the AP values for all classes. The results of the detection experiments were compared using fast-RCNN as a reference, and are shown in Table 1 below:
TABLE 1 comparison tables of the invention based on modified ALI and Faster-RCNN
Figure BDA0003305978700000073
As can be seen from table 1: the normal AP value of the pin is improved to 0.8110 from the original 0.7870, the defective AP value of the pin is improved to 0.6506 from the original 0.5760, and the MAP value is improved to 0.7308 from the original 0.6815, which shows that the detection accuracy of the model is superior to that of fast-RCNN.
According to the invention, through PR curves of two models, in the graphs shown in FIGS. 3 and 4, a blue curve is a PR curve of fast-RCNN, an orange curve is a PR curve of ALI + fast-RCNN, and from the local view, although the accuracy of the fast-RCNN is higher than that of the ALI + fast-RCNN in a partial region with lower recall rate, the orange curve is always higher than that of the blue curve after the recall rate is increased, and the accuracy of the ALI + fast-RCNN is obviously higher than that of the fast-RCNN; overall, the area enclosed by the orange curve is larger than that of the blue curve on both types of targets to be detected, indicating that the ALI + fast-RCNN has better detection performance on the pin detection compared with the fast-RCNN.
Fig. 5(a) and 5(b) show partial detection effects, and it can be seen from fig. 5(a) and 5(b) that the effect of the model on distribution line pin detection is obvious.

Claims (7)

1. The distribution line pin defect detection method based on the improved ALI and the fast-RCNN is characterized by comprising the following steps of: collecting pin defect images of the distribution lines, manually marking the pin defect image data, and constructing a training sample set; building a network structure: the first is an antagonistic learning inference network, the second is a fast-RCNN network, detection model training is carried out according to the obtained training samples, and after the specified steps are trained, the training is completed; inputting the image to be detected into a trained counterstudy inference network, and outputting a high-quality reconstructed pin image; and finally, completing the defect identification through the trained fast-RCNN network.
2. The distribution line pin defect detection method based on the improved ALI and the fast-RCNN is characterized by comprising the following steps of:
step 1: acquiring pin defect images of a distribution line, manually labeling a pin defect image data set, and constructing a training sample set;
step 2: building a network structure model:
the first one is an adversarial inference learning ALI network model, the basic structure of which comprises an inference network, a generation network and a judger;
the second one is a fast-RCNN network model, and the basic structure of the model comprises a feature extraction network CNN, a target detection network RCNN and a region suggestion network RPN;
and step 3: performing countermeasure inference learning ALI network model training according to the obtained training sample;
and 4, step 4: inputting the image to be detected into a trained confrontation inference learning ALI network model, outputting a reconstructed pin image, and finally completing defect identification through a trained fast-RCNN network model.
3. The improved ALI and fast-RCNN based distribution line pin defect detection method of claim 2, wherein: in the step 1, original images of pin defect images of the distribution line are acquired by unmanned aerial vehicle inspection, and the states of the pins are divided into normal, falling and rusty states; manually marking a pin defect image data set, wherein the marking format is VOC, dividing original image data into normal and defect types, marking the pin as normal under the condition of complete pin, and marking the other situations as defects; and (4) cutting the original image data to construct a corresponding training sample set.
4. The improved ALI and fast-RCNN based distribution line pin defect detection method of claim 2, wherein: in the step 2, in the countermeasure inference learning ALI network model,
generating a network to generate a pseudo sample according with the distribution of the original sample;
the judger identifies the pseudo samples generated by the generation network and the samples extracted in the training set;
the structure of the inference network, the structure of the generation network and the structure of the judger are all full convolution neural networks;
the counterinference learning ALI network model firstly trains the distribution line pin image data, and determines the weights of the inference network, the generation network and the internal convolution layer neurons of the judger.
5. The improved ALI and fast-RCNN based distribution line pin defect detection method of claim 2, wherein: in the step 3, training of the countermeasure inference learning ALI network model is performed, the inference network adopts batch type training, in order to prevent the training model from generating an overfitting phenomenon, penalty parameters are set for the loss function, and the training process is completed, specifically as follows:
(a) the method comprises the following steps The inference network adopts batch type training, a batch block with a fixed size in a sample T is randomly selected as input x in each training, and then processed image data x 'is obtained as input of the inference network through a processing process C (x' | x);
(b) the method comprises the following steps The infer network and generate network processes are represented by the following equations:
z'＝ω1x'+b1
zh＝f(z')
x”＝ω2z+b2
xt＝g(x”)
wherein z ishTo infer the output of the network, xtTo generate a network output; total parameter λ ═ { ω ═ ω1,b1,ω2,b2}; f and g are sigmod activation functions, which indicate that the neuron is activated when the output of the neuron approaches to 1, and is inhibited when the output of the neuron is 0; the goal of inferring and generating networks is to minimize the input x and generate the network output xtAverage reconstruction error therebetween;
in order to prevent the over-fitting phenomenon generated by the training model, a penalty term is set for a loss function, and the loss function is as follows:
Figure FDA0003305978690000021
the average activation degree of convolutional layer neurons i in the hidden layer is:
Figure FDA0003305978690000022
wherein:
Figure FDA0003305978690000023
represents the degree of activation of the convolutional layer neuron i of the hidden layer,
Figure FDA0003305978690000024
deducing the network and generating the activation degree of a convolutional layer neuron i in a network hidden layer when inputting x;
pi′＝p
wherein p is a sparse parameter;
the above formula shows that the average activation of neurons i in convolutional layers of the inference network and the generation network hidden layer is close topi'; an additional penalty parameter is set in the optimization objective function to achieve the purpose of sparsity limitation, namely p is realized by minimizing the penalty parameteri' Effect close to p, penalty parameter is represented by:
Figure FDA0003305978690000031
wherein: s is the number of hidden neurons in the convolutional layer in the hidden layer, and i is the neurons in the convolutional layer in the hidden layer in sequence, so the loss function after setting sparsity limit is as follows:
Figure FDA0003305978690000032
wherein: beta is a weight parameter for controlling the sparsity penalty parameter;
the generation network takes random sampling from p (z) obeying normal distribution as input, and x is obtained by the generation networkh；
X, xhAnd xtAn input judger for generating a network output xhAnd xtRecognizing from the real sample x that the generating network should fool the judger as much as possible;
executing competitive learning by the countermeasure inference learning ALI network model and the fast-RCNN network model, and continuously updating parameters; the penalty function for the discriminator is:
L＝log(T(x))+log(1-T(xh))+log(1-T(xt))；
and repeating the specified times to finish the model training.
6. The improved ALI and fast-RCNN based distribution line pin defect detection method of claim 2, wherein: in the step 4, 80% of image data is selected from the data as a training set, 20% of image data is selected as a verification set, and the epoch is trained for 100 times; inputting the verification image data into the trained model and calculating the normal distribution of the verification image data, and finally comparing the reconstructed pin image data through the trained fast-RCNN network to judge the defect identification condition of the distribution line pin image data.
7. The improved ALI and fast-RCNN based distribution line pin defect detection method of claim 6, wherein: selecting accuracy, recall rate, AP value and MAP value as evaluation indexes:
1): the accuracy rate represents the proportion of correct detection in the detected faults; the recall rate indicates the proportion of all faults detected; wherein: n is a radical ofTPJudging the correct number, N, of selected faulty pinsFPTo determine the number of errors, NFNNumber of failed pins not detected:
Figure FDA0003305978690000033
Figure FDA0003305978690000034
2): sequencing the detected target boundary frames from high to low according to the confidence level, drawing a P-R curve by taking the recall rate as an abscissa and the accuracy rate as an ordinate, wherein the area of the curve in a coordinate axis range is an AP value; the MAP is an average of the AP values for all classes.
CN202111203745.4A 2021-10-15 2021-10-15 Distribution line pin defect detection method based on improved ALI and fast-RCNN Active CN114049305B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111203745.4A CN114049305B (en) 2021-10-15 2021-10-15 Distribution line pin defect detection method based on improved ALI and fast-RCNN

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111203745.4A CN114049305B (en) 2021-10-15 2021-10-15 Distribution line pin defect detection method based on improved ALI and fast-RCNN

Publications (2)

Publication Number Publication Date
CN114049305A true CN114049305A (en) 2022-02-15
CN114049305B CN114049305B (en) 2024-07-16

Family

ID=80205202

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111203745.4A Active CN114049305B (en) 2021-10-15 2021-10-15 Distribution line pin defect detection method based on improved ALI and fast-RCNN

Country Status (1)

Country Link
CN (1) CN114049305B (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114882216A (en) * 2022-04-18 2022-08-09 华南理工大学 Garment button quality detection method, system and medium based on deep learning
CN114998309A (en) * 2022-07-12 2022-09-02 启东万惠机械制造有限公司 Unsupervised mechanical part surface defect detection method and unsupervised mechanical part surface defect detection system
CN116071744A (en) * 2023-01-10 2023-05-05 山东省气候中心 Mature-period tomato identification method and system based on Faster RCNN network

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111650204A (en) * 2020-05-11 2020-09-11 安徽继远软件有限公司 Transmission line hardware defect detection method and system based on cascade target detection
CN111914937A (en) * 2020-08-05 2020-11-10 湖北工业大学 Lightweight improved target detection method and detection system

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111650204A (en) * 2020-05-11 2020-09-11 安徽继远软件有限公司 Transmission line hardware defect detection method and system based on cascade target detection
CN111914937A (en) * 2020-08-05 2020-11-10 湖北工业大学 Lightweight improved target detection method and detection system

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
张晓航: "基于GAN的改进RPN输电线路细小金具缺陷检测方法", 《电子器件》, 20 December 2021 (2021-12-20) *
朱超平;杨永斌;: "基于改进的Faster-RCNN模型的汽车轮毂表面缺陷在线检测算法研究", 表面技术, no. 06, 20 June 2020 (2020-06-20) *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114882216A (en) * 2022-04-18 2022-08-09 华南理工大学 Garment button quality detection method, system and medium based on deep learning
CN114882216B (en) * 2022-04-18 2024-04-30 华南理工大学 Garment button attaching quality detection method, system and medium based on deep learning
CN114998309A (en) * 2022-07-12 2022-09-02 启东万惠机械制造有限公司 Unsupervised mechanical part surface defect detection method and unsupervised mechanical part surface defect detection system
CN116071744A (en) * 2023-01-10 2023-05-05 山东省气候中心 Mature-period tomato identification method and system based on Faster RCNN network
CN116071744B (en) * 2023-01-10 2023-06-30 山东省气候中心 Mature-period tomato identification method and system based on Faster RCNN network

Also Published As

Publication number Publication date
CN114049305B (en) 2024-07-16

Similar Documents

Publication Publication Date Title
CN107609525B (en) Remote sensing image target detection method for constructing convolutional neural network based on pruning strategy
WO2021134871A1 (en) Forensics method for synthesized face image based on local binary pattern and deep learning
CN114049305A (en) Distribution line pin defect detection method based on improved ALI and fast-RCNN
CN108256482B (en) Face age estimation method for distributed learning based on convolutional neural network
CN109272500B (en) Fabric classification method based on adaptive convolutional neural network
CN107194418B (en) Rice aphid detection method based on antagonistic characteristic learning
CN110929847A (en) Converter transformer fault diagnosis method based on deep convolutional neural network
CN110543906B (en) Automatic skin recognition method based on Mask R-CNN model
CN111008959A (en) Grading ring defect detection method, device, medium and equipment based on aerial image
CN112557034B (en) Bearing fault diagnosis method based on PCA _ CNNS
CN108986091A (en) Casting defect image detecting method based on depth Hash network
CN114359695A (en) Insulator breakage identification method based on uncertainty estimation
CN112164033A (en) Abnormal feature editing-based method for detecting surface defects of counternetwork texture
CN112733747A (en) Identification method, system and device for relieving falling fault of valve pull rod
CN112288700A (en) Rail defect detection method
CN115587290A (en) Aero-engine fault diagnosis method based on variational self-coding generation countermeasure network
CN111382709A (en) Insulator image detection method based on unmanned aerial vehicle inspection
CN109740672B (en) Multi-stream feature distance fusion system and fusion method
Saleh et al. Tire defect detection model using machine learning
CN114219763A (en) Infrared picture detection method for abnormal heating point of power distribution equipment based on fast RCNN algorithm
CN116152194A (en) Object defect detection method, system, equipment and medium
CN110750876A (en) Bearing data model training and using method
CN114821174B (en) Content perception-based transmission line aerial image data cleaning method
CN116304789A (en) VP inclinometer fault diagnosis method and device
Yang et al. NAM net: meta-network with normalization-based attention for few-shot 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
GR01 Patent grant