CN112164040A - Steel surface defect identification method based on semi-supervised deep learning algorithm - Google Patents
Steel surface defect identification method based on semi-supervised deep learning algorithm Download PDFInfo
- Publication number
- CN112164040A CN112164040A CN202010982319.4A CN202010982319A CN112164040A CN 112164040 A CN112164040 A CN 112164040A CN 202010982319 A CN202010982319 A CN 202010982319A CN 112164040 A CN112164040 A CN 112164040A
- Authority
- CN
- China
- Prior art keywords
- steel surface
- sample
- samples
- neural network
- 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.)
- Granted
Links
- 229910000831 Steel Inorganic materials 0.000 title claims abstract description 75
- 239000010959 steel Substances 0.000 title claims abstract description 75
- 230000007547 defect Effects 0.000 title claims abstract description 39
- 238000000034 method Methods 0.000 title claims abstract description 32
- 238000013135 deep learning Methods 0.000 title claims abstract description 20
- 238000012549 training Methods 0.000 claims abstract description 46
- 238000012795 verification Methods 0.000 claims abstract description 12
- 238000013528 artificial neural network Methods 0.000 claims description 29
- 230000006870 function Effects 0.000 claims description 19
- 239000000203 mixture Substances 0.000 claims description 6
- 230000037361 pathway Effects 0.000 claims description 5
- 238000010008 shearing Methods 0.000 claims description 5
- 238000010606 normalization Methods 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 3
- 238000003062 neural network model Methods 0.000 abstract description 20
- 230000000694 effects Effects 0.000 abstract description 3
- 238000013145 classification model Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 238000007689 inspection Methods 0.000 description 3
- 238000010801 machine learning Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000002779 inactivation Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
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
-
- 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
- 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
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- 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
-
- 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
- G06T2207/20076—Probabilistic image processing
-
- 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
- G06T2207/20081—Training; Learning
-
- 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
- G06T2207/20084—Artificial neural networks [ANN]
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Quality & Reliability (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to the technical field of image classification and identification, in particular to a steel surface defect identification method based on a semi-supervised deep learning algorithm, which comprises the following steps: A) acquiring an image of the surface of the steel and obtaining initial image sample data; B) dividing initial image sample data into a training data set and a verification data set according to a preset proportion, selecting 20% of the training data set to perform manual marking to serve as a labeled sample set, and taking the rest as a non-labeled sample set; C) inputting the training data set in the step B) into a model classifier for iterative training, and verifying by using samples in a verification sample set in each iteration N steps until a preset threshold value is reached; D) and inputting the new steel surface image into the model classifier to obtain a defect identification result of the steel surface image. The substantial effects of the invention are as follows: the neural network model can adapt to sample data without labels, the sample data with labels required by the classifier is greatly reduced, and the training cost of the classifier is reduced.
Description
Technical Field
The invention relates to the technical field of image classification and identification, in particular to a steel surface defect identification method based on a semi-supervised deep learning algorithm.
Background
The automatic identification of the steel surface defects based on deep learning is one of research hotspots in steel quality inspection, but most of the current deep learning quality inspection methods focus on a supervised learning algorithm which needs large-scale marking samples. In some practical cases, it is difficult to collect and mark enough data samples for the supervised learning algorithm model, so that the application and development of deep learning in the field of steel surface quality inspection are influenced. For example, chinese patent CN109242825A, published 2019, 1 month, 18 days, a method and an apparatus for identifying defects on a steel surface based on a deep learning technique, the method includes: collecting corresponding steel surface defect pictures according to the type of the steel production line; marking the defects according to a marking process of a marking platform to generate a training sample set and a verification sample set; designing a neural network model according to the types and characteristics of the steel surface defects; training a neural network model according to the training sample set and the training process of the neural network; evaluating each index of the neural network model according to the calibration sample set and the evaluation method of the neural network model, and optimizing the neural network model; and according to the equipment performance and the size of the neural network model, carrying out the deployment and application of the neural network model in a distributed mode. The technical scheme adopts the fully labeled samples to train the classification model, and the efficiency is low. The semi-supervised deep learning utilizes marked samples and unmarked samples to carry out model training, and can well overcome the problem that a large number of marked samples are difficult to obtain. Therefore, it is necessary to develop a classification model construction technique using semi-supervised learning.
Disclosure of Invention
The technical problem to be solved by the invention is as follows: the defect identification is inaccurate in calculation due to insufficient labeled sample data in the existing steel surface defect machine identification. The method can reduce the number of labeled samples required by classification model training, fully utilizes sample data without labels, and reduces the difficulty of classification model establishment.
In order to solve the technical problems, the technical scheme adopted by the invention is as follows: the steel surface defect identification method based on the semi-supervised deep learning algorithm comprises the following steps: A) acquiring an image of the surface of the steel and obtaining initial image sample data; B) dividing initial image sample data into a training data set and a verification data set according to a preset proportion, turning and shearing the training data set for enhancement, selecting 20% of the training data set for manual marking to serve as a labeled sample set, and taking the rest as a non-labeled sample set to add labels to the verification data set; C) inputting the training data set in the step B) into a model classifier for iterative training, verifying by using samples in a verification sample set in every iteration N steps, and recording the accuracy of the model classifier until the accuracy reaches a preset threshold; D) inputting the new steel surface image into the model classifier trained in the step C) to obtain a defect identification result of the steel surface image. The number of samples is increased by turning and shearing the training data set, and manual marking is performed by selecting 20%, so that the cost of sample marking is reduced.
Preferably, in the step a), the steel surface image is preprocessed: A1) carrying out normalization processing on the steel surface image; A2) and scaling the normalized steel surface image to enable the pixel width of the shorter side of the normalized steel surface image to be a preset value m, and then cutting the longer side of the normalized steel surface image to enable the steel surface image to be converted into a format of m multiplied by 3. The normalized steel surface image is easier to be processed by a classifier. Wherein m is preferably 32x32, the preferred reason is the pixel size of the steel surface defect, which can be generally shown by 32x32 pixel size, and the meaningless data required to be processed by the classifier is reduced.
Preferably, in step C), the model classifier is a neural network, and the method for iteratively training the neural network includes: C1) setting a threshold ηt1、ηt2And ηt3(ii) a C2) Taking a sample x from a labeled sample set or an unlabeled sample set1Input to the neural network to obtain an output result f of the neural networkθ(x1),fθ(x1) Including the probability of k defects, if sample x1Belongs to a set of labeled samples fθ(x1)|k≤ηt1Then sample x is sampled1Is the neural network f, is the loss function L, thetaθ(x) If the sample x is the training parameter of1Belongs to a set of unlabeled samples, fθ(x2)|k<ηt2And f isθ(x2)|k<ηt3Then sample x is sampled2The training result of (a) is incorporated into the loss function L; C3) calculating the gradient g from the loss function Lθ,Updating a sliding average parameter theta', theta ═ theta +(1-alpha) theta, alpha is a sliding average rate, and then calculating the neural network fθ(x) Is equal to Step (theta', g)θ) The number of iterations t is added to 1, and the process is re-executed from step C1). Whether the loss function L is included is improved, so that the neural network model can adapt to sample data without labels.
Preferably, in step C2), sample x is obtained1The pathway of (a) further comprises: taking two samples x simultaneously from a set of labeled or unlabeled samplesjAnd xl(ii) a Two samples xjAnd xlMixing is carried out xm=Mixλ(xj,xl)=λxj+(1-λ)xlAnd the label is mixed by interpolation, will be provided withAs a sample x1And inputting the neural network. The number of samples can be increased through sample mixing, the existing labeled sample data is fully used, and the sample data required by training of the neural network model is reduced.
Preferably, step C2)Obtaining a sample x1The pathway of (a) further comprises: taking two samples u simultaneously from a set of unlabeled samplesjAnd ulTwo samples x1And x2Are mixed, um=Mixλ(uj,ul)=λuj+(1-λ)ul(ii) a Computing speculative tagsAnd ujand ulRespectively representing two samples of the source unlabeled sample set, carrying out interpolation mixing on the labels,will be provided withAs a sample x1And inputting the neural network. The sample mixing can increase the number of samples, fully utilize the existing sample data and reduce the sample data required by the training of the neural network model.
Preferably, in step C2), a preset number of samples are taken from the labeled sample set and input to the neural network, and then samples are taken from the labeled sample set or the unlabeled sample set for iteration. For steel surface unlabeled data we use training signal enhancement. For the unmarked data on the steel surface, the neural network model is low in identification precision at the beginning and is easy to misjudge the unmarked data on the steel surface, so that the unmarked sample on the steel surface can be fully utilized after the neural network model has certain precision and is judged, and the selected unmarked sample on the steel surface has purposiveness in adding training
Preferably, in step C2), when the accuracy of the neural network model reaches a set threshold, a sample is taken from the labeled sample or unlabeled sample set for iteration.
Preferably, in step C1), a threshold value is set Wherein T is the current iteration number, T is the total iteration number, and k is the number of the steel surface defect types.
Preferably, in step C2), the loss function L is a cross-entropy loss function or a squared error loss function. The cross entropy loss function or the square error loss function can reflect the loss caused by the parameters of the neural network model during iteration.
The substantial effects of the invention are as follows: the neural network model can adapt to sample data without a label, so that the classifier can be trained by combining a labeled sample set and a non-labeled sample set to obtain the identification result of the steel surface defect, the labeled sample data required by the classifier is greatly reduced, the training cost of the classifier is reduced, and the application range of machine learning in the identification of the steel surface defect is expanded.
Drawings
FIG. 1 is a block diagram illustrating a method for identifying defects on a steel surface according to an embodiment.
Detailed Description
The following provides a more detailed description of the present invention, with reference to the accompanying drawings.
The first embodiment is as follows:
a steel surface defect identification method based on a semi-supervised deep learning algorithm is shown in figure 1 and comprises the following steps:
A) and acquiring an image of the surface of the steel and obtaining initial image sample data. Preprocessing the steel surface image: A1) carrying out normalization processing on the steel surface image; A2) and scaling the normalized steel surface image to enable the pixel width of the shorter side of the normalized steel surface image to be a preset value m, and then cutting the longer side of the normalized steel surface image to enable the steel surface image to be converted into a format of m multiplied by 3. The normalized steel surface image is easier to be processed by a classifier. Wherein m takes the value of 32x32, the pixel size of the steel surface defect can be usually shown by the pixel size of 32x32, and the meaningless data required to be processed by the classifier is reduced.
B) Dividing initial image sample data into a training data set and a verification data set according to a preset proportion, turning and shearing the training data set for enhancement, selecting 20% of the training data set for manual marking to serve as a labeled sample set, and taking the rest as a non-labeled sample set to add labels to the verification data set.
C) Inputting the training data set in the step B) into a neural network for iterative training, verifying by using samples in a verification sample set in every iteration N steps, and recording the accuracy of the model classifier until the accuracy reaches a preset threshold.
The method for iteratively training the neural network comprises the following steps: C1) setting a threshold ηt1、ηt2And ηt3Setting a threshold value Wherein T is the current iteration number, T is the total iteration number, and k is the number of the types of the steel surface defects; C2) taking a sample x from a labeled sample set or an unlabeled sample set1Input to the neural network to obtain an output result f of the neural networkθ(x1),fθ(x1) Including the probability of k defects, if sample x1Belongs to a set of labeled samples fθ(x1)|k≤ηt1Then sample x is sampled1Is the neural network f, is the loss function L, thetaθ(x) If the sample x is the training parameter of1Belongs to a set of unlabeled samples, fθ(x2)|k<ηt2And f isθ(x2)|k<ηt3Then sample x is sampled2The training result of (a) is incorporated into the loss function L; C3) calculating the gradient g from the loss function Lθ,Updating a sliding average parameter theta', theta ═ theta +(1-alpha) theta, alpha is a sliding average rate, and then calculating the neural network fθ(x) Is equal to Step (theta', g)θ) The number of iterations t is added to 1, and the process is re-executed from step C1). Whether the loss function L is included is improved, so that the neural network model can adapt to sample data without labels.
D) Inputting the new steel surface image into the model classifier trained in the step C) to obtain a defect identification result of the steel surface image. The number of samples is increased by turning and shearing the training data set, and manual marking is performed by selecting 20%, so that the cost of sample marking is reduced.
In step C2), after a preset number of samples are taken from the labeled sample set and input to the neural network, samples are taken from the labeled sample set or the unlabeled sample set for iteration. For steel surface unlabeled data we use training signal enhancement. For the unmarked data on the steel surface, the neural network model is low in identification precision at the beginning and is easy to misjudge the unmarked data on the steel surface, so that the unmarked sample on the steel surface can be fully utilized after the neural network model has certain precision, and the selected unmarked sample on the steel surface has purposiveness in training.
The beneficial technical effects of this embodiment are: the neural network model can adapt to sample data without a label, so that the classifier can be trained by combining a labeled sample set and a non-labeled sample set to obtain the identification result of the steel surface defect, the labeled sample data required by the classifier is greatly reduced, the training cost of the classifier is reduced, and the application range of machine learning in the identification of the steel surface defect is expanded.
Example two:
in the embodiment of the method for identifying the defects of the steel surface based on the semi-supervised deep learning algorithm, in step C2), a sample x is obtained1The pathway of (a) further comprises: taking two samples x simultaneously from a set of labeled or unlabeled samplesjAnd xl(ii) a Two samples xjAnd xlMixing is carried out xm=Mixλ(xj,xl)=λxj+(1-λ)xlAnd the label is mixed by interpolation,will be provided withAs a sample x1And inputting the neural network. Taking two samples u simultaneously from a set of unlabeled samplesjAnd ulTwo samples x1And x2Are mixed, um=Mixλ(uj,ul)=λuj+(1-λ)ul(ii) a Computing speculative tagsAnd ujand ulRespectively representing two samples of the source unlabeled sample set, carrying out interpolation mixing on the labels,will be provided withAs a sample x1Input neural netLinking the collaterals.
The neural network established in this embodiment includes: the first layer of network structure is a convolutional layer with a reception field size of 3 × 3, a convolutional kernel of 16, a step size of 1, and parameters of the convolutional layer of 3 × 3 × 3 × 16+16, namely 448, and the output is a feature map with 16 channels and a size of 32 × 32. The second and third tier network structures are residual structures with two 3x3 convolutional layers. The second layer of the network structure is a convolutional layer with a 3 × 3 convolutional kernel of 32 reception field size and a step size of 2, and outputs a feature map with 32 channels and a size of 16 × 16. The third layer of the network structure is a convolutional layer with a 3 × 3 convolutional kernel of 32 reception field size and a step size of 1, and outputs a feature map with 32 channels and a size of 16 × 16. The fourth and fifth layer network structures are residual structures with two 3 × 3 convolutional layers. The fourth layer of the network structure is a convolutional layer with a 3 × 3 convolutional kernel of 32 reception field size and a step size of 1, and outputs a feature map with 32 channels and a size of 16 × 16. The fifth layer network structure is a convolutional layer with a reception field size of 3 × 3 and a convolutional layer with a convolutional kernel of 32, the step size of 1, and a feature map with 32 channels and a size of 16 × 16 is output. The sixth and seventh layer network structures use a residual structure with two 3 × 3 convolutional layers. The sixth layer of the network structure is a convolutional layer with a 3 × 3 convolutional kernel of 64 reception field size and a step size of 2, and outputs a feature map with 64 channels and a size of 8 × 8. The seventh layer network structure is a convolutional layer with a reception field size of 3 × 3, a convolutional kernel of 64, a step size of 1, and a feature map with 64 channels and a size of 8 × 8 as output. The eighth and ninth tier network architectures employ a residual architecture with two 3x3 convolutional layers. The eighth layer of the network structure is a convolutional layer with a 3 × 3 convolutional kernel of 64 reception field size and a step size of 1, and outputs a feature map with 64 channels and a size of 8 × 8. The ninth layer of the network structure is a convolutional layer with a 3 × 3 convolutional kernel of 64 reception field size and a step size of 1, and outputs a feature map with 64 channels and a size of 8 × 8.
The tenth and eleventh network structures use a residual structure with two 3 × 3 convolutional layers. The tenth layer of the network structure is a convolutional layer with a reception field size of 3 × 3 and a convolutional layer with a convolutional kernel of 128, and the characteristic diagram with a step size of 2 is output and has a size of 4 × 4 and 128 channels. The eleventh layer of the network structure is a convolutional layer with a reception field size of 3 × 3 and a convolutional kernel of 128, and the convolutional layer with a step size of 1 outputs a feature map with 128 channels and a size of 4 × 4.
The twelfth and thirteenth layer architectures are residual architectures with two 3x3 convolutional layers. The twelfth layer of the network structure is a convolutional layer with a reception field size of 3 × 3 and a convolutional layer with a convolutional kernel of 128, and the characteristic diagram with a step size of 2 and an output of 128 channels and a size of 4 × 4. The thirteenth layer of the network structure is a convolutional layer with a 3 × 3 convolutional kernel of 128 reception field size and a step size of 1, and outputs a feature map with 128 channels and a size of 4 × 4.
Finally, the dropout and softmax classification layers, with a 3 × 3 receptive field size of maximum pooling with step size of 2 and random inactivation of 0.1, are used to calculate the probability that the output belongs to each class.
Compared with the embodiment, the method and the device can more fully utilize the existing sample data, and further reduce the number of the samples required by the neural network training. The rest steps are the same as the first embodiment.
The above embodiment is only a preferred embodiment of the present invention, and is not intended to limit the present invention in any way, and other variations and modifications may be made without departing from the technical scope of the claims.
Claims (8)
1. A steel surface defect identification method based on a semi-supervised deep learning algorithm is characterized in that,
the method comprises the following steps:
A) acquiring an image of the surface of the steel and obtaining initial image sample data;
B) dividing initial image sample data into a training data set and a verification data set according to a preset proportion, turning and shearing the training data set for enhancement, selecting 20% of the training data set for manual marking to serve as a labeled sample set, and taking the rest as a non-labeled sample set to add labels to the verification data set;
C) inputting the training data set in the step B) into a model classifier for iterative training, verifying by using samples in a verification sample set in every iteration N steps, and recording the accuracy of the model classifier until the accuracy reaches a preset threshold;
D) inputting the new steel surface image into the model classifier trained in the step C) to obtain a defect identification result of the steel surface image.
2. The steel surface defect identification method based on semi-supervised deep learning algorithm as recited in claim 1,
in the step A), the steel surface image is preprocessed:
A1) carrying out normalization processing on the steel surface image;
A2) and scaling the normalized steel surface image to enable the pixel width of the shorter side of the normalized steel surface image to be a preset value m, and then cutting the longer side of the normalized steel surface image to enable the steel surface image to be converted into a format of m multiplied by 3.
3. The steel surface defect identification method based on semi-supervised deep learning algorithm as recited in claim 1,
in step C), the model classifier is a neural network, and the method for performing iterative training on the neural network includes: C1) setting a threshold ηt1、ηt2And ηt3;
C2) Taking a sample x from a labeled sample set or an unlabeled sample set1Input to the neural network to obtain an output result f of the neural networkθ(x1),fθ(x1) Including the probability of k defects, if sample x1Belongs to a set of labeled samples fθ(x1)|k≤ηt1Then sample x is sampled1Is the neural network f, is the loss function L, thetaθ(x) The training parameters of (a) are set,
if sample x1Belongs to a set of unlabeled samples, fθ(x2)|k<ηt2And f isθ(x2)|k<ηt3Then sample x is sampled2The training result of (a) is incorporated into the loss function L;
C3) calculating the gradient g from the loss function Lθ,Updating a sliding average parameter theta', theta ═ theta +(1-alpha) theta, alpha is a sliding average rate, and then calculating the neural network fθ(x) Is equal to Step (theta', g)θ) The number of iterations t is added to 1, and the process is re-executed from step C1).
4. The steel surface defect identification method based on semi-supervised deep learning algorithm as recited in claim 3,
in step C2), sample x is obtained1The pathway of (a) further comprises:
taking two samples x simultaneously from a set of labeled or unlabeled samplesjAnd xl;
Two samples xjAnd xlMixing is carried out xm=Mixλ(xj,xl)=λxj+(1-λ)xl,
5. The steel surface defect identification method based on semi-supervised deep learning algorithm as recited in claim 3,
in step C2), sample x is obtained1The pathway of (a) further comprises:
taking two samples u simultaneously from a set of unlabeled samplesjAnd ulTwo samples x1And x2Are mixed, um=Mixλ(uj,ul)=λuj+(1-λ)ul;
Computing speculative tagsAndujand ulRespectively representing two samples of the source unlabeled sample set, carrying out interpolation mixing on the labels,
6. The steel surface defect identification method based on semi-supervised deep learning algorithm as recited in claim 3,
in step C2), a preset number of samples are taken from the labeled sample set and input into the neural network, and then samples are taken from the labeled sample set or the unlabeled sample set for iteration.
8. The steel surface defect identification method based on semi-supervised deep learning algorithm as recited in claim 3,
in step C2), the loss function L is a cross entropy loss function or a squared error loss function.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010982319.4A CN112164040B (en) | 2020-09-17 | Steel surface defect identification method based on semi-supervised deep learning algorithm |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010982319.4A CN112164040B (en) | 2020-09-17 | Steel surface defect identification method based on semi-supervised deep learning algorithm |
Publications (2)
Publication Number | Publication Date |
---|---|
CN112164040A true CN112164040A (en) | 2021-01-01 |
CN112164040B CN112164040B (en) | 2024-05-10 |
Family
ID=
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113947601A (en) * | 2021-12-20 | 2022-01-18 | 武汉鑫国茂包装有限公司 | Plastic product surface defect detection method and system based on semi-supervised learning |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20180068429A1 (en) * | 2015-04-15 | 2018-03-08 | Institute Of Automation Chinese Academy Of Sciences | Image Steganalysis Based on Deep Learning |
CN108764281A (en) * | 2018-04-18 | 2018-11-06 | 华南理工大学 | A kind of image classification method learning across task depth network based on semi-supervised step certainly |
CN111489326A (en) * | 2020-01-13 | 2020-08-04 | 杭州电子科技大学 | Copper foil substrate surface defect detection method based on semi-supervised deep learning |
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20180068429A1 (en) * | 2015-04-15 | 2018-03-08 | Institute Of Automation Chinese Academy Of Sciences | Image Steganalysis Based on Deep Learning |
CN108764281A (en) * | 2018-04-18 | 2018-11-06 | 华南理工大学 | A kind of image classification method learning across task depth network based on semi-supervised step certainly |
CN111489326A (en) * | 2020-01-13 | 2020-08-04 | 杭州电子科技大学 | Copper foil substrate surface defect detection method based on semi-supervised deep learning |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113947601A (en) * | 2021-12-20 | 2022-01-18 | 武汉鑫国茂包装有限公司 | Plastic product surface defect detection method and system based on semi-supervised learning |
CN113947601B (en) * | 2021-12-20 | 2022-03-22 | 武汉鑫国茂包装有限公司 | Plastic product surface defect detection method and system based on semi-supervised learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111046784B (en) | Document layout analysis and identification method and device, electronic equipment and storage medium | |
CN111126386B (en) | Sequence domain adaptation method based on countermeasure learning in scene text recognition | |
CN104966097A (en) | Complex character recognition method based on deep learning | |
WO2016165082A1 (en) | Image stego-detection method based on deep learning | |
CN111861978A (en) | Bridge crack example segmentation method based on Faster R-CNN | |
CN105654129A (en) | Optical character sequence recognition method | |
CN111428071A (en) | Zero-sample cross-modal retrieval method based on multi-modal feature synthesis | |
CN112232371B (en) | American license plate recognition method based on YOLOv3 and text recognition | |
CN110648310A (en) | Weak supervision casting defect identification method based on attention mechanism | |
CN113779260B (en) | Pre-training model-based domain map entity and relationship joint extraction method and system | |
CN107818321A (en) | A kind of watermark date recognition method for vehicle annual test | |
CN114880468A (en) | Building specification examination method and system based on BilSTM and knowledge graph | |
CN113420619A (en) | Remote sensing image building extraction method | |
CN115131747A (en) | Knowledge distillation-based power transmission channel engineering vehicle target detection method and system | |
CN111984790B (en) | Entity relation extraction method | |
CN112364837A (en) | Bill information identification method based on target detection and text identification | |
CN115019183B (en) | Remote sensing image model migration method based on knowledge distillation and image reconstruction | |
CN112579777A (en) | Semi-supervised classification method for unlabelled texts | |
CN115357718B (en) | Method, system, device and storage medium for discovering repeated materials of theme integration service | |
CN109255794B (en) | Standard part depth full convolution characteristic edge detection method | |
CN112164040A (en) | Steel surface defect identification method based on semi-supervised deep learning algorithm | |
CN112164040B (en) | Steel surface defect identification method based on semi-supervised deep learning algorithm | |
CN116089944A (en) | Cross-platform application program abnormality detection method and system based on transfer learning | |
CN110427973B (en) | Classification method for ambiguity-oriented annotation samples | |
CN114863163A (en) | Method and system for cell classification based on cell image |
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 |