CN114240892B - Knowledge distillation-based unsupervised industrial image anomaly detection method and system - Google Patents

Knowledge distillation-based unsupervised industrial image anomaly detection method and system Download PDF

Info

Publication number
CN114240892B
CN114240892B CN202111555291.7A CN202111555291A CN114240892B CN 114240892 B CN114240892 B CN 114240892B CN 202111555291 A CN202111555291 A CN 202111555291A CN 114240892 B CN114240892 B CN 114240892B
Authority
CN
China
Prior art keywords
image
network
anomaly
knowledge distillation
industrial
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.)
Active
Application number
CN202111555291.7A
Other languages
Chinese (zh)
Other versions
CN114240892A (en
Inventor
沈卫明
曹云康
宋亚楠
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.)
Huazhong University of Science and Technology
Original Assignee
Huazhong University of Science and Technology
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 Huazhong University of Science and Technology filed Critical Huazhong University of Science and Technology
Priority to CN202111555291.7A priority Critical patent/CN114240892B/en
Publication of CN114240892A publication Critical patent/CN114240892A/en
Application granted granted Critical
Publication of CN114240892B publication Critical patent/CN114240892B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

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
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00—Geometric image transformations in the plane of the image
    • G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
    • G06T3/4023—Scaling of whole images or parts thereof, e.g. expanding or contracting based on decimating pixels or lines of pixels; based on inserting pixels or lines of pixels
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/70—Denoising; Smoothing

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Pure & Applied Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Computational Mathematics (AREA)
  • Computing Systems (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Biomedical Technology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Algebra (AREA)
  • Databases & Information Systems (AREA)
  • Evolutionary Biology (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Quality & Reliability (AREA)
  • Image Analysis (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses an unsupervised industrial image anomaly detection method and system based on knowledge distillation, and belongs to the technical field of industrial image processing. The method comprises two stages of training and testing, wherein the multi-scale knowledge distillation and multi-scale anomaly fusion are formed, the multi-scale knowledge distillation comprises a teacher network and a student network, self-adaptive difficult cases are used for mining difficult cases samples, and the student network is optimized by using pixel and context similarity among the difficult cases. In the training stage, only normal industrial images are utilized to carry out knowledge distillation from a teacher network to a student network, and iterative optimization is carried out on student network parameters, so that the depth characteristics of the student network and the normal industrial product images extracted by the teacher network are similar; in the test stage, depth features of the test image are extracted respectively, and regression errors among the features can be used for image anomaly segmentation and detection. The method effectively improves the performance of unsupervised industrial image anomaly detection, reduces labor cost, and improves the automation and intellectualization level of quality inspection of the production line.

Description

Knowledge distillation-based unsupervised industrial image anomaly detection method and system
Technical Field
The invention belongs to the technical field of industrial image processing, and further relates to an unsupervised industrial image anomaly detection method and system based on knowledge distillation.
Background
The automatic detection of the quality of industrial products is of great importance, and the industrial image anomaly detection technology can improve the product quality detection performance and production efficiency of the production line. Currently most industries, for example: the steel, textile, semiconductor industries and the like, the detection of the product quality still mainly depends on manual experience, and the detection personnel are easy to fatigue after working for a long time, so that false detection and omission detection are caused. The industrial image anomaly detection technology based on image processing is developed, and can replace manual work to effectively detect the quality of industrial products. The traditional industrial image anomaly detection main flow based on image processing comprises the following steps: the method comprises the steps of collecting images, extracting features and scoring anomalies, wherein the feature extraction part is highly dependent on expert knowledge, and is difficult to generalize to different scenes. The industrial image anomaly detection method based on deep learning does not need manual priori, can accelerate the deployment of product quality inspection, improves the detection performance and the detection speed, and is becoming a mainstream method gradually.
However, because the frequency of occurrence of anomalies in the industrial production line is low, the collected normal data is far greater than the abnormal data, and therefore the unsupervised deep learning anomaly detection method is significant.
Disclosure of Invention
Aiming at the defects or improvement demands of the prior art, the invention provides an unsupervised industrial image anomaly detection method based on knowledge distillation, which aims to detect industrial product image anomalies by utilizing the strong feature expression capability of deep learning, does not need to mark anomaly data in the parameter optimization stage, effectively improves the detection efficiency, reduces the manufacturing cost and has adaptability to different types of industrial product image anomalies.
In order to achieve the above purpose, the invention provides an unsupervised industrial image anomaly detection method based on knowledge distillation, which comprises the following steps:
Training phase:
Training a normal industrial product image serving as a training sample image by using the sample image to generate an abnormality detection model, wherein the abnormality detection model comprises a teacher network and a student network, the teacher network is used for extracting first features of an input image, and the student network is used for extracting second features of the input image;
extracting a difficult sample from the first feature and the second feature through self-adaptive difficult mining;
utilizing the difficult sample to improve an objective function of a student network in a knowledge distillation mode, and training to generate an anomaly detection model;
Testing:
inputting the image to be detected into a trained anomaly detection model to obtain anomaly scores of different scales, and fusing the anomaly scores of all scales by using multiplication to obtain a final anomaly score map to finish image anomaly detection.
The abnormality detection model can be divided into a multi-scale knowledge distillation module and a multi-scale abnormality fusion module, wherein the multi-scale knowledge distillation module comprises a teacher network and a student network.
The deep neural network trained on the large-scale natural image dataset has high discrimination of the extracted deep features, and can be used as an effective industrial image feature extractor. Normal industrial image features extracted by pre-training deep neural networks tend to be distributed in a compact manifold of high-dimensional feature space. The abnormal industrial image features are distributed more dispersedly and are far from the normal feature manifold. Modeling the distribution of the normal feature manifold using a model G (θ):
Where M represents the normal feature manifold. Theoretically, when G (θ) is able to accurately describe a normal feature manifold, G (x; θ) can accurately determine whether test data belongs to the manifold. However, the general model is difficult to accurately give the classification result, but gives probability, and needs to manually give a threshold value for further truncation:
G(x；θ)＝y (2)
In general, the larger y indicates the more likely that the data is anomalous. Since training data is limited and distributed in actual training, it is difficult to directly learn G (θ) to describe a normal feature manifold. Knowledge distillation can be used to implicitly learn a normal feature manifold, and the corresponding model G can be expressed as G (T, S), where T represents a teacher network, is a pre-training network with high discriminant, and S represents a student network with the same network structure as the teacher network. The corresponding abnormality discrimination model is:
wherein, F T and F S respectively represent the feature extractor of the teacher network and the student network, and formula (3) is essentially to calculate the feature difference extracted by the teacher network and the student network. In order for the formula (3) to satisfy the condition of the formula (2), namely: when the test data is normal data, the abnormal score is small; when the test data is abnormal data, the abnormal score is large, the characteristics of normal data extracted by the T and the S are required to be similar as far as possible, and the characteristics of the abnormal data are required to be different as far as possible. Through knowledge distillation on normal data, the characteristics of the normal data extracted by the T and the S can be made to be similar as much as possible; meanwhile, S does not learn the characteristic of the abnormal data extracted by T, and the characteristic difference of the abnormal data extracted by the S is larger. The knowledge distillation step is described in detail below.
Given image I, the corresponding image feature f T,fS is extracted by S, T:
fT＝FT(I) (4)
fS＝FS(I) (5)
Let f T,fS be h×w in resolution and D in dimension. Normalization of f T,fS:
considering different scale features extracted from deep neural network A feature map representing the first layer.
In order to make the normal data characteristics obtained by extracting T and S be similar as possible, the method can be directly optimized Defining pixel level loss as:
Wherein, Indicating the total number of scales, and H (l)×W(l) indicates the size of each scale. By optimizing the pixel level loss of the normal sample features, the normal data features extracted by T and S can be made similar. However, in the practical training process, the parameter quantity of the selected T and S is usually larger, and the similarity between normal industrial images is extremely high, so that enough knowledge is difficult to extract, and the fitting problem occurs in the training process. Namely, the normal data characteristics of the training set extracted by the T and the S are very similar, but larger regression errors still exist in the actual use process, so that the abnormal detection performance is reduced. In order to alleviate the overfitting problem, the method uses the context similarity loss and the self-adaptive difficult-case mining to further optimize the training process.
In the knowledge distillation training process, the loss function used determines how the student network learns the knowledge of the teacher network. To alleviate the overfitting problem, the student network may be constrained from extracting contextual similarity of features with the teacher network. Intuitively, if the context similarity between the student network and the teacher network is high, they should have the same similarity relationship between depth features extracted from the same input image. Since the depth features are normalized to the unit hypersphere, their similarity relationship can be defined by cosine similarity:
Wherein Q (f) ∈R N×D denotes a pair A stretching operation is performed, where D represents the feature dimension and N represents the number of features. G (f) represents a similarity calculation, and values of G (f T)(l) and G (f S)(l), with coordinates (i, j)) encode the similarity of the i-th vector in Q (f T)(l)) to the j-th vector in Q (f S)(l):
where N (l)＝H(l)×W(l) represents the number of vectors contained in Q T (l),QS (l). Taking into account the context similarity, the final loss function can be defined as:
L(fT,fS)＝Lps(fT,fS)+λLcs(fT,fS) (13)
where λ is the hyper-parameter weight that controls pixel similarity and context similarity loss.
In the training process, the difficult sample should play a larger role, while the simple sample does not need continuous optimization. The existing method adopts a fixed threshold value to judge whether any sample is a difficult sample, but the fixed threshold value is difficult to select, and the required threshold values among different tasks are different. The invention uses the self-adaptive difficult-case mining strategy to dynamically mine the difficult cases in real time. More specifically, a difficult mining strategy for measuring the difficulty level of a sample in real time is provided. The difficult mining strategy may be defined as follows:
m(l,t)＝αm(l,t-1)+(1-α)(μ(l,t)+βσ(l,t)) (14)
Wherein, And (3) extracting depth characteristic differences with the characteristic map position (i, j) by the teacher network and the student network for the t iteration. Mu (l,t),σ(l,t) isThe mean and variance of (3) so that the difficulty level of the sample can be dynamically judged based on mu (l,t),σ(l,t). The formula (14) uses an exponential moving average method to avoid extreme noise generated in each iteration process, so that the change of the dynamic threshold m (l,t) is smoother. Alpha is a super-parameter for controlling the smoothness degree of m (l,t), and beta is a super-parameter for controlling the difficulty degree of digging to obtain a sample. In general, the larger β, the fewer the number of samples mined, and the greater the difficulty in the training batch. Difficult cases mining can be extracted by the following formula:
Wherein, (i h,jh) represents the finally extracted difficult sample coordinates, and the corresponding teacher network characteristics and student network characteristics are represented as After the difficult cases are mined, the loss function is calculated by only using the mined difficult case samples, so that the difficult case samples in the training process play a larger role, and meaningless optimization of simple samples is avoided. The loss function after difficult mining can be defined as:
After training, the student network may extract features similar to those of the teacher network that extract normal data, i.e., the student network has implicitly described the normal data manifold extracted by the teacher network. At this time, the anomaly score may be calculated according to equation (3). Specifically, the anomaly score is calculated as follows:
To obtain better performance, multiplications are used to fuse the outlier scores for each scale:
wherein Upsample () is an up-sampling function, and up-samples the abnormal score maps of different scales to the resolution of the input image through quadratic linear interpolation. AS (f) is the final anomaly score map, which can provide a reference for subsequent decisions.
In another aspect, the present invention provides an unsupervised industrial image anomaly detection system based on knowledge distillation, comprising: a computer readable storage medium and a processor;
The computer-readable storage medium is for storing executable instructions;
The processor is used for reading executable instructions stored in the computer readable storage medium and executing the method for detecting the unsupervised industrial image abnormality based on knowledge distillation.
The unsupervised industrial image anomaly detection method based on knowledge distillation establishes an end-to-end deep learning anomaly detection model, can be used for solving the problem of cold start of a production line due to insufficient anomaly samples at early stage, improves the performance of unsupervised industrial image anomaly detection, effectively improves the quality detection automation level of the production line based on image analysis, and can provide services for subsequent supervised anomaly detection and anomaly diagnosis.
Drawings
FIG. 1 is a schematic diagram of an unsupervised industrial image anomaly detection method based on knowledge distillation.
Detailed Description
The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention. In addition, the technical features of the embodiments of the present invention described below may be combined with each other as long as they do not collide with each other.
The invention provides an unsupervised industrial image anomaly detection method based on knowledge distillation, which is shown in fig. 1 and comprises the following steps:
Training phase:
Training a normal industrial product image serving as a training sample image by using the sample image to generate an abnormality detection model, wherein the abnormality detection model comprises a teacher network and a student network, the teacher network is used for extracting first features of an input image, and the student network is used for extracting second features of the input image;
extracting a difficult sample from the first feature and the second feature through self-adaptive difficult mining;
utilizing the difficult sample to improve an objective function of a student network in a knowledge distillation mode, and training to generate an anomaly detection model;
Testing:
inputting the image to be detected into a trained anomaly detection model to obtain anomaly scores of different scales, and fusing the anomaly scores of all scales by using multiplication to obtain a final anomaly score map to finish image anomaly detection.
The abnormality detection model can be divided into a multi-scale knowledge distillation module and a multi-scale abnormality fusion module, wherein the multi-scale knowledge distillation module comprises a teacher network and a student network. The deep neural network trained on the large-scale natural image dataset has high discrimination of the extracted deep features, and can be used as an effective industrial image feature extractor. Normal industrial image features extracted by pre-training deep neural networks tend to be distributed in a compact manifold of high-dimensional feature space. The abnormal industrial image features are distributed more dispersedly and are far from the normal feature manifold. Modeling the distribution of the normal feature manifold using a model G (θ):
Where M represents the normal feature manifold. Theoretically, when G (θ) is able to accurately describe a normal feature manifold, G (x; θ) can accurately determine whether test data belongs to the manifold. However, the general model is difficult to accurately give the classification result, but gives probability, and needs to manually give a threshold value for further truncation:
G(x；θ)＝y (2)
In general, the larger y indicates the more likely that the data is anomalous. Since training data is limited and distributed in actual training, it is difficult to directly learn G (θ) to describe a normal feature manifold. Knowledge distillation can be used to implicitly learn a normal feature manifold, and the corresponding model G can be expressed as G (T, S), where T represents a teacher network, is a pre-training network with high discriminant, and S represents a student network with the same network structure as the teacher network. The corresponding abnormality discrimination model is:
wherein, F T and F S respectively represent the feature extractor of the teacher network and the student network, and formula (3) is essentially to calculate the feature difference extracted by the teacher network and the student network. In order for the formula (3) to satisfy the condition of the formula (2), namely: when the test data is normal data, the abnormal score is small; when the test data is abnormal data, the abnormal score is large, the characteristics of normal data extracted by the T and the S are required to be similar as far as possible, and the characteristics of the abnormal data are required to be different as far as possible. Through knowledge distillation on normal data, the characteristics of the normal data extracted by the T and the S can be made to be similar as much as possible; meanwhile, S does not learn the characteristic of the abnormal data extracted by T, and the characteristic difference of the abnormal data extracted by the S is larger. The knowledge distillation step is described in detail below.
Given image I, the corresponding image feature f T,fS is extracted by S, T:
fT＝FT(I) (4)
fS＝FS(I) (5)
Let f T,fS be h×w in resolution and D in dimension. Normalization of f T,fS:
considering different scale features extracted from deep neural network A feature map representing the first layer.
In order to make the normal data characteristics obtained by extracting T and S be similar as possible, the method can be directly optimized Defining pixel level loss as:
Wherein, Indicating the total number of scales, and H (l)×W(l) indicates the size of each scale. By optimizing the pixel level loss of the normal sample features, the normal data features extracted by T and S can be made similar. However, in the practical training process, the parameter quantity of the selected T and S is usually larger, and the similarity between normal industrial images is extremely high, so that enough knowledge is difficult to extract, and the fitting problem occurs in the training process. Namely, the normal data characteristics of the training set extracted by the T and the S are very similar, but larger regression errors still exist in the actual use process, so that the abnormal detection performance is reduced. In order to alleviate the overfitting problem, the method uses the context similarity loss and the self-adaptive difficult-case mining to further optimize the training process.
In the knowledge distillation training process, the loss function used determines how the student network learns the knowledge of the teacher network. To alleviate the overfitting problem, the student network may be constrained from extracting contextual similarity of features with the teacher network. Intuitively, if the context similarity between the student network and the teacher network is high, they should have the same similarity relationship between depth features extracted from the same input image. Since the depth features are normalized to the unit hypersphere, their similarity relationship can be defined by cosine similarity:
Wherein Q (f) ∈R N×D denotes a pair A stretching operation is performed, where D represents the feature dimension and N represents the number of features. G (f) represents a similarity calculation, and values of G (f T)(l) and G (f S)(l), with coordinates (i, j)) encode the similarity of the i-th vector in Q (f T)(l)) to the j-th vector in Q (f S)(l):
where N (l)＝H(l)×W(l) represents the number of vectors contained in Q T (l),QS (l). Taking into account the context similarity, the final loss function can be defined as:
L(fT,fS)＝Lps(fT,fS)+λLcs(fT,fS) (13)
where λ is the hyper-parameter weight that controls pixel similarity and context similarity loss.
In the training process, the difficult sample should play a larger role, while the simple sample does not need continuous optimization. The existing method adopts a fixed threshold value to judge whether any sample is a difficult sample, but the fixed threshold value is difficult to select, and the required threshold values among different tasks are different. The invention uses the self-adaptive difficult-case mining strategy to dynamically mine the difficult cases in real time, and only uses the difficult-case samples obtained by mining to calculate the loss, as shown in AHSM (Adaptively Hard Sample Mining) in fig. 1. More specifically, a difficult mining strategy for measuring the difficulty level of a sample in real time is provided. The difficult mining strategy may be defined as follows:
m(l,t)＝αm(l,t-1)+(1-α)(μ(l,t)+βσ(l,t)) (14)
Wherein, And (3) extracting depth characteristic differences with the characteristic map position (i, j) by the teacher network and the student network for the t iteration. Mu (l,t),σ(l,t) isThe mean and variance of (3) so that the difficulty level of the sample can be dynamically judged based on mu (l,t),σ(l,t). The formula (14) uses an exponential moving average method to avoid extreme noise generated in each iteration process, so that the change of the dynamic threshold m (l,t) is smoother. Alpha is a super-parameter for controlling the smoothness degree of m (l,t), and beta is a super-parameter for controlling the difficulty degree of digging to obtain a sample. In general, the larger β, the fewer the number of samples mined, and the greater the difficulty in the training batch. Difficult cases mining can be extracted by the following formula:
Wherein, (i h,jh) represents the finally extracted difficult sample coordinates, and the corresponding teacher network characteristics and student network characteristics are represented as After the difficult cases are mined, the loss function is calculated by only using the mined difficult case samples, so that the difficult case samples in the training process play a larger role, and meaningless optimization of simple samples is avoided. The loss function after difficult mining can be defined as:
After training, the student network may extract features similar to those of the teacher network that extract normal data, i.e., the student network has implicitly described the normal data manifold extracted by the teacher network. At this time, the anomaly score may be calculated according to equation (3). Specifically, the anomaly score is calculated as follows:
To obtain better performance, multiplications are used to fuse the outlier scores for each scale:
wherein Upsample () is an up-sampling function, and up-samples the abnormal score maps of different scales to the resolution of the input image through quadratic linear interpolation. AS (f) is the final anomaly score map, which can provide a reference for subsequent decisions.
It will be readily appreciated by those skilled in the art that the foregoing description is merely a preferred embodiment of the invention and is not intended to limit the invention, but any modifications, equivalents, improvements or alternatives falling within the spirit and principles of the present patent are intended to be included within the scope of the present invention.

Claims (4)

1. An unsupervised industrial image anomaly detection method based on knowledge distillation, which is characterized by comprising the following steps:
Training phase:
Training a normal industrial product image serving as a training sample image by using the sample image to generate an abnormality detection model, wherein the abnormality detection model comprises a teacher network and a student network, the teacher network is used for extracting first features of an input image, and the student network is used for extracting second features of the input image;
extracting a difficult sample from the first feature and the second feature through self-adaptive difficult mining;
Utilizing the difficult sample to improve an objective function of a student network in a knowledge distillation mode, and training to generate an anomaly detection model; the objective function is:
Where L ps is the pixel similarity penalty, L cs is the context similarity penalty, The first characteristic and the second characteristic corresponding to the difficult sample are represented, and lambda is the hyper-parameter weight for controlling the pixel similarity loss and the context similarity loss;
The pixel similarity loss is:
Wherein, Indicating the total number of scales, H (l)×W(l) indicating the size of each scale;
The context similarity loss is:
Wherein, Representing the total number of scales, H (l)×W(l) representing the size of each scale, Q (f) εR N×D representing the pairPerforming a stretching operation, wherein D represents feature dimensions, N represents feature numbers, G (f) represents similarity calculation, and values of G (f T)(l) and G (f S)(l) with coordinates (i, j) encode the similarity of the ith vector in Q (f T)(l) and the jth vector in Q (f S)(l));
Testing:
Inputting the image to be detected into a trained anomaly detection model to obtain anomaly scores of different scales, and fusing the anomaly scores of all scales by using multiplication to obtain a final anomaly score map to finish image anomaly detection; the anomaly score is calculated as follows:
The anomaly scores for the various scales are fused using multiplication:
Wherein AS (f) is the final outlier plot, upsample () is an upsampling function, upsampling outlier plots of different scales to the resolution of the input image by quadratic linear interpolation, Indicating the total number of degrees of scale.
2. The method of claim 1, wherein the first and second features f T,fS are represented as:
fT＝FT(I)
fS＝FS(I)
Normalization:
Wherein, F T and F S respectively represent feature extractors of a teacher network and a student network, and I is an input image.
3. The method of claim 2, wherein the refractory samples are represented as:
m(l,t)＝αm(l,t-1)+(1-α)(μ(l,t)+βσ(l,t))
Wherein, A feature map representing a teacher network and a student network of the first layer,For the t iteration, the teacher network and the student network extract depth feature differences with feature map positions (i, j), m (l,t) as a dynamic threshold value and mu (l,t)、σ(l,t) asAlpha is a super parameter for controlling the smoothness degree of m (l,t), and beta is a super parameter for controlling the difficulty degree of the sample obtained by excavation.
4. An unsupervised industrial image anomaly detection system based on knowledge distillation, comprising: a computer readable storage medium and a processor;
The computer-readable storage medium is for storing executable instructions;
The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the knowledge distillation-based unsupervised industrial image anomaly detection method of any one of claims 1 to 3.
CN202111555291.7A 2021-12-17 2021-12-17 Knowledge distillation-based unsupervised industrial image anomaly detection method and system Active CN114240892B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111555291.7A CN114240892B (en) 2021-12-17 2021-12-17 Knowledge distillation-based unsupervised industrial image anomaly detection method and system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111555291.7A CN114240892B (en) 2021-12-17 2021-12-17 Knowledge distillation-based unsupervised industrial image anomaly detection method and system

Publications (2)

Publication Number Publication Date
CN114240892A CN114240892A (en) 2022-03-25
CN114240892B true CN114240892B (en) 2024-07-02

Family

ID=80758577

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111555291.7A Active CN114240892B (en) 2021-12-17 2021-12-17 Knowledge distillation-based unsupervised industrial image anomaly detection method and system

Country Status (1)

Country Link
CN (1) CN114240892B (en)

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114742799B (en) * 2022-04-18 2024-04-26 华中科技大学 Industrial scene unknown type defect segmentation method based on self-supervision heterogeneous network
CN114782694B (en) * 2022-06-21 2022-09-30 中国科学技术大学 Unsupervised anomaly detection method, system, device and storage medium
CN115239638A (en) * 2022-06-28 2022-10-25 厦门微图软件科技有限公司 Industrial defect detection method, device and equipment and readable storage medium
EP4322106A4 (en) 2022-06-30 2024-02-28 Contemporary Amperex Technology Co., Limited Defect detection method and apparatus
CN116028891B (en) * 2023-02-16 2023-07-14 之江实验室 Industrial anomaly detection model training method and device based on multi-model fusion
CN116630286B (en) * 2023-05-31 2024-02-13 博衍科技(珠海)有限公司 Method, device, equipment and storage medium for detecting and positioning image abnormality
CN116993694B (en) * 2023-08-02 2024-05-14 江苏济远医疗科技有限公司 Non-supervision hysteroscope image anomaly detection method based on depth feature filling

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021197223A1 (en) * 2020-11-13 2021-10-07 平安科技（深圳）有限公司 Model compression method, system, terminal, and storage medium
WO2021248868A1 (en) * 2020-09-02 2021-12-16 之江实验室 Knowledge distillation-based compression method for pre-trained language model, and platform

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021248868A1 (en) * 2020-09-02 2021-12-16 之江实验室 Knowledge distillation-based compression method for pre-trained language model, and platform
WO2021197223A1 (en) * 2020-11-13 2021-10-07 平安科技（深圳）有限公司 Model compression method, system, terminal, and storage medium

Also Published As

Publication number Publication date
CN114240892A (en) 2022-03-25

Similar Documents

Publication Publication Date Title
CN114240892B (en) Knowledge distillation-based unsupervised industrial image anomaly detection method and system
Zhao et al. Deep multi-scale convolutional transfer learning network: A novel method for intelligent fault diagnosis of rolling bearings under variable working conditions and domains
Zhao et al. A visual long-short-term memory based integrated CNN model for fabric defect image classification
CN113256066A (en) PCA-XGboost-IRF-based job shop real-time scheduling method
CN112819802B (en) Method for supervising and predicting blast furnace condition abnormality based on tuyere information deep learning
CN112964469B (en) Online fault diagnosis method for rolling bearing under variable load of transfer learning
CN110728694B (en) Long-time visual target tracking method based on continuous learning
CN109685765B (en) X-ray film pneumonia result prediction device based on convolutional neural network
CN111753918B (en) Gender bias-removed image recognition model based on countermeasure learning and application
Liang et al. Comparison detector for cervical cell/clumps detection in the limited data scenario
CN108765374B (en) Method for screening abnormal nuclear area in cervical smear image
Zhang et al. Imbalanced data based fault diagnosis of the chiller via integrating a new resampling technique with an improved ensemble extreme learning machine
Freytag et al. Labeling examples that matter: Relevance-based active learning with gaussian processes
CN112784872B (en) Cross-working condition fault diagnosis method based on open set joint transfer learning
CN112017204A (en) Tool state image classification method based on edge marker graph neural network
CN107194413A (en) A kind of differentiation type based on multi-feature fusion cascades the target matching method of display model
CN116051479A (en) Textile defect identification method integrating cross-domain migration and anomaly detection
CN117576079A (en) Industrial product surface abnormality detection method, device and system
CN113222149A (en) Model training method, device, equipment and storage medium
Nguyen et al. Physics-infused fuzzy generative adversarial network for robust failure prognosis
Li et al. Gadet: A geometry-aware x-ray prohibited items detector
CN113128564A (en) Typical target detection method and system based on deep learning under complex background
CN116383747A (en) Anomaly detection method for generating countermeasure network based on multi-time scale depth convolution
CN113469977B (en) Flaw detection device, method and storage medium based on distillation learning mechanism
Liang et al. Figure-ground image segmentation using genetic programming and feature selection

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