CN115082435B - Defect detection method based on self-supervision momentum contrast - Google Patents
Defect detection method based on self-supervision momentum contrast Download PDFInfo
- Publication number
- CN115082435B CN115082435B CN202210858635.XA CN202210858635A CN115082435B CN 115082435 B CN115082435 B CN 115082435B CN 202210858635 A CN202210858635 A CN 202210858635A CN 115082435 B CN115082435 B CN 115082435B
- Authority
- CN
- China
- Prior art keywords
- defect
- sample
- image
- positive
- momentum
- 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
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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/28—Quantising the image, e.g. histogram thresholding for discrimination between background and foreground patterns
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
- G06V10/765—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects using rules for classification or partitioning the feature space
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/772—Determining representative reference patterns, e.g. averaging or distorting patterns; Generating dictionaries
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- 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]
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- General Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention relates to the technical field of single-classification image detection, and discloses a defect detection method based on self-supervision momentum comparison, which comprises the following steps: generating an original pseudo-defect image; cutting a small rectangular area with variable size and aspect ratio from the original pseudo-defect image, randomly rotating the small rectangular area and dithering pixel values, and then pasting the small rectangular area to the random position of the positive sample to form a pseudo-defect image; positive and negative sample characteristics which can obviously reflect the existence of defects are generated by utilizing a comparison learning loss function driving model, so that a normal sample and a defect sample are separated, a characteristic dictionary is maintained by utilizing a momentum comparison method, and the characteristics of the positive and negative samples are learned while the characteristics keep consistency; and calculating the defect abnormal score by using a Gaussian density estimator, and judging whether the whole image is a defect image. The method can simulate the irregularity, the continuity and the diversity of the defect data as accurately as possible, and improve the detection accuracy of the model according to the method.
Description
Technical Field
The invention relates to the technical field of single-classification image detection, in particular to a defect detection method based on self-supervision momentum comparison.
Background
Anomaly detection is intended to detect instances of anomalies and defect patterns that differ from what is seen for normal instances. Many of the problems from different vision applications are anomaly detection, including manufacturing defect detection, medical image analysis, and video surveillance. Unlike typical supervised classification problems, anomaly detection faces unique challenges.
First, due to the nature of the problem, it is difficult to obtain large amounts of anomaly data, whether tagged or untagged. Second, the defect region may be a small and subtle high resolution image, so the difference between normal and abnormal patterns tends to be fine grained. Second, due to limited access to the anomaly data, the construction of an anomaly detector is typically done under semi-supervised or single classification settings using only normal data. Finally, since the distribution of abnormal patterns is unknown in advance, the training models only learn the existing normal examples, and therefore the training models cannot well determine whether the test sample is abnormal.
In summary, the invention provides a defect detection method based on self-supervision momentum comparison, which can construct defect data, simulate the irregularity, the continuity and the diversity of the defect data as accurately as possible, and improve the detection accuracy of a model according to the method so as to solve the problem of low accuracy of abnormal detection.
Disclosure of Invention
The invention aims to provide a defect detection method based on self-supervision momentum comparison.
The invention is realized by the following technical scheme: a defect detection method based on self-supervision momentum contrast comprises the following steps:
s1, constructing a Gaussian noise sample image as an original pseudo-defect image according to an original positive sample, a randomly generated mean value and a variance in a data set;
s2, generating a negative sample by using a sample mixing operation;
s3, generating positive and negative sample characteristics which can obviously reflect the existence of defects by using a comparison learning loss function driving model, separating a normal sample from a defect sample, maintaining a characteristic dictionary by using a momentum comparison method, and learning the characterization of the positive and negative samples while keeping the consistency of the characteristics;
and S4, calculating a defect abnormal score by using a Gaussian density estimator, judging whether the whole image is a defect image, dividing the image into image blocks according to a fixed step length when the image is tested to position defects, extracting features from all the image blocks, calculating an abnormal score for each image block, and transmitting the abnormal score to each pixel by using a Gaussian smoothing method.
The invention discloses a defect detection method based on self-supervision momentum comparison, which combines an original positive sample in a data set, utilizes a randomly generated mean value and variance to jointly construct a Gaussian noise sample image, utilizes a Berlin noise generator to generate a noise image, carries out binarization through a uniformly randomly sampled threshold value to form a defect distribution map, and utilizes the positive sample and the Gaussian noise sample to combine to generate an original pseudo-defect image according to the defect distribution map; cutting a small rectangular area with variable size and aspect ratio from the original pseudo-defect image, randomly rotating and dithering the pixel value of the small rectangular area, and then pasting the small rectangular area to the random position of the positive type sample to form a pseudo-defect image, namely a negative type sample; positive and negative sample characteristics which can obviously reflect the existence of defects are generated by utilizing a comparison learning loss function driving model, so that a normal sample and a defect sample are separated as far as possible, a larger characteristic dictionary is maintained by utilizing a momentum comparison method, and the characteristics of the positive and negative samples are learned while the characteristics keep consistency; calculating the abnormal score of the defect by using a Gaussian density estimator, judging whether the whole image is the defect image, dividing the image into image blocks according to a fixed step length when the image positioning defect is tested, extracting features from all the image blocks, calculating the abnormal score of each image block, and transmitting the abnormal score to each pixel by using a Gaussian smoothing method. The invention provides a novel pseudo defect construction method and a training frame based on positive and negative samples, which can effectively detect defects on the aspect of single classification. The pseudo-defect sample construction method accurately simulates the irregularity, the coherence and the diversity of the defect sample, and ensures the further improvement of the learning performance of the neural network model.
In order to better implement the present invention, further, the step S1 includes:
s11, combining the original positive samples in the data set, and constructing a Gaussian noise sample image by using a randomly generated mean value and variance;
step S12, generating a noise image by using a Berlin noise generator;
s13, performing binarization processing on the noise image through a uniformly randomly sampled threshold value to form a defect distribution map;
and S14, generating an original pseudo-defect image by combining the normal sample and the Gaussian noise sample according to the defect distribution map.
In order to better implement the present invention, further, the step S2 includes:
randomly selecting a cutting area from an original pseudo-defect image, cutting a small rectangular area with variable size and aspect ratio, randomly rotating the small rectangular area and dithering pixel values, and then pasting the small rectangular area to a random position of a positive sample to form a pseudo-defect image as a negative sample.
In order to better implement the present invention, further, the step S1 includes:
s11, combining the original positive samples in the data set, and constructing a Gaussian noise sample image by using a randomly generated mean value and variance;
step S12, generating a noise image by using a Berlin noise generator;
s13, performing binarization processing on the noise image through a uniformly and randomly sampled threshold value to form a defect distribution map;
and S14, generating an original pseudo-defect image by combining the normal sample and the Gaussian noise sample according to the defect distribution map.
In order to better implement the present invention, further, the step S2 includes:
a negative class sample is generated using a sample mixing operation.
In order to better implement the present invention, further, the step S3 includes:
constructing a corresponding negative sample for each positive sample through sample mixing operation, and generating a negative sample set;
combining the positive and negative sample sets to create a sample space;
generating a query view and a dictionary view by a random data expansion method, and sending the query view and the dictionary view into a corresponding query network and a corresponding dictionary network to obtain a pair of embedded vectors q and k;
and adding a classifier driving model after the network is queried for training, performing momentum updating by using the dictionary network and the query network, and performing comparison learning according to the embedded vector acquired by the dictionary network and the embedded vector acquired by the query network.
In order to better implement the present invention, further, the step S4 includes:
when the image positioning defect is tested, dividing the image into image blocks according to a fixed step length, extracting features from all the image blocks, calculating the abnormal score of each image block, and transmitting the calculated abnormal score to each pixel by using a Gaussian smoothing method;
randomly selecting an embedded vector of a sample in a training set to participate in the calculation of a total probability formula;
calculating the mean value and the variance of each embedded vector, and obtaining a total probability formula according to the obtained mean value and variance;
and for the test sample, calculating a defect abnormity score according to a total probability formula, and judging whether the whole image is a defect image according to the defect abnormity score to finish defect detection.
Compared with the prior art, the invention has the following advantages and beneficial effects:
(1) The invention provides a construction method of defect data, which can simulate the irregularity, the continuity and the diversity of the defect data as accurately as possible and improve the detection accuracy of a model according to the method;
(2) The invention utilizes the comparison learning framework to drive the model to generate positive and negative sample characteristics which can obviously reflect the existence of defects, so that the positive and negative samples are separated as far as possible;
(3) The invention maintains a larger feature dictionary by using a momentum comparison method, and learns the characterization of positive and negative samples while keeping the consistency of the features.
Drawings
The invention is further described with reference to the following figures and examples, all of which are intended to be covered by the present disclosure and the scope of the invention.
Fig. 1 is a schematic diagram of defect data generation in a defect detection method based on auto-supervision momentum contrast according to the present invention.
Fig. 2 is a schematic flow chart of a defect detection method based on auto-supervision momentum contrast according to the present invention.
Detailed Description
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it should be understood that the described embodiments are only a part of the embodiments of the present invention, and not all embodiments, and therefore should not be considered as a limitation to the scope of protection. All other embodiments, which can be obtained by a person skilled in the art without any inventive step based on the embodiments of the present invention, are within the scope of the present invention.
In the embodiment of the invention, the mixnoise operation is a sample mixing operation, the Perlin noise generator is a Berlin noise generator, the query view is a query view, the key view is a dictionary view, the key network is a dictionary network, and the query network is a query network.
Example 1:
the defect detection method based on the auto-supervised momentum comparison of the present embodiment is a schematic diagram of defect data generation as shown in fig. 1, and two examples are illustrated to show how defect data is generated. In order to simulate the irregularity, continuity and diversity of the defect data as accurately as possible, the Berlin noise generator is used to generate noise with the same size as the sampleAn acoustic image, a defect distribution diagram is formed by binarization of the noise image through a threshold value of uniform random sampling
And randomly generating a mean value and a variance for each positive sample in the data set, generating Gaussian noise according to the values of the mean value and the variance and the size of the positive sample, and superposing the Gaussian noise on the positive sample to construct a Gaussian noise sample image A. The Gaussian noise sample image is distributed according to the defect distribution map
Performing texture display and matching with the positive type sample
Mixing to form defect anomalies beyond the distribution to obtain an original pseudo-defect image
. From the original pseudo-defect image
And cutting a small rectangular area with variable size and aspect ratio, randomly rotating the small rectangular area and dithering pixel values, and then pasting the small rectangular area to the random position of the positive sample to form a pseudo-defect image, namely a negative sample. As shown in fig. 2, in the present invention, first, only normal samples without defects in the single classification defect detection data set, i.e., the original positive type samples in the data set, are constructed, and then the defect samples (negative type samples) are constructed, so that the positive type samples in the data set are used at first. Combining an original positive sample in a data set, utilizing a numerical value generated by a random function as a mean value and a variance, utilizing a random function random to randomly generate a numerical value as a mean value and a variance, constructing a Gaussian noise sample image with the same size as an original image according to the randomly generated mean value and variance, generating a noise image by utilizing a Berlin noise generator, performing binarization by using a uniformly randomly sampled threshold value to form a defect distribution diagram, and performing binarization according to defectsAnd the distribution graph generates an original pseudo-defect image by combining the positive type sample and the Gaussian noise sample. Secondly, cutting a small rectangular area with variable size and aspect ratio from the original pseudo-defect image, randomly rotating and dithering the pixel value of the small rectangular area, randomly rotating the rectangular area for simulating the irregularity of the defect sample by the cut small rectangular area, and then pasting the small rectangular area to the random position of the positive sample to form the pseudo-defect image, namely the negative sample. Secondly, positive and negative sample characteristics which can obviously reflect the existence of defects are generated by using a comparison learning loss function driving model, so that a normal sample and a defect sample are separated as much as possible, a larger characteristic dictionary is maintained by using a momentum comparison method, and the characteristics of the positive and negative samples are learned while the characteristics keep consistency. Finally, a Gaussian density estimator is used for calculating defect abnormal scores, the abnormal scores have two uses, and the first method is that the embedded vector of the whole image calculates the abnormal scores to judge whether the image has defects; and the other method is that the abnormal score is utilized to assist the image in positioning, whether the whole image is a defect image is judged, when the image positioning defect is tested, the image is divided into image blocks according to a fixed step length, the features are extracted from all the image blocks, the abnormal score is calculated for each image block, and the abnormal score is spread to each pixel by using a Gaussian smoothing method.
Example 2:
this embodiment is further optimized based on embodiment 1, in which a Perlin noise generator (berlin noise generator) is used to generate a noise image and capture various abnormal shapes, binarization is performed by a uniform random sampling threshold value to form a defect distribution map, and an original pseudo-defect image is generated by combining a normal sample and a gaussian noise sample according to the defect distribution map.
The invention is provided with
In order to be a sample space, the sample space,
whereinn is the number of samples in the dataset.
Generating a Perlin noise image (Berlin noise image) P by using a Perlin noise generator (Berlin noise generator), and binarizing the noise image P by uniformly and randomly sampling a threshold value to form a defect distribution map
。
Randomly generating a mean value and a variance value for each positive type sample in the data set, generating Gaussian noise according to the values of the mean value and the variance value and the size of the positive type sample, and superposing the Gaussian noise on the positive type sample subjected to any three data enhancement operations (the data enhancement operations comprise sharpness, brightness change, color change and tone separation) to construct a constructed Gaussian noise sample image A.
The Gaussian noise sample image is distributed according to the defect distribution map
Performing texture display and matching with the positive type sample
Mixing to form defect anomalies beyond the distribution to obtain an original pseudo-defect image
. Original pseudo-defect image
Is defined as
(1);
Wherein the content of the first and second substances,
is composed of
The inverse of (a) is,
is a dot-product operation, and the operation,
is a transparency parameter upon mixing, the parameter being selected from
Interval uniform sampling, i.e.
And A is the constructed Gaussian noise sample image.
Other parts of this embodiment are the same as embodiment 1, and thus are not described again.
Example 3:
this embodiment is further optimized on the basis of the above embodiment 1 or 2, from the original pseudo-defect image
And cutting a small rectangular area with variable size and aspect ratio, randomly rotating the small rectangular area and dithering a pixel value, and then pasting the small rectangular area to the random position of the positive type sample to form a pseudo-defect image, namely the negative type sample.
The invention defines a mixnoise operation (sample mixing operation) to generate a negative sample, which is essentially a data enhancement strategy, and particularly generates a local irregular noise pattern and places the local irregular noise pattern into a positive sample to try to simulate defect data in real life.
The mixnoise operation (sample mixing operation) proposed by the invention is to obtain the original pseudo-defect image
Cutting a small rectangular area with randomly variable size and aspect ratio, randomly rotating the small rectangular area and dithering pixel values, and then pasting the converted small rectangular area to a positive sample
Thereby forming a final pseudo-defect image, i.e., a negative type sample, represented as:
wherein, in the step (A),
in order to be a negative class sample set,
is the nth negative class sample, and n is the number of samples.
Other parts of this embodiment are the same as those of embodiment 1 or 2, and thus are not described again.
Example 4:
in this embodiment, a comparison learning loss function is used to drive the model to generate positive and negative type sample features that can obviously reflect the existence of defects, so that the normal sample and the negative type sample are separated as much as possible, a larger feature dictionary is maintained by using a momentum comparison method, and the features of the positive and negative type samples are learned while the features are kept consistent.
First, a negative type sample set is constructed for each positive type sample through a mixnoise operation (sample mixing operation)
Wherein, in the process,
in order to be a negative class sample set,
for the nth negative sample class, n is the number of samples, and the positive and negative sample class sets are combined to create a sample space
Expressed as:
where n is the number of samples, due to sample space
Is created by combining positive and negative samples, so the total number of the samples is 2n, and the random data expansion method is adopted
Two views, namely, the query view and the key view, are generated and then fed into a query network
Heke network (dictionary network)
Obtaining a pair of embedded vectors
And
here, both the query network (query network) and the key network (dictionary network) are Resnet50 networks without classifiers, i.e., a common convolutional layer, a BN layer, and a ReLU activation function followed by four residual modules. For query network (query network)
Has two functions of obtaining embedded vector and training input data, and for the following functions, the invention is applied to query network
A classifier is added later to drive model training, expressed as:
(2) Wherein, in the step (A),
is a function of the loss of the classification,
the normal sample is put into a query network (query network) after data enhancement operation, the prediction and the original label are obtained through a classifier to carry out cross entropy operation,
the positive type sample is put into a query network (query network) after being subjected to data enhancement operation by using a negative type sample obtained by mixnoise operation, the prediction and the original label are subjected to cross entropy operation by using a classifier,
for cross entropy loss, for contrast learning, through the query network (query network)
Obtaining the embedded vector of the positive sample with the category of 0 and the rest embedded vectors for comparison learning, firstly constructing a comparison embedded library
And is represented as:
（3）；
wherein, the first and the second end of the pipe are connected with each other,
and
is the query view (query view) and key view of all samples of the current batchThe graph (dictionary view) is respectively input into embedded vector sets obtained by a query network (query network) and a key network (dictionary network),
a queue of embedded vectors, i.e., a feature dictionary, is stored for training. During training, a positive type sample with the category of 0 in the current batch is randomly selected
Embedded vector of, unite
All the positive samples in (1) form a positive sample set
Training a model, a loss function for a specific training
Comprises the following steps:
（4）；
wherein the content of the first and second substances,
，
is a parameter of the temperature of the liquid crystal,
is from a positive sample set
The embedding vector of class 0 obtained in (c),
is an embedded vector obtained from the query network (query network),
is a contrast embedding warehouse
All the embedding vectors obtained from the key network (dictionary network);
the overall loss function in the network framework is:
（5）；
by using
The parameters of the query network (query network) are updated, and the key network (dictionary network) and the query network (query network) are updated by momentum, which is expressed as:
（6）；
wherein the content of the first and second substances,
is a parameter of the key network (dictionary network),
is a parameter of the query network (query network),
is a momentum coefficient. Finally, a queue is maintained to store the latest embedded vector
. Adding a classifier driving model behind a query network (inquiry network) for training, performing momentum updating by using a key network (dictionary network) and the query network (inquiry network), and acquiring an embedded vector according to the key network (dictionary network) and an embedded vector acquired by the query network (inquiry network)The input vectors are compared and learned, the key network (dictionary network) is updated through query network (inquiry network) parameters, all embedded vectors output by the two networks are combined for comparison and learning, and the 0/1 prediction output by the classifier can assist in correct defect judgment. Each sample is provided with 0/1 labels, namely a positive class sample and a defective negative class sample, and the query network (inquiry network) in the learning network is compared and classified.
Other parts of this embodiment are the same as any of embodiments 1 to 3, and thus are not described again.
Example 5:
in this embodiment, a gaussian density estimator is used to calculate the defect anomaly score so as to determine whether the sample has defects or not. When the test sample is used for defect positioning, an image is divided into image blocks according to a fixed step length, features are extracted from all the image blocks, abnormal score calculation is carried out on each image block, and the abnormal score is spread to each pixel by using a Gaussian smoothing method. For arbitrary samples
The defect judgment is carried out, and the anomaly detection is realized by using Gaussian density estimation as follows:
（7）；
wherein is
And
is the mean and variance obtained from the embedded vectors of the training data, m is the number of selected embedded vectors,
is a query network. The gaussian density estimator can determine whether the image is a defect image. When the image positioning defect is tested, the image is divided into image blocks according to a fixed step length, embedded vectors are extracted from all the image blocks, the abnormal score calculation is carried out on each image block, and the abnormal score is spread to each pixel by using a Gaussian smoothing method.
Other parts of this embodiment are the same as any of embodiments 1 to 4, and thus are not described again.
The above description is only a preferred embodiment of the present invention, and is not intended to limit the present invention in any way, and all simple modifications and equivalent variations of the above embodiments according to the technical spirit of the present invention are included in the scope of the present invention.
Claims (4)
1. A defect detection method based on self-supervision momentum contrast is characterized by comprising the following steps:
s1, constructing a Gaussian noise sample image as an original pseudo-defect image according to an original positive sample, a randomly generated mean value and a variance in a data set;
s2, generating a negative type sample by using a mixnoise operation;
s3, generating positive and negative sample characteristics which can obviously reflect the existence of defects by using a comparison learning loss function driving model, separating a normal sample from a negative sample, maintaining a characteristic dictionary by using a momentum comparison method, and learning the characterization of the positive and negative samples while keeping the consistency of the characteristics;
s4, calculating a defect abnormal score by using a Gaussian density estimator, and judging whether the whole image is a defect image according to the defect abnormal score to finish defect detection;
when the image positioning defect is tested, dividing the image into image blocks according to a fixed step length, extracting features from all the image blocks, calculating the abnormal score of each image block, and transmitting the calculated abnormal score to each pixel by using a Gaussian smoothing method;
randomly selecting an embedded vector of a sample in a training set to participate in the calculation of a total probability formula;
calculating the mean value and the variance of each embedded vector, and obtaining a total probability formula according to the obtained mean value and variance;
and for the test sample, calculating a defect abnormity score according to a total probability formula, and judging whether the whole image is a defect image according to the defect abnormity score to finish defect detection.
2. The method for defect detection based on the auto-supervised momentum contrast as recited in claim 1, wherein the step S1 comprises:
s11, combining the original positive samples in the data set, and constructing a Gaussian noise sample image by using a randomly generated mean value and variance;
step S12, generating a noise image by using a Perlin noise generator;
s13, performing binarization processing on the noise image through a uniformly and randomly sampled threshold value to form a defect distribution map;
and S14, generating an original pseudo-defect image by combining the normal sample and the Gaussian noise sample according to the defect distribution map.
3. The method for defect detection based on the auto-supervised momentum contrast as recited in claim 1, wherein the step S2 comprises:
randomly selecting a cutting area from an original pseudo-defect image, cutting a small rectangular area with variable size and aspect ratio, randomly rotating the small rectangular area and dithering pixel values, and then pasting the small rectangular area to a random position of a positive sample to form a pseudo-defect image as a negative sample.
4. The method for defect detection based on the auto-supervised momentum contrast as claimed in claim 1, wherein the step S3 comprises:
constructing a corresponding negative type sample for each positive type sample through a mixnoise operation, and generating a negative type sample set;
combining the positive and negative sample sets to create a sample space;
generating a query view and a key view by a random data expansion method, and sending the query view and the key view into a corresponding query network and a corresponding key network to obtain a pair of embedded vectors q and k;
and adding a classifier driving model behind the query network for training, performing momentum updating by using the key network and the query network, and performing comparison learning according to the embedded vector acquired by the key network and the embedded vector acquired by the query network.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210858635.XA CN115082435B (en) | 2022-07-21 | 2022-07-21 | Defect detection method based on self-supervision momentum contrast |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210858635.XA CN115082435B (en) | 2022-07-21 | 2022-07-21 | Defect detection method based on self-supervision momentum contrast |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115082435A CN115082435A (en) | 2022-09-20 |
| CN115082435B true CN115082435B (en) | 2022-11-22 |
Family
ID=83243835
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210858635.XA Active CN115082435B (en) | 2022-07-21 | 2022-07-21 | Defect detection method based on self-supervision momentum contrast |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115082435B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115690094B (en) * | 2022-12-12 | 2023-05-30 | 常州微亿智造科技有限公司 | Industrial defect detection method and system based on self-supervision network |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112966755A (en) * | 2021-03-10 | 2021-06-15 | 深圳市固电电子有限公司 | Inductance defect detection method and device and readable storage medium |
| CN112991330A (en) * | 2021-04-19 | 2021-06-18 | 征图新视(江苏)科技股份有限公司 | Knowledge distillation-based positive sample industrial defect detection method |
| CN113313684A (en) * | 2021-05-28 | 2021-08-27 | 北京航空航天大学 | Video-based industrial defect detection system under dim light condition |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11354778B2 (en) * | 2020-04-13 | 2022-06-07 | Google Llc | Systems and methods for contrastive learning of visual representations |
| US11386302B2 (en) * | 2020-04-13 | 2022-07-12 | Google Llc | Systems and methods for contrastive learning of visual representations |
| CN111982910B (en) * | 2020-07-06 | 2022-10-25 | 华南理工大学 | Weak supervision machine vision detection method and system based on artificial defect simulation |
| CN112365468A (en) * | 2020-11-11 | 2021-02-12 | 南通大学 | AA-gate-Unet-based offshore wind power tower coating defect detection method |
| CN112464879B (en) * | 2020-12-10 | 2022-04-01 | 山东易视智能科技有限公司 | Ocean target detection method and system based on self-supervision characterization learning |
| CN113032458A (en) * | 2021-03-23 | 2021-06-25 | 中国人民解放军63920部队 | Method and device for determining abnormality of spacecraft |
| CN114281985A (en) * | 2021-09-30 | 2022-04-05 | 腾讯科技（深圳）有限公司 | Sample feature space enhancement method and device |
-
2022
- 2022-07-21 CN CN202210858635.XA patent/CN115082435B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112966755A (en) * | 2021-03-10 | 2021-06-15 | 深圳市固电电子有限公司 | Inductance defect detection method and device and readable storage medium |
| CN112991330A (en) * | 2021-04-19 | 2021-06-18 | 征图新视(江苏)科技股份有限公司 | Knowledge distillation-based positive sample industrial defect detection method |
| CN113313684A (en) * | 2021-05-28 | 2021-08-27 | 北京航空航天大学 | Video-based industrial defect detection system under dim light condition |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115082435A (en) | 2022-09-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107330453B (en) | Pornographic image identification method based on step-by-step identification and fusion key part detection | |
| CN110852316B (en) | Image tampering detection and positioning method adopting convolution network with dense structure | |
| CN112183788A (en) | Domain adaptive equipment operation detection system and method | |
| US11686689B2 (en) | Automatic optimization of an examination recipe | |
| CN107169417B (en) | RGBD image collaborative saliency detection method based on multi-core enhancement and saliency fusion | |
| CN111027590B (en) | Breast cancer data classification method combining deep network features and machine learning model | |
| TW202221536A (en) | Generating training data usable for examination of a semiconductor specimen | |
| CN115082435B (en) | Defect detection method based on self-supervision momentum contrast | |
| CN114550110A (en) | Vehicle weight identification method and system based on unsupervised domain adaptation | |
| CN114972216A (en) | Construction method and application of texture surface defect detection model | |
| CN116188516A (en) | Training method of defect data generation model | |
| Dalal et al. | Autoregressive models: What are they good for? | |
| CN114820541A (en) | Defect detection method based on reconstructed network | |
| Li et al. | Fabric defect segmentation system based on a lightweight GAN for industrial Internet of Things | |
| CN117058716A (en) | Cross-domain behavior recognition method and device based on image pre-fusion | |
| CN113112482A (en) | PCB defect detection method based on attention mechanism network | |
| CN115546689A (en) | Video time sequence abnormal frame detection method based on unsupervised frame correlation | |
| CN114821174B (en) | Content perception-based transmission line aerial image data cleaning method | |
| CN115409822A (en) | Industrial part surface anomaly detection method based on self-supervision defect detection algorithm | |
| CN115661539A (en) | Less-sample image identification method embedded with uncertainty information | |
| CN111259859B (en) | Pedestrian re-recognition method based on combined variable picture generation | |
| CN112861670B (en) | Transmission line hardware detection method and system | |
| Ziga et al. | Non-parametric texture synthesis using texture classification | |
| Mo et al. | Contrastive adaptive frequency decomposition network guided by haze discrimination for real-world image dehazing | |
| TW202022883A (en) | Machine continuous learning method of neural network object classifier and related monitoring camera apparatus |
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 | ||
| GR01 | Patent grant |