CN117765363A - Image anomaly detection method and system based on lightweight memory bank - Google Patents
Image anomaly detection method and system based on lightweight memory bank Download PDFInfo
- Publication number
- CN117765363A CN117765363A CN202410194599.0A CN202410194599A CN117765363A CN 117765363 A CN117765363 A CN 117765363A CN 202410194599 A CN202410194599 A CN 202410194599A CN 117765363 A CN117765363 A CN 117765363A
- Authority
- CN
- China
- Prior art keywords
- image
- memory bank
- feature map
- anomaly
- product
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 62
- 238000000034 method Methods 0.000 claims abstract description 38
- 238000000605 extraction Methods 0.000 claims abstract description 29
- 230000004927 fusion Effects 0.000 claims abstract description 29
- 230000002159 abnormal effect Effects 0.000 claims abstract description 16
- 238000007781 pre-processing Methods 0.000 claims abstract description 14
- 238000012549 training Methods 0.000 claims description 22
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 claims description 15
- 238000004364 calculation method Methods 0.000 claims description 12
- 230000008569 process Effects 0.000 claims description 11
- 238000010586 diagram Methods 0.000 claims description 10
- 238000011176 pooling Methods 0.000 claims description 9
- 238000010845 search algorithm Methods 0.000 claims description 3
- 230000011218 segmentation Effects 0.000 claims description 2
- 238000012545 processing Methods 0.000 abstract description 2
- 239000000047 product Substances 0.000 description 21
- 230000006870 function Effects 0.000 description 12
- 230000005856 abnormality Effects 0.000 description 9
- 238000012360 testing method Methods 0.000 description 6
- 238000010606 normalization Methods 0.000 description 4
- 238000012795 verification Methods 0.000 description 4
- 238000013507 mapping Methods 0.000 description 3
- 238000004891 communication Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000004806 packaging method and process Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
- 239000013589 supplement Substances 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
Landscapes
- Image Analysis (AREA)
Abstract
The invention relates to the technical field of digital image processing anomaly detection, and provides an image anomaly detection method and system based on a lightweight memory bank. The method comprises the steps of obtaining a product image to be detected, preprocessing and dividing the product image to obtain a plurality of image blocks; based on each image block, extracting features by adopting a feature extraction network to obtain a fusion feature map; calculating an anomaly score for each image block; selecting a maximum anomaly score from the anomaly scores of each image block of the product image to be tested; calculating the distance between the fusion feature map of the image block and all feature maps stored in a memory bank based on the fusion feature map of the image block corresponding to the maximum anomaly score, and selecting the minimum distance; weighting the maximum anomaly score based on the minimum distance to obtain a final anomaly score; judging whether the abnormal score of the whole image is larger than a set threshold value, if so, the image of the product to be detected is abnormal, otherwise, the image of the product to be detected is normal.
Description
Technical Field
The invention relates to the technical field of digital image processing anomaly detection, in particular to an image anomaly detection method and system based on a light-weight memory bank.
Background
The statements in this section merely provide background information related to the present disclosure and may not necessarily constitute prior art.
The anomaly detection method of the industrial image is various, and a reconstruction-based method, a stream-based normalization method, a density estimation-based method, a memory bank-based method, a generation countermeasure network (GAN) -based method and the like are common, and each method has specific advantages and application scenes.
The memory bank-based anomaly detection method is a very efficient method, and PatchCore is a representative work in the field of industrial anomaly detection based on the memory bank method. The method realizes accurate abnormality detection by using a memory bank for storing normal sample characteristics. In addition, the memory-based abnormality detection method has many advantages, and does not need a large number of abnormal samples to train, only needs to extract and store the characteristics of normal samples, and compares the characteristics of the input samples with the characteristics in the memory during detection so as to judge whether the input samples are abnormal.
When image anomaly detection is carried out, the anomaly detection method based on the memory bank generally uses a pre-trained model based on the ImageNet to carry out feature extraction, but in the industrial field, large differences exist between industrial products and environments and an ImageNet data set, so that the detection accuracy is not facilitated to be improved. In addition, the conventional memory-based method requires a large amount of memory space for the memory bank because of the large number of features to be stored, and also takes a large amount of reasoning time because of the comparison with the large number of stored features during the test.
Disclosure of Invention
The invention provides an image anomaly detection method and system based on a lightweight memory bank, which solve the problems that the memory bank occupies too large storage space and the like in the field of image anomaly detection and have application capability for industrial anomaly detection.
In order to achieve the above purpose, the present invention adopts the following technical scheme:
The first aspect of the invention provides an image anomaly detection method based on a lightweight memory bank.
an image anomaly detection method based on a lightweight memory bank comprises the following steps:
acquiring a product image to be detected, preprocessing and dividing the product image to obtain a plurality of image blocks;
based on each image block, extracting features by adopting a feature extraction network to obtain a fusion feature map;
calculating an anomaly score for each image block;
Selecting a maximum anomaly score from the anomaly scores of each image block of the product image to be tested; calculating the distance between the fusion feature map of the image block and all feature maps stored in a memory bank based on the fusion feature map of the image block corresponding to the maximum anomaly score, and selecting the minimum distance;
Weighting the maximum anomaly score based on the minimum distance to obtain a final anomaly score;
judging whether the abnormal score of the whole image is larger than a set threshold value, if so, the image of the product to be detected is abnormal, otherwise, the image of the product to be detected is normal.
further, the preprocessing and segmentation process includes: preprocessing a product image to be processed by adopting a bilinear interpolation method to obtain tensors; and dividing the tensor according to the preset size to obtain a plurality of image blocks with the same size.
further, after the fusion feature map is obtained, pooling dimension reduction is performed on the fusion feature map.
Further, the feature extraction network comprises an initialization convolution module, a main branch network and an auxiliary branch network; the initialization convolution network is used for extracting a characteristic diagram of the image block; the main branch network is used for extracting the features of different layers of the feature map to obtain a main branch feature map; the auxiliary branch network is used for extracting features of different scales of the feature map to obtain an auxiliary branch feature map; and splicing the main branch feature map and the auxiliary branch feature map to obtain a fusion feature map.
Still further, the main branch network includes three residual blocks connected in sequence, the three residual blocks having the same structure and sequentially increasing in number of channels.
still further, the auxiliary branch network includes two convolution modules and a pooling module.
further, the process of calculating the anomaly score of each image block includes: and calculating the anomaly score of each image block by adopting a nearest neighbor search algorithm.
further, the memory bank is obtained by using an existing product image training feature extraction network.
furthermore, in the training process, the fusion feature images obtained by training are stored in a memory bank, the similarity of each fusion feature image and the feature images in the memory bank is calculated in the training process, and the fusion feature images with the similarity value smaller than a preset threshold value are added into the memory bank.
The second aspect of the invention provides an image anomaly detection system based on a lightweight memory bank.
an image anomaly detection system based on a lightweight memory bank, comprising:
a data acquisition module configured to: acquiring a product image to be detected, preprocessing and dividing the product image to obtain a plurality of image blocks;
a feature extraction module configured to: based on each image block, extracting features by adopting a feature extraction network to obtain a fusion feature map;
an anomaly calculation module configured to: calculating an anomaly score for each image block;
a distance calculation module configured to: selecting a maximum anomaly score from the anomaly scores of each image block of the product image to be tested; calculating the distance between the fusion feature map of the image block and all feature maps stored in a memory bank based on the fusion feature map of the image block corresponding to the maximum anomaly score, and selecting the minimum distance;
a weighting module configured to: weighting the maximum anomaly score based on the minimum distance to obtain a final anomaly score;
An anomaly detection module configured to: judging whether the abnormal score of the whole image is larger than a set threshold value, if so, the image of the product to be detected is abnormal, otherwise, the image of the product to be detected is normal.
compared with the prior art, the invention has the beneficial effects that:
The invention solves the problems of overlarge occupied storage space of the memory bank in the field of image anomaly detection and the like, and has application capability for industrial anomaly detection.
The invention designs the characteristic extraction network in the industrial scene, and can effectively extract the characteristics of the industrial product image.
The invention provides a weighted anomaly score calculation method which can effectively reduce erroneous judgment and further improve the accuracy of image anomaly detection.
The invention provides a light-weight image anomaly detection scheme which can realize rapid and accurate anomaly detection of industrial images.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the invention.
FIG. 1 is a detection flow chart of an anomaly detection model shown in the present invention;
FIG. 2 is a diagram of a feature extraction network of an anomaly detection model shown in the present invention;
FIG. 3 is a flow chart of a lightweight memory bank module for constructing an anomaly detection model shown in the present invention;
fig. 4 is a flowchart of anomaly score calculation for the anomaly detection model shown in the present invention.
Detailed Description
The invention will be further described with reference to the drawings and examples.
It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the invention. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.
it is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of exemplary embodiments according to the present invention. As used herein, the singular is also intended to include the plural unless the context clearly indicates otherwise, and furthermore, it is to be understood that the terms "comprises" and/or "comprising" when used in this specification are taken to specify the presence of stated features, steps, operations, devices, components, and/or combinations thereof.
It is noted that the flowcharts and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of the present disclosure. It should be noted that each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the flowchart illustrations and/or block diagrams, and combinations of blocks in the flowchart illustrations and/or block diagrams, can be implemented by special purpose hardware-based systems which perform the specified functions or operations, or combinations of special purpose hardware and computer instructions.
Example 1
As shown in fig. 1, the present embodiment provides an image anomaly detection method based on a lightweight memory bank, and the present embodiment is illustrated by applying the method to a server, and it can be understood that the method may also be applied to a terminal, and may also be applied to a system and a terminal, and implemented through interaction between the terminal and the server. The server can be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, and can also be a cloud server for providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network servers, cloud communication, middleware services, domain name services, security services CDNs, basic cloud computing services such as big data and artificial intelligent platforms and the like. The terminal may be, but is not limited to, a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, etc. The terminal and the server may be directly or indirectly connected through wired or wireless communication, and the present application is not limited herein. In this embodiment, the method includes the steps of:
S1, shooting a large number of industrial product images in an industrial scene, preprocessing, and finishing into a data set for training a feature extraction network;
s2, building a feature extraction network, and training the feature extraction network;
s3, arranging images for anomaly detection in an industrial scene into a data set, preprocessing, and dividing the data set into a training set, a verification set and a test set;
s4, a feature extraction module of the anomaly detection model is completed by using a pre-training network;
S5, fusing the multi-scale features and reducing the dimension;
s6, constructing a lightweight memory bank through a self-defined symmetrical kernel function, and storing representative characteristics;
s7, after training is finished, adopting a weighted anomaly score calculation method to prevent erroneous judgment during image anomaly detection;
S8, selecting a model with fewer storage characteristic diagrams and higher accuracy for packaging deployment.
Specifically, S1 includes:
product images are taken from different angles and conditions in an industrial environment to increase the richness of the data set and the taken images are classified and labeled.
The images were divided into training, validation and test sets at a ratio of 70%, 15% and 15%. The file path and corresponding label of each image are recorded with one folder to store the image and one CSV file, respectively.
Specifically, S2 includes:
As shown in fig. 2, the feature extraction network includes an initialization convolution module and two branch modules, the main branch including three residual blocks designed to capture features of different levels of an image.
The initialization convolution module comprises a3 x 3 convolution layer and a normalization layer. For input image tensorsthe process is described as follows:
。
Wherein,representing the convolution kernel as a convolution of 3 x 3,/>Representation normalization,/>Is a feature map obtained after the initial convolution.
for three residual blocks, each residual block contains the same structure, but sequentially increases in number of channels. Taking the first residual block as an example, taking the feature map obtained by initial convolutionAs input, sequentially:
；
；
。
Wherein,representing the convolution kernel as a convolution of 3 x 3,/>A convolution with a convolution kernel of 1 x 1 is represented,representing a Relu activation function,/>representing normalization operations,/>、/>、/>output corresponding to three operations in the first residual block,/>, respectivelyIs the output corresponding to the first residual block.
And then sequentially obtain、/>the second and third residual block outputs are characterized, respectively. />, in the first residual block、/>、/>Merging, record/>the feature extraction function of the second and third residual blocks is similarly denoted as/>, respectively, for the feature extraction function of the first residual blockAnd/>。
The feature extraction network captures different scale features of the image by using auxiliary branches, which are different in scale and type from the features of the main branches, and can be regarded as a supplement to the main branches, thereby improving its robustness and accuracy.
specifically, the auxiliary branches mainly comprise two convolution modules and a pooling module, and the characteristic diagram obtained after the convolution is initializedThe auxiliary branch may be described as the following procedure:
；
；
。
Wherein,representing the output of the first convolution module,/>Representing the output of the second layer,/>representing maximum pooling,/>Representing the output of the auxiliary branch.
Features of the main branchFeatures of auxiliary branches/>and splicing to obtain final features, mapping the features to obtain probabilities of various categories by using a full-connection layer, and training a network by using a cross entropy loss and Adam optimizer.
And saving the trained weight as a pth file for the subsequent abnormality detection task.
specifically, S3 includes:
the preprocessing operation of the picture for anomaly detection in the industrial scene is mainly to adjust the image size. The purpose is to ensure that the size of each input image is equal, thereby maintaining data consistency. Assume an original image∈target image/>∈/>mapping is performed by the following formula：
。
Wherein,Is the coordinates in the original image,/>is the corresponding coordinates in the new image,/>、/>Is the width and height of the original image and W, H is the width and height of the target image.
Since this mapping may not always fall on integer pixel coordinates, interpolation is required to determine the value of each pixel in the new image, and the anomaly detection model uses bilinear interpolation to adjust the image to a specified width and height (256256). Specifically, interpolation is achieved by the following formula:
。
Wherein,Representing bilinear interpolation,/>Is the coordinates in the original image and,Is the coordinates in the target image.
Because of the use of unsupervised learning, the training set only contains images without anomalies and no labels are needed, and in the test set and the verification set, the training set contains images without anomalies as well as images without anomalies and has corresponding labels. After the pretreatment, the training set, the verification set and the test set are divided according to the proportion of 60%, 20% and 20%.
Specifically, S4 includes:
the image for abnormality detection is converted into tensors after preprocessing, and then the tensors are divided into a series of patch forms. For a set of tensors∈/>The method comprises the steps of carrying out a first treatment on the surface of the Where B is the batch size, C is the channel number, and H and W are the height and width of the image, respectively. Given patch size/>Sum step/>dividing the tensor corresponding to each picture according to the following formula:
。
Wherein,is the patch index in the vertical direction,/>，/>Is a patch index in the horizontal direction,/>，/>、/>the steps in the vertical and horizontal directions are indicated,Indicating the vertical and horizontal dimensions of the patch.
the anomaly detection model sets the size of each patch to be (16, 16), the step size to be (16, 16), and the entire image to be uniformly covered without overlapping portions. After the representation of each patch is obtained, the anomaly detection model will perform feature extraction on each patch using a pre-trained feature extraction network.after initializing the convolution layer, get/>Three residual blocks, after which the main branch is taken, are shown in the following formula:
。
Wherein,、/>、/>representing the outputs of the first, second and third residual blocks respectively,、/>And/>representing its corresponding feature extraction function,/>is a feature map obtained after initializing the convolutional layer.
Obtaining features by auxiliary branchingThe process is as follows:
。
Wherein,、/>And/>respectively correspond to three operations in the auxiliary branch,/>and initializing a characteristic diagram obtained after the convolution layer.
specifically, S5 includes:
Firstly, extracting a network according to pre-training features to obtain a fused feature map by an abnormality detection model, and carrying out feature map extraction on the fused feature map、/>Splicing to generate/>, which contains multi-scale information：
。
Wherein,representing stitching along the channel dimension.
For fusion featuresthe dimension is reduced by the pooling operation while preserving important spatial information, in particular, the following pooling will be used to obtain:
。
Wherein,Representing adaptive average pooling operations,/>Representing the characteristics of the patch.
specifically, S6 includes:
as shown in FIG. 3, the anomaly detection model uses custom symmetric kernel functions to construct a lightweight memory bank that stores only a small number of representative features from the beginning of training to the end of training. Each patch featurekey information of an image can be accurately captured and recorded as a feature to be embedded/>。
In order to realize a lightweight memory bank, the anomaly detection model uses a self-defined kernel function to calculate the similarity, so that the representativeness of storage characteristics is ensured, and the number of the stored characteristic graphs is obviously reduced. Assume that the feature set stored in the memory bank at this time is. The anomaly detection model is calculated/>, by the following formulaAnd/>similarity of features in/>and selecting the maximum value/>, of the similarity：
。
Wherein,，/>Is/>Is a feature of/>Representation/>And/>Similarity of/>is an adjustable parameter that controls the sensitivity of the kernel function,/>Is the maximum of all similarities.
Similarity degreethe lower the feature difference is, the larger the anomaly detection model is by ensuring/>below a preset threshold/>only when this condition is met will a new embedded vector be added to the memory bank, guaranteeing the representativeness of the stored feature. The process can be expressed by the following formula:
。
Wherein,is possible to add to/>S represents the similarity maximum.
If all ofAre all lower than/>then/>For/>Otherwise is/>(representing no vector added). Finally, update the collection/>, of stored features in the memory library：/>。
specifically, S7 includes:
as shown in FIG. 4, after training, the anomaly detection model first obtains the features of the image patch during the verification and test phasesThen, a weighted anomaly score calculation method is adopted to prevent erroneous judgment when image anomaly detection is realized.
if one patch in the image is abnormal, the whole image is regarded as abnormal, and the abnormality detection model realizes fine-grained abnormality detection. The anomaly detection model firstly obtains the anomaly score of each patch through the existing nearest neighbor search method:
。
Wherein,is a nearest neighbor search algorithm,/>A set of anomaly scores comprising a patch,/>Is the patch feature to be detected.
Will beis marked as/>Its corresponding anomaly score/>is marked as/>. In some cases,/>The higher probability is only caused by some external condition (such as illumination, etc.), and the anomaly score/>, which is obtained only based on the nearest neighbor searchTo determine the true degree of abnormality of the image is insufficient.
for each ofselect/>Selecting the maximum value/>Its corresponding embedded feature is/>. By calculating/>The degree of abnormality of the patch is further evaluated from all other embedded distances in the memory bank, calculated/>, by the following formulaEuclidean distance/>, to all stored features in a memory bankand get the minimum value/>The formula is as follows:
。
Wherein,representative/>Value in the ith dimension,/>representing the value of the jth feature in the memory bank in the ith dimension, and D representing the dimension of the feature space.
anomaly detection model pair by using softmax functionWeighting is performed to reduce false positives, such as the following formula:
。
Wherein,Representing the number of patches,/>Is/>, of the ith patch，/>Is the ith patch，/>Is the anomaly score for the entire image.
Specifically, S8 includes:
the feature images stored in the memory library are controlled by controlling the threshold value of the similarity, the size of the memory library is effectively reduced, and a model with fewer stored feature images and higher accuracy is selected according to the detection result to perform package deployment.
Example two
the embodiment provides an image anomaly detection system based on a lightweight memory bank.
an image anomaly detection system based on a lightweight memory bank, comprising:
a data acquisition module configured to: acquiring a product image to be detected, preprocessing and dividing the product image to obtain a plurality of image blocks;
a feature extraction module configured to: based on each image block, extracting features by adopting a feature extraction network to obtain a fusion feature map;
an anomaly calculation module configured to: calculating an anomaly score for each image block;
a distance calculation module configured to: selecting a maximum anomaly score from the anomaly scores of each image block of the product image to be tested; calculating the distance between the fusion feature map of the image block and all feature maps stored in a memory bank based on the fusion feature map of the image block corresponding to the maximum anomaly score, and selecting the minimum distance;
a weighting module configured to: weighting the maximum anomaly score based on the minimum distance to obtain a final anomaly score;
An anomaly detection module configured to: judging whether the abnormal score of the whole image is larger than a set threshold value, if so, the image of the product to be detected is abnormal, otherwise, the image of the product to be detected is normal.
here, the data acquisition module, the feature extraction module, the anomaly calculation module, the distance calculation module, the weighting module, and the anomaly detection module are the same as the examples and the application scenarios implemented by the steps in the first embodiment, but are not limited to the disclosure of the first embodiment. It should be noted that the modules described above may be implemented as part of a system in a computer system, such as a set of computer-executable instructions.
The above description is only of the preferred embodiments of the present invention and is not intended to limit the present invention, but various modifications and variations can be made to the present invention by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (10)
1. An image anomaly detection method based on a lightweight memory bank is characterized by comprising the following steps:
acquiring a product image to be detected, preprocessing and dividing the product image to obtain a plurality of image blocks;
based on each image block, extracting features by adopting a feature extraction network to obtain a fusion feature map;
calculating an anomaly score for each image block;
Selecting a maximum anomaly score from the anomaly scores of each image block of the product image to be tested; calculating the distance between the fusion feature map of the image block and all feature maps stored in a memory bank based on the fusion feature map of the image block corresponding to the maximum anomaly score, and selecting the minimum distance;
Weighting the maximum anomaly score based on the minimum distance to obtain a final anomaly score;
judging whether the abnormal score of the whole image is larger than a set threshold value, if so, the image of the product to be detected is abnormal, otherwise, the image of the product to be detected is normal.
2. The method for detecting image anomalies based on a lightweight memory bank according to claim 1, wherein the preprocessing and segmentation process includes: preprocessing a product image to be processed by adopting a bilinear interpolation method to obtain tensors; and dividing the tensor according to the preset size to obtain a plurality of image blocks with the same size.
3. the method for detecting image anomalies based on a lightweight memory bank according to claim 1, further comprising, after obtaining the fused feature map, pooling and dimension-reducing the fused feature map.
4. The method for detecting image anomalies based on a lightweight memory bank according to claim 1, wherein the feature extraction network comprises an initialization convolution module, a main branch network, and an auxiliary branch network; the initialization convolution network is used for extracting a characteristic diagram of the image block; the main branch network is used for extracting the features of different layers of the feature map to obtain a main branch feature map; the auxiliary branch network is used for extracting features of different scales of the feature map to obtain an auxiliary branch feature map; and splicing the main branch feature map and the auxiliary branch feature map to obtain a fusion feature map.
5. the method for detecting image anomalies based on a lightweight memory bank according to claim 4, wherein the main branch network includes three residual blocks connected in sequence, the three residual blocks being identical in structure and increasing in number of channels in sequence.
6. The method for detecting image anomalies based on a lightweight memory bank according to claim 4, wherein the auxiliary branch network comprises two convolution modules and one pooling module.
7. The method for detecting image anomalies based on a lightweight memory bank according to claim 1, wherein the process of calculating the anomaly score for each image block includes: and calculating the anomaly score of each image block by adopting a nearest neighbor search algorithm.
8. The method for detecting image anomalies based on a lightweight memory bank of claim 1, wherein said memory bank is obtained by employing an existing product image training feature extraction network.
9. The method for detecting image anomalies based on a lightweight memory according to claim 8, wherein in the training process, the fusion feature images obtained by training are stored in the memory, and in the training process, the similarity between each fusion feature image and the feature images in the memory is calculated, and the fusion feature images with similarity values smaller than a preset threshold are added into the memory.
10. an image anomaly detection system based on a lightweight memory bank, comprising:
a data acquisition module configured to: acquiring a product image to be detected, preprocessing and dividing the product image to obtain a plurality of image blocks;
a feature extraction module configured to: based on each image block, extracting features by adopting a feature extraction network to obtain a fusion feature map;
an anomaly calculation module configured to: calculating an anomaly score for each image block;
a distance calculation module configured to: selecting a maximum anomaly score from the anomaly scores of each image block of the product image to be tested; calculating the distance between the fusion feature map of the image block and all feature maps stored in a memory bank based on the fusion feature map of the image block corresponding to the maximum anomaly score, and selecting the minimum distance;
a weighting module configured to: weighting the maximum anomaly score based on the minimum distance to obtain a final anomaly score;
An anomaly detection module configured to: judging whether the abnormal score of the whole image is larger than a set threshold value, if so, the image of the product to be detected is abnormal, otherwise, the image of the product to be detected is normal.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410194599.0A CN117765363A (en) | 2024-02-22 | 2024-02-22 | Image anomaly detection method and system based on lightweight memory bank |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410194599.0A CN117765363A (en) | 2024-02-22 | 2024-02-22 | Image anomaly detection method and system based on lightweight memory bank |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117765363A true CN117765363A (en) | 2024-03-26 |
Family
ID=90314732
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410194599.0A Pending CN117765363A (en) | 2024-02-22 | 2024-02-22 | Image anomaly detection method and system based on lightweight memory bank |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117765363A (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112508137A (en) * | 2021-02-09 | 2021-03-16 | 南方电网数字电网研究院有限公司 | Transformer abnormality detection method and device, computer equipment and storage medium |
| CN112905844A (en) * | 2021-03-23 | 2021-06-04 | 广东工业大学 | Video voice retrieval method |
| CN115578585A (en) * | 2022-09-30 | 2023-01-06 | 成都数之联科技股份有限公司 | Industrial image anomaly detection method, system, computer device and storage medium |
| CN115830006A (en) * | 2023-02-03 | 2023-03-21 | 山东锋士信息技术有限公司 | Anomaly detection method for improving hypersphere space learning based on neighbor comparison |
| CN117557872A (en) * | 2024-01-12 | 2024-02-13 | 苏州大学 | Unsupervised anomaly detection method and device for optimizing storage mode |
| CN117576079A (en) * | 2023-12-15 | 2024-02-20 | 华中科技大学 | Industrial product surface abnormality detection method, device and system |
-
2024
- 2024-02-22 CN CN202410194599.0A patent/CN117765363A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112508137A (en) * | 2021-02-09 | 2021-03-16 | 南方电网数字电网研究院有限公司 | Transformer abnormality detection method and device, computer equipment and storage medium |
| CN112905844A (en) * | 2021-03-23 | 2021-06-04 | 广东工业大学 | Video voice retrieval method |
| CN115578585A (en) * | 2022-09-30 | 2023-01-06 | 成都数之联科技股份有限公司 | Industrial image anomaly detection method, system, computer device and storage medium |
| CN115830006A (en) * | 2023-02-03 | 2023-03-21 | 山东锋士信息技术有限公司 | Anomaly detection method for improving hypersphere space learning based on neighbor comparison |
| CN117576079A (en) * | 2023-12-15 | 2024-02-20 | 华中科技大学 | Industrial product surface abnormality detection method, device and system |
| CN117557872A (en) * | 2024-01-12 | 2024-02-13 | 苏州大学 | Unsupervised anomaly detection method and device for optimizing storage mode |
Non-Patent Citations (1)
| Title |
|---|
| 张小红: "《基于深度学习的道路场景语义分割研究》", 15 February 2024 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111950453B (en) | Random shape text recognition method based on selective attention mechanism | |
| CN107529650B (en) | Closed loop detection method and device and computer equipment | |
| CN109840556B (en) | Image classification and identification method based on twin network | |
| CN112150493B (en) | Semantic guidance-based screen area detection method in natural scene | |
| CN112115783A (en) | Human face characteristic point detection method, device and equipment based on deep knowledge migration | |
| CN109871821B (en) | Pedestrian re-identification method, device, equipment and storage medium of self-adaptive network | |
| CN113505792B (en) | Multi-scale semantic segmentation method and model for unbalanced remote sensing image | |
| CN111814794A (en) | Text detection method and device, electronic equipment and storage medium | |
| CN113095370B (en) | Image recognition method, device, electronic equipment and storage medium | |
| CN110991321B (en) | Video pedestrian re-identification method based on tag correction and weighting feature fusion | |
| CN113076871A (en) | Fish shoal automatic detection method based on target shielding compensation | |
| JP2017062778A (en) | Method and device for classifying object of image, and corresponding computer program product and computer-readable medium | |
| CN111160407A (en) | Deep learning target detection method and system | |
| CN110598788A (en) | Target detection method and device, electronic equipment and storage medium | |
| CN113487610B (en) | Herpes image recognition method and device, computer equipment and storage medium | |
| CN116580257A (en) | Feature fusion model training and sample retrieval method and device and computer equipment | |
| CN113780145A (en) | Sperm morphology detection method, sperm morphology detection device, computer equipment and storage medium | |
| CN116977674A (en) | Image matching method, related device, storage medium and program product | |
| CN115147418A (en) | Compression training method and device for defect detection model | |
| CN111931572B (en) | Target detection method for remote sensing image | |
| CN111582057B (en) | Face verification method based on local receptive field | |
| CN113704276A (en) | Map updating method and device, electronic equipment and computer readable storage medium | |
| CN113627481A (en) | Multi-model combined unmanned aerial vehicle garbage classification method for smart gardens | |
| CN113129332A (en) | Method and apparatus for performing target object tracking | |
| CN116129417A (en) | Digital instrument reading detection method based on low-quality 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 |