CN109872313A - A kind of method for detecting surface defects of products based on depth convolution self-encoding encoder - Google Patents
A kind of method for detecting surface defects of products based on depth convolution self-encoding encoder Download PDFInfo
- Publication number
- CN109872313A CN109872313A CN201910118229.8A CN201910118229A CN109872313A CN 109872313 A CN109872313 A CN 109872313A CN 201910118229 A CN201910118229 A CN 201910118229A CN 109872313 A CN109872313 A CN 109872313A
- Authority
- CN
- China
- Prior art keywords
- picture
- encoding encoder
- depth convolution
- convolution self
- training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Landscapes
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of method for detecting surface defects of products based on depth convolution self-encoding encoder comprising following steps: step 1, and the non-defective unit picture of defect is not included using imaging device acquisition；Step 2 is trained using depth convolution self-encoding encoder network model and training dataset；Step 3 rebuilds each slice map that test data is concentrated using the complete depth convolution self-encoding encoder of training, and calculate rebuild before with reconstruction after two picture pixels difference difference parameter；Step 4 rebuilds the slice use of picture to be detected, calculates the difference parameter of the difference of two picture pixels after rebuilding preceding and reconstruction；Step 5: carrying out the diminution of multiple multiplying powers to non-defective unit figure, the non-defective unit figure of each reduction magnification repeat more than process.
Description
Technical field
The present invention relates to machine vision appearance detection fields, more particularly to a kind of production based on depth convolution self-encoding encoder
Product detection method of surface flaw.
Background technique
Machine vision is the technology and methods for replacing human eye with machine to measure and detect.Machine vision technique can be used for
Substitution is artificial to carry out automated production quality testing.Ordinary circumstance, which carries out product quality detection using machine vision, can mainly divide
It is detected at size detection and appearance.
In machine vision product appearance detection field.Work is needed when developing new projects using traditional non-neural network method
Cheng Shi debugs a large amount of Image Processing parameter, and which results in new projects' development cycle is long.And if it is desired to replacing different patterns
Product (such as the cloth for changing a kind of pattern) when engineer may be needed to debug Image Processing parameter again, in some cases
This, which may cause this kind of Machine Vision Inspecting System, can not put into actual production use.
Development with neural network in the application of field of machine vision, occur many neural network based having prison
Machine vision method is superintended and directed, for example is based on convolutional neural networks (CNN) or BP neural network, the application of such methods solves perhaps
More non-very high product appearance test problems of the insurmountable complexity of neural network machine visible sensation method of tradition.It is this kind of to have prison
Superintending and directing method, there is also apparent defects.First the disadvantage is that the picture for needing to collect a large amount of non-defective unit and defective products carries out nerve
Network training could improve recognition correct rate, however be not easy to collect largely in the case where most machine vision actually uses scene
Defective products picture, the quantity of usual non-defective unit can need to control when hands-on not of the same race far more than the quantity of defective products
It is excessive that the training picture ratio of class is unable to difference.Under normal circumstances in order to by the misjudgment rate of supervised learning neural network drop
The low training figure that ten times of quantity are generally required for original 1/10th.So having supervision machine view using neural network based
Feel method can encounter the big problem of development difficulty in practical item development, encounter the item for needing to be replaced frequently different pattern products
Engineer may be needed to resurvey the image re -training of a large amount of novel type products when mesh, in some cases this kind of machine view
Feel that detection technique can not be applied in actual production.
In neural network supervised learning field, using the neural network of depth convolution in standard in terms of handling digital picture
In exactness and speed better than do not have using depth convolution neural network (such as BP neural network), after study with actual test
It was found that being better than not encoding certainly using depth convolution using effect of the depth convolution self-encoding encoder for Digital Image Processing
Device.Machine learning method can be made to cannot be considered in terms of image usually using the network model and picture pretreatment mode of different parameters
It is global and local.
Summary of the invention
Technical problem to be solved by the invention is to provide a kind of surface defects of products based on depth convolution self-encoding encoder
Detection method can greatly improve the treatment effect to processing digital picture using convolution self-encoding encoder.
The present invention is that above-mentioned technical problem is solved by following technical proposals: one kind being based on depth convolution self-encoding encoder
Method for detecting surface defects of products, which is characterized in that itself the following steps are included:
Step 1 is not included the non-defective unit picture of defect using imaging device acquisition, non-defective unit picture is divided into two by a certain percentage
Part, for a part for training depth convolution self-encoding encoder, another part is used for test depth convolution self-encoding encoder training effect
And the defect for calculating reference distinguishes threshold value；All pictures are sliced, training dataset and test data set are obtained；
Step 2 is trained using depth convolution self-encoding encoder network model and training dataset；
Step 3 carries out weight to each slice map that test data is concentrated using the complete depth convolution self-encoding encoder of training
Build, and calculate rebuild before with reconstruction after two picture pixels difference difference parameter；
Step 4 rebuilds the slice use of picture to be detected, calculate reconstruction before with rebuild after two picture pixels it
The difference parameter of difference；Picture is divided into non-defective unit and defective products according to the value of difference parameter, and obtains defects detection result；
Step 5: carrying out the diminution of multiple multiplying powers to non-defective unit figure, and the non-defective unit figure of each reduction magnification repeats above
Process；Step 1 is repeated to step 4 until having handled all zoom ratios, a certain can be calculated under different reduction ratios
Whether picture to be detected includes defect；Picture to be detected will also carry out the diminution of corresponding ratio in repeating process；Thus it is tieing up
Hold neural network structure it is constant in the case where the object surface structure under different scale is detected.
Preferably, the step 2 is trained using depth convolution self-encoding encoder neural network and training dataset.
Preferably, the picture to be detected is to be clapped in the process using industrial camera or other imaging devices using actually detected
That takes the photograph is used for the whether defective picture of testing product.
Preferably, the imaging device industrial camera.
Preferably, the step 5 is scanned picture to be detected using sliding window method, persistently in window
It is sliced picture and carries out defects detection.
Preferably, the step 3 finds reference of the maximum difference parameter as Threshold segmentation.
Preferably, the depth convolution self-encoding encoder, which uses, is divided into coded portion and the two-part network architecture of decoding.
The positive effect of the present invention is that: the product defects inspection proposed by the present invention based on depth convolution self-encoding encoder
Survey method can apply to highly complex texture image, such as wood texture and textile fiber structure, these are highly complex, high
The abstract image of degree is that the conventional machines visible sensation method of nerual network technique is not used to be difficult to accurately handle.The present invention uses
Convolution self-encoding encoder can greatly improve the treatment effect to processing digital picture.There is supervision neural network method to need artificial acquisition
Defect map, the acquisition difficulty of defect map is much higher than the acquisition difficulty of non-defective unit figure under normal circumstances, this imitates the exploitation for the project that restricts
Rate and the development cost for greatly improving project.There is supervision neural network method can not in the project that some testing products often more become
The product of automatic study novel type, leads to not the exploitation for these projects.Method proposed by the present invention belongs to unsupervised
It practises, training neural network need to only provide the non-defective unit picture for being easy to acquire.It also can be automatic after the product to be detected of replacement novel type
It relearns.Traditional supervised learning neural network method needs manually to carry out the big of defect after the product of replacement new varieties
The defect picture collection of amount is to guarantee to detect accuracy.The acquisition difficulty of defect picture is high under normal circumstances, generates defective
Quantity can be far below the quantity of non-defective unit, and be difficult to collect the defect of A wide selection of colours and designs.Using technology proposed by the present invention more
The product to be detected for renewing kind only need to acquire the non-defective unit picture easily obtained later under manual oversight, can be automatically performed training
It is calculated with parameter, starts the defects detection for carrying out new varieties product.The present invention is optimized and changes on the basis of depth convolution
Self-encoding encoder network structure that is good, having used algorithm to generate adjusts the complexity of neural network freely to solve
The product appearance detection project of different level of abstractions and complexity.Some defects are macroscopically being easier discovery, some defects
It is microcosmic it is upper be easier to find, by original image being carried out the diminution of different proportion and slice is cut out, use identical neural network
Model training goes out the neural network weight under different zoom ratio, and method proposed by the present invention is allowed to have higher accuracy, and
And the number of parameters for needing to debug is decreased, it is more intelligent and efficient.
Detailed description of the invention
Fig. 1 is the network architecture schematic diagram of depth convolution self-encoding encoder.
Fig. 2 is a kind of network structure of depth convolution self-encoding encoder disclosed by the embodiments of the present invention.
Fig. 3 is showing for the result that the depth convolution self-encoding encoder after present invention training rebuilds three slice maps respectively
It is intended to.
Fig. 4 is that the present invention is based on the flow charts of the method for detecting surface defects of products of depth convolution self-encoding encoder.
Specific embodiment
Present pre-ferred embodiments are provided with reference to the accompanying drawing, in order to explain the technical scheme of the invention in detail.
As shown in Figures 1 to 4, the present invention is based on the method for detecting surface defects of products of depth convolution self-encoding encoder include with
Lower step:
Step 1 is not included the non-defective unit picture of defect using imaging device acquisition, non-defective unit picture is divided into two by a certain percentage
Part, for a part for training depth convolution self-encoding encoder, another part is used for test depth convolution self-encoding encoder training effect
And the defect for calculating reference distinguishes threshold value；All pictures are sliced, training dataset and test data set are obtained.Tool
Steps are as follows for body: non-defective unit picture is the non-defective unit picture not comprising defect shot using industrial camera or other imaging devices,
It can be black and white picture or multichannel picture, such as color image.Non-defective unit image is divided into two parts by a certain percentage, one
Part is for training self-encoding encoder, and another part is used to test self-encoding encoder training effect and the defect for calculating reference is distinguished
Threshold value.All pictures are sliced, training non-defective unit slice map and test non-defective unit slice map are generated, between slice and slice
Allow to exist and be overlapped, thus to obtain training dataset and test data set.The size and depth convolution being sliced under normal circumstances are certainly
Encoder neural network input image size is consistent, and special circumstances, which cannot keep, unanimously then to be needed to contract to slice picture
It puts to match neural network input image size.
Step 2 is trained using depth convolution self-encoding encoder network model and training dataset, and self-encoding encoder is one
Kind is mainly used for Data Dimensionality Reduction or extracts the unsupervised learning neural network of feature, and convolution self-encoding encoder is in traditional self-encoding encoder
On the basis of combine convolutional neural networks convolution sum pondization operation, when handling image data, effect is more preferable.Depth convolution is certainly
The network architecture schematic diagram of encoder as shown in Figure 1, be divided into coded portion and decoding two parts, coded portion by picture compression at
Characteristic pattern is restored reconstruction figure by characteristic pattern, decoded portion.Use example depth convolution as shown in Figure 2 encodes certainly when hands-on
Device neural network and training dataset are trained, and neural network weight is saved after the completion of training.Training with python language and
Keras is example as backend, it is necessary first to which then specific neural network, Zhi Houshe are built in the library python for importing dependence
The path that trained picture file is clipped in computer hard disc is set, training parameter is finally set, starts to train.
Step 3 carries out each slice map that test data is concentrated using the complete depth convolution self-encoding encoder of training
Rebuild, and calculate rebuild before with reconstruction after two picture pixels difference difference parameter L1 norm, it is poor to find maximum
Reference of the different parameter L1 norm as Threshold segmentation.Step 3 is rebuild using python language and keras as example, is needed first
Then the mind of preservation after the completion of training is built with specific neural network identical when training and is loaded into the library python for importing dependence
Through network weight, setting needs the picture file rebuild to be clipped in the path on computer hard disc later, finally with the neural network into
Row image reconstruction, the picture once rebuild.
The each slice map that test data is concentrated is rebuild using training complete depth convolution self-encoding encoder, Fig. 3
For the result rebuild to three different slice maps.Calculate original image on reconstruction figure respective pixel luminance difference it is absolute
Value, and calculate whole pixel absolute value of the difference and sum.The namely difference of original image and the difference of the upper respective pixel brightness of reconstruction figure
Different parameter L1 norm, calculation equation such as following formula (1):
... ... ... ... ... (1)
WhereinFor the brightness value of some pixel in original image,To rebuild the brightness value for scheming some upper pixel,NRepresent every slice
Pixel quantity on figure in total.A is original image, and b is to rebuild figure, and i is the serial number of pixel on picture, is started counting from 1.║a-b║1 Table
Show the L1 norm of original image and reconstruction figure.
The above operation is applied and is concentrated on all slices in test data, the maximum of wherein difference parameter L1 norm is found
Value can be used as the threshold reference for distinguishing non-defective unit and defective products.
Step 4 rebuilds the slice use of picture to be detected, calculates two picture pictures after rebuilding preceding and reconstruction
The difference parameter L1 norm of the difference of element；Picture is divided into non-defective unit and defective products according to the value of difference parameter L1 norm, and is obtained
Defects detection result.The picture that product to be detected is acquired using industrial camera rebuilds the slice of picture, calculates before rebuilding
And the difference parameter L1 norm of the difference of two picture pixels after reconstruction；Picture divided into according to the value of difference parameter L1 norm good
Product and defective products, and obtain defects detection result.
Picture to be detected be using during actually detected using industrial camera or other imaging devices shoot for examining
Survey the whether defective picture of product.Default use is schemed identical parameter with processing training and is sliced to picture to be detected.Make
Slice is rebuild with training complete depth convolution self-encoding encoder, the network after training can be by the textile fiber knot with defect
If composition is reduced to, there is no the states that fibre structure when defect is most possibly presented.If it should be noted that by non-defective unit
Slice map input this neural network, the result of output and input are almost the same.Original image is compared by algorithm and is rebuild between figure
Difference degree can be used to whether judge in picture comprising defect.The above operation is applied on all slices, with step 3
In calculated maximum L1 norm be reference threshold, for judging whether current slice includes defect and generate examining report.It will
The absolute value picture of original image and reconstruction figure difference is synthesized, and obtains defects detection effect picture, defects detection effect picture is to detection
Product surface is classified as defect area and is marked, and can be used for manually rechecking.
Step 5: carrying out the diminution of multiple multiplying powers to non-defective unit figure, the non-defective unit figure of each reduction magnification repeat with
On process.The diminution of multiple multiplying powers is carried out to non-defective unit figure, for example saving length-width ratio for non-defective unit figure size reduction is original two
/ mono-, a quarter and 1/8th.Step 1 is repeated to step 4 until having handled all zoom ratios, can be calculated
Whether a certain picture to be detected includes defect under different reduction ratios out；Picture to be detected will also carry out pair in repeating process
Answer the diminution of ratio；It is possible thereby to maintain neural network structure it is constant in the case where to different scale under object surface structure
It is detected.
Technical solution of the present invention is applied and is organized on wire defect detection system in longitude and latitude by the present embodiment.For trained non-defective unit
Image is the grayscale image that 10 resolution ratio for being acquired and being cut by industrial camera is 640*384.The specific mistake of the present embodiment
Journey is as follows:
One, longitude and latitude compiles the preparation of image data set:
Pair warp and weft compiles non-defective unit image and classifies, and 80% non-defective unit picture is used to train, and remaining 20% for testing.To picture
Data set carries out slicing treatment, and the size of slice is 64*64 pixel, and stepping is 32 pixels, then every resolution ratio is 640*384's
Image can be cut into the image of 209 64*64, form training set.Convolution self-encoding encoder training dataset is 1672 longitudes and latitudes
Slice picture is compiled, test data set is 418 slice pictures.
Two, carry out neural metwork training
Using depth convolution self-encoding encoder as shown in Figure 2, the number of iterations 300, batch size (batch size) is 128, choosing
It uses two-value cross entropy as loss function, is carried out using 1672 training with through weft-knitted slice map piece, that is, training dataset
Training.Training saves the weighted data of neural network after completing.
Three, rebuild simultaneously computational discrimination threshold value to test data set
Test data set is rebuild using neural network after training, Fig. 3 is that three different slice maps are carried out with weight
The result built is for reference, it can be seen that the textile fiber structure chart with defect can be reduced to vacation by the network after training
The state that fibre structure is most possibly presented when defect is such as not present.The difference between original image and reconstruction figure is calculated using L1 norm
Off course degree.The above calculating is carried out to whole test data sets, maximizing improves 10% using on the basis of this maximum value as figure
In piece whether include defect discrimination threshold.
Four, it is scaled to original data set, repeat above step
Trained non-defective unit image length and width are used for by 10 while being reduced into original a quarter, repeat above-mentioned process until calculating
It is spellbound through network weight and discrimination threshold.First-stage reduction has been carried out in the present embodiment, and several different must contract can also be set more
Small scale repeats process, obtains the neural network weight and discrimination threshold under different diminution ratios.This step is to improve inspection
Survey the effect of different scale defect.
Five, carry out on-line checking
The network weight obtained using selected depth convolution self-encoding encoder neural network model, training, and to test data set
The discrimination threshold being calculated can carry out detecting through weft knitting online.Using industrial camera shooting longitude and latitude compilation piece, use
Sliding window (sliding window) method is scanned picture to be detected, persistently carries out to the slice picture in window scarce
Fall into detection.Neural network weight and differentiation weight after scaled to picture to be detected under the corresponding diminution ratio of application carry out
Defects detection.The defects detection finally concluded under different diminution ratios generates examining report.
If the product for replacing a kind of new varieties carries out defects detection, manual oversight is needed to resurvey ten multiple nothings big
Area defect image, the defect that small probability generates are mixed into training dataset and will not significantly affect to final detection quality generation, it
Software will be automatically performed re -training according to above-mentioned method afterwards, completes parameter and calculates, and start the defect inspection of new varieties product
It surveys.
Particular embodiments described above, the technical issues of to solution of the invention, technical scheme and beneficial effects carry out
It is further described, it should be understood that the above is only a specific embodiment of the present invention, is not limited to
The present invention, all within the spirits and principles of the present invention, any modification, equivalent substitution, improvement and etc. done should be included in this
Within the protection scope of invention.
Claims (7)
1. a kind of method for detecting surface defects of products based on depth convolution self-encoding encoder, which is characterized in that it includes following step
It is rapid:
Step 1 is not included the non-defective unit picture of defect using imaging device acquisition, non-defective unit picture is divided into two by a certain percentage
Part, for a part for training depth convolution self-encoding encoder, another part is used for test depth convolution self-encoding encoder training effect
And the defect for calculating reference distinguishes threshold value；All pictures are sliced, training dataset and test data set are obtained；
Step 2 is trained using depth convolution self-encoding encoder network model and training dataset；
Step 3 carries out weight to each slice map that test data is concentrated using the complete depth convolution self-encoding encoder of training
Build, and calculate rebuild before with reconstruction after two picture pixels difference difference parameter；
Step 4 rebuilds the slice use of picture to be detected, calculate reconstruction before with rebuild after two picture pixels it
The difference parameter of difference；Picture is divided into non-defective unit and defective products according to the value of difference parameter, and obtains defects detection result；
Step 5: carrying out the diminution of multiple multiplying powers to non-defective unit figure, and the non-defective unit figure of each reduction magnification repeats above
Process；Step 1 is repeated to step 4 until having handled all zoom ratios, a certain can be calculated under different reduction ratios
Whether picture to be detected includes defect；Picture to be detected will also carry out the diminution of corresponding ratio in repeating process；Thus it is tieing up
Hold neural network structure it is constant in the case where the object surface structure under different scale is detected.
2. the method for detecting surface defects of products as described in claim 1 based on depth convolution self-encoding encoder, which is characterized in that
The step 2 is trained using depth convolution self-encoding encoder neural network and training dataset.
3. the method for detecting surface defects of products as described in claim 1 based on depth convolution self-encoding encoder, which is characterized in that
The picture to be detected is to be produced using what industrial camera or other imaging devices were shot for detecting using during actually detected
The whether defective picture of product.
4. the method for detecting surface defects of products as described in claim 1 based on depth convolution self-encoding encoder, which is characterized in that
The imaging device industrial camera.
5. the method for detecting surface defects of products as described in claim 1 based on depth convolution self-encoding encoder, which is characterized in that
The step 5 is scanned picture to be detected using sliding window method, persistently carries out defect to the slice picture in window
Detection.
6. the method for detecting surface defects of products as described in claim 1 based on depth convolution self-encoding encoder, which is characterized in that
The step 3 finds reference of the maximum difference parameter as Threshold segmentation.
7. the method for detecting surface defects of products as described in claim 1 based on depth convolution self-encoding encoder, which is characterized in that
The depth convolution self-encoding encoder, which uses, is divided into coded portion and the two-part network architecture of decoding.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910118229.8A CN109872313A (en) | 2019-02-15 | 2019-02-15 | A kind of method for detecting surface defects of products based on depth convolution self-encoding encoder |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910118229.8A CN109872313A (en) | 2019-02-15 | 2019-02-15 | A kind of method for detecting surface defects of products based on depth convolution self-encoding encoder |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109872313A true CN109872313A (en) | 2019-06-11 |
Family
ID=66918865
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910118229.8A Pending CN109872313A (en) | 2019-02-15 | 2019-02-15 | A kind of method for detecting surface defects of products based on depth convolution self-encoding encoder |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109872313A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110570396A (en) * | 2019-08-07 | 2019-12-13 | 华中科技大学 | industrial product defect detection method based on deep learning |
| CN111047550A (en) * | 2019-11-04 | 2020-04-21 | 睿视智觉（厦门）科技有限公司 | Product abnormity detection method, device and equipment based on machine vision |
| CN111062918A (en) * | 2019-12-10 | 2020-04-24 | 歌尔股份有限公司 | Anomaly detection method and device based on computer vision |
| CN112288741A (en) * | 2020-11-23 | 2021-01-29 | 四川长虹电器股份有限公司 | Product surface defect detection method and system based on semantic segmentation |
| CN113129272A (en) * | 2021-03-30 | 2021-07-16 | 广东省科学院智能制造研究所 | Defect detection method and device based on denoising convolution self-encoder |
| CN113256602A (en) * | 2021-06-10 | 2021-08-13 | 中科云尚(南京)智能技术有限公司 | Unsupervised fan blade defect detection method and system based on self-encoder |
| CN113822836A (en) * | 2020-06-05 | 2021-12-21 | 英业达科技有限公司 | Method of marking an image |
| CN114022409A (en) * | 2021-09-30 | 2022-02-08 | 电子科技大学 | Coated medicine surface defect detection algorithm based on deep learning |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107578003A (en) * | 2017-08-29 | 2018-01-12 | 中国科学院遥感与数字地球研究所 | A kind of remote sensing images transfer learning method based on GEOGRAPHICAL INDICATION image |
| CN108562589A (en) * | 2018-03-30 | 2018-09-21 | 慧泉智能科技（苏州）有限公司 | A method of magnetic circuit material surface defect is detected |
| CN108961217A (en) * | 2018-06-08 | 2018-12-07 | 南京大学 | A kind of detection method of surface flaw based on positive example training |
-
2019
- 2019-02-15 CN CN201910118229.8A patent/CN109872313A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107578003A (en) * | 2017-08-29 | 2018-01-12 | 中国科学院遥感与数字地球研究所 | A kind of remote sensing images transfer learning method based on GEOGRAPHICAL INDICATION image |
| CN108562589A (en) * | 2018-03-30 | 2018-09-21 | 慧泉智能科技（苏州）有限公司 | A method of magnetic circuit material surface defect is detected |
| CN108961217A (en) * | 2018-06-08 | 2018-12-07 | 南京大学 | A kind of detection method of surface flaw based on positive example training |
Non-Patent Citations (1)
| Title |
|---|
| SHUANG MEI 等: "Automatic Fabric Defect Detection with a Multi-Scale Convolutional DenoisingAutoencoder Network Model", 《SENSORS》 * |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110570396A (en) * | 2019-08-07 | 2019-12-13 | 华中科技大学 | industrial product defect detection method based on deep learning |
| CN110570396B (en) * | 2019-08-07 | 2022-03-15 | 华中科技大学 | Industrial product defect detection method based on deep learning |
| CN111047550A (en) * | 2019-11-04 | 2020-04-21 | 睿视智觉（厦门）科技有限公司 | Product abnormity detection method, device and equipment based on machine vision |
| CN111062918A (en) * | 2019-12-10 | 2020-04-24 | 歌尔股份有限公司 | Anomaly detection method and device based on computer vision |
| CN111062918B (en) * | 2019-12-10 | 2023-11-21 | 歌尔股份有限公司 | Abnormality detection method and device based on computer vision |
| US12002195B2 (en) | 2019-12-10 | 2024-06-04 | Goertek Inc. | Computer vision-based anomaly detection method, device and electronic apparatus |
| CN113822836A (en) * | 2020-06-05 | 2021-12-21 | 英业达科技有限公司 | Method of marking an image |
| CN112288741A (en) * | 2020-11-23 | 2021-01-29 | 四川长虹电器股份有限公司 | Product surface defect detection method and system based on semantic segmentation |
| CN113129272A (en) * | 2021-03-30 | 2021-07-16 | 广东省科学院智能制造研究所 | Defect detection method and device based on denoising convolution self-encoder |
| CN113256602A (en) * | 2021-06-10 | 2021-08-13 | 中科云尚(南京)智能技术有限公司 | Unsupervised fan blade defect detection method and system based on self-encoder |
| CN114022409A (en) * | 2021-09-30 | 2022-02-08 | 电子科技大学 | Coated medicine surface defect detection algorithm based on deep learning |
| CN114022409B (en) * | 2021-09-30 | 2023-04-18 | 电子科技大学 | Coated medicine surface defect detection algorithm based on deep learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109872313A (en) | A kind of method for detecting surface defects of products based on depth convolution self-encoding encoder | |
| CN108961217B (en) | Surface defect detection method based on regular training | |
| Ying et al. | From patches to pictures (PaQ-2-PiQ): Mapping the perceptual space of picture quality | |
| CN107123111B (en) | Deep residual error network construction method for mobile phone screen defect detection | |
| CN111402226A (en) | Surface defect detection method based on cascade convolution neural network | |
| CN110473173A (en) | A kind of defect inspection method based on deep learning semantic segmentation | |
| CN108562589A (en) | A method of magnetic circuit material surface defect is detected | |
| CN111383209A (en) | Unsupervised flaw detection method based on full convolution self-encoder network | |
| CN106875373A (en) | Mobile phone screen MURA defect inspection methods based on convolutional neural networks pruning algorithms | |
| CN108573222A (en) | The pedestrian image occlusion detection method for generating network is fought based on cycle | |
| CN111798409A (en) | Deep learning-based PCB defect data generation method | |
| CN110889838A (en) | Fabric defect detection method and device | |
| CN113538375A (en) | PCB defect detection method based on YOLOv5 | |
| CN113129272A (en) | Defect detection method and device based on denoising convolution self-encoder | |
| CN113160123A (en) | Leather defect detection method, system and device based on YOLOv5 | |
| US20190272627A1 (en) | Automatically generating image datasets for use in image recognition and detection | |
| CN115861323B (en) | Leather defect detection method based on refined segmentation network | |
| CN111339902A (en) | Liquid crystal display number identification method and device of digital display instrument | |
| CN111242911A (en) | Method and system for determining image definition based on deep learning algorithm | |
| CN109410158B (en) | Multi-focus image fusion method based on convolutional neural network | |
| CN109840905A (en) | Power equipment rusty stain detection method and system | |
| CN114119500A (en) | Yarn dyed fabric defect area detection method based on generation countermeasure network | |
| CN116778269A (en) | Method for constructing product surface defect detection model based on self-encoder reconstruction | |
| CN115861614A (en) | Method and device for automatically generating semantic segmentation graph based on down jacket image | |
| CN110111332A (en) | Collagent casing for sausages defects detection model, detection method and system based on depth convolutional neural networks |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190611 |