CN110415238A - Diaphragm spots detection method based on reversed bottleneck structure depth convolutional network - Google Patents
Diaphragm spots detection method based on reversed bottleneck structure depth convolutional network Download PDFInfo
- Publication number
- CN110415238A CN110415238A CN201910700273.XA CN201910700273A CN110415238A CN 110415238 A CN110415238 A CN 110415238A CN 201910700273 A CN201910700273 A CN 201910700273A CN 110415238 A CN110415238 A CN 110415238A
- Authority
- CN
- China
- Prior art keywords
- image
- depth convolutional
- convolutional network
- bottleneck structure
- reversed
- 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.)
- Withdrawn
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Quality & Reliability (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of diaphragm spots detection methods based on reversed bottleneck structure depth convolutional network to realize the detection and calibration to flaw point in diaphragm by using reversed bottleneck structure depth convolutional network is based on.This method includes the parts such as Image Acquisition, image segmentation, data mark, network training, spots detection and image mosaic.Method provided by the invention takes full advantage of depth convolutional network for the validity of image characteristics extraction, and reversed bottleneck structure can greatly reduce the quantity of parameter in the case where detection accuracy is guaranteed, to realize the purpose for rapidly and accurately detecting flaw point in diaphragm.
Description
Technical field
The present invention relates to the fields such as deep learning and computer vision, specifically design a kind of deep based on reversed bottleneck structure
Spend the diaphragm spots detection method of convolutional network.
Background technique
With the fast development of electronics technology sector, various portable devices such as laptop, tablet computer, mobile phone etc.
It is widely used in daily life, human-computer interaction interface window --- display screen is then particularly important.Liquid crystal display is because aobvious
Show quality it is high, without electromagnetic radiation, effective area is wide, low in energy consumption the advantages that, be almost used for all portable devices.Liquid crystal display
On optical diaphragm both play a part of protecting liquid crystal display or affect the clarity of display.In the production process of optical diaphragm
In, dust, scratch, print unevenness are even etc. can all cause the flaw of optical diaphragm, these flaws directly affect the display of screen
Effect, so, the flaw defect detection for diaphragm be it is highly important, it is directly related to the final performance and quality of product.
And these flaw points are difficult to identify that for human eye in process of production, the biography that most domestic enterprise uses at present
System method such as statistic law, Spectrum Method etc. is not all high there is accuracy of identification and detects the problems such as speed is slow, is unable to satisfy work
The requirement of industry production accuracy and real-time.
Currently, depth convolutional network is applied to computer vision field, and with apparent advantage increasingly by all circles
Researchers' has deep love for, and the short time emerges one after another for all kinds of algorithms and network model of image classification, and with quickish
Speed improvement classification performance, for traditional algorithm, accuracy of identification and detection speed have and are substantially improved.
Summary of the invention
Goal of the invention: in order to solve the problems, such as that traditional flaw detection method accuracy of identification is low, slow-footed, the present invention provides
A kind of diaphragm spots detection method based on reversed bottleneck structure depth convolutional network, by using deep based on reversed bottleneck structure
Convolutional network is spent, the detection and calibration to flaw point in diaphragm are realized.
Technical solution: to achieve the above object, the technical solution adopted by the present invention are as follows:
Diaphragm spots detection method based on reversed bottleneck structure depth convolutional network, by trained network model,
Flaw point in diaphragm is detected.For diaphragm defective, first progress Image Acquisition and data mark, picture is put
Enter in network and be trained, the image for the diaphragm for needing to detect is sent into network after the completion of training, network determines whether exist
Flaw.Specifically comprise the following steps:
(1) membrane image defective is collected, data set is made as；
(2) collected flaw membrane image is divided into a series of small images；
(3) small image is judged and is marked: by the small image labeling containing flaw point in region as NG, nothing in region
The small image labeling of flaw point is OK；
(4) according to flaw data obtained, the parameter of regulating networks carries out reversed bottleneck structure depth convolutional network
Training；
(5) image detected will be needed successively to be cut to a series of small images, and numbered；
(6) the small image of step (5) is inputted into network, loads the weight parameter that training is completed, obtains to every small picture
Judging result (NG, OK), and record result be NG small image number；
It (7) is original input picture size by small image mosaic according to the number of small images all in step (6), according to knot
Fruit is that the number of the small image of NG goes out the region collimation mark where the small image, which is flaw point position.
In above-mentioned steps, step (1)~(4) are data prediction, network training step, and step (5)~(7) are flaw inspection
Survey step.
The reversed bottleneck structure depth convolutional network, reversed bottleneck structure depth convolutional network is mainly by 16 reversed bottles
Neck convolution module, 2 convolutional layers, 1 pond layer and 1 full articulamentum composition.Wherein each reversed bottleneck convolution module is by 2
Convolutional layer (convolution kernel is 1 × 1) and 1 depth separate convolutional layer (convolution kernel is 3 × 3 or 5 × 5) composition.
Judgement is carried out to small image in the step (3) to include the following steps: with mark
(a1) the file storage for creating entitled NG has the small image of flaw point, and by being overturn, being translated to image
And it adjusts contrast and realizes data enhancing, EDS extended data set；
(a2) file for creating entitled OK stores the small image of flawless fault, and by being overturn, being translated to image
And it adjusts contrast and realizes data enhancing, EDS extended data set；
(a3) image in NG file is put into training set file train respectively according to the ratio of 8:2 and verifying collects
NG file under file val；
(a4) image in OK file is put into training set file train respectively according to the ratio of 8:2 and verifying collects
OK file under file val.
The specific steps that the step (4) is trained reversed bottleneck structure depth convolutional network are as follows:
(b1) pre-training is carried out to reversed bottleneck structure depth convolutional network on ImageNet data set；
(b2) freeze the feature extraction layer of reversed bottleneck structure depth convolutional network, modification classification layer parameter；
(b3) network training is carried out to training set, is collected after the completion of training using verifying and carry out accuracy evaluation, until penalty values
No longer decline, precision is no longer promoted；
(b4) defrosting feature extraction layer continues to train, and is collected after the completion of training using verifying and carries out accuracy evaluation, until loss
Value no longer declines, and precision is no longer promoted；Otherwise regulating networks parameter continues to train.
Regulating networks parameter in the step (4), refers specifically to:
(c1) size of learning rate LR, momentum Momentum, weight attenuation rate WeightDecay are adjusted；
(c2) size of every batch of training samples number BatchSize is adjusted；
(c3) size of the number of iterations Epoch of entire data set is adjusted.
The utility model has the advantages that
Effect 1: by data enhancement operations, solving the problems, such as that amount of training data is few, reduces needed for data set production
Cost.
Effect 2: reversed bottleneck structure convolution module is used, parameter amount is greatly reduced, improves precision, accelerate network
Training speed and test speed, and improve the accuracy rate of detection.
Effect 3: carrying out pre-training using ImageNet, solve training samples number it is few when feature extraction is difficult asks
Topic.Pass through network 1000 type objects on ImageNet through row feature extraction and classification before training diaphragm flaw defect detection
The pre-training weight on ImageNet is loaded, present networks have powerful ability in feature extraction.
Detailed description of the invention
Fig. 1 is the step flow diagram of embodiment；
Fig. 2 is the overall structure figure of reversed bottleneck structure depth convolutional network；
Fig. 3 is reversed bottleneck structure convolution module structure chart；
Fig. 4 is the effect picture of detection diaphragm spots of the invention.
Specific embodiment
The present invention will be further explained with reference to the accompanying drawing.
Diaphragm spots detection method based on reversed bottleneck structure depth convolutional network, by trained network model,
Flaw point in diaphragm is detected.For diaphragm defective, first progress Image Acquisition and data mark, picture is put
Enter in network and be trained, the image for the diaphragm for needing to detect is sent into network after the completion of training, network determines whether exist
Flaw.Specifically comprise the following steps:
(1) membrane image defective is collected, data set is made as；
(2) collected flaw membrane image is divided into a series of small images；
(3) small image is judged and is marked: by the small image labeling containing flaw point in region as NG, nothing in region
The small image labeling of flaw point is OK；
(4) according to flaw data obtained, the parameter of regulating networks carries out reversed bottleneck structure depth convolutional network
Training；
(5) image detected will be needed successively to be cut to a series of small images, and numbered；
(6) the small image of step (5) is inputted into network, loads the weight parameter that training is completed, obtains to every small picture
Judging result (NG, OK), and record result be NG small image number；
It (7) is original input picture size by small image mosaic according to the number of small images all in step (6), according to knot
Fruit is that the number of the small image of NG goes out the region collimation mark where the small image, which is flaw point position.
In above-mentioned steps, step (1)~(4) are data prediction, network training step, and step (5)~(7) are flaw inspection
Survey step.
As shown in Fig. 2, reversed bottleneck structure depth convolutional network mainly by 16 reversed bottleneck convolution modules (MBConv),
2 convolutional layers (Conv), 1 pond layer (Pooling) and 1 full articulamentum (FC) composition.
As shown in figure 3, each reversed bottleneck convolution module can be divided by 2 convolutional layers (convolution kernel is 1 × 1) and 1 depth
From convolutional layer (DWConv, convolution kernel are 3 × 3 or 5 × 5) composition.The identical point of reversed bottleneck structure and traditional bottleneck structure exists
In: model compression can be carried out by 1 × 1 convolution kernel, parameter amount is greatly decreased in the case where losing few precision,
Network operations speed is promoted, computing resource is saved, is also beneficial to be deployed on general mobile platform.Reversed bottleneck structure and biography
The difference of system bottleneck structure is: traditional bottleneck structure first uses 1 × 1 convolution to reduce the dimension of the characteristic pattern of input, then
The convolution operation for carrying out 3 × 3, is finally again become larger dimension with 1 × 1 convolution.And the reversed bottleneck structure that the present invention uses is then
It is that first the characteristic pattern dimension of input is become larger with 1 × 1 convolution, then separates convolution with 3 × 3 depth and do convolution algorithm, finally
Its dimension is reduced using 1 × 1 convolution algorithm, after 1 × 1 convolution algorithm at this time, does not use ReLU6 activation primitive, and
It is, to retain more features information, to guarantee the ability to express of model using linear activation primitive.
Below with reference to example, the present invention will be further explained.
Embodiment:
As shown in Figure 1, a kind of diaphragm spots detection method based on reversed bottleneck structure depth convolutional network.
Step 1: preparing data set
The image with diaphragm defective is acquired from workshop, data set is made, which includes number Zhang Liyong work
Industry grade camera carries out the picture of shooting acquisition.
Step 2: collected flaw membrane image is divided into a series of small images.
Since collected initial pictures are very big, and flaw therein point is very small, if being directly input to big image
Be trained in network, not only increased the difficulty of training and detection, but also caused computing resource waste, and be difficult to data set into
Rower note.Therefore, it carries out Pixel-level to original image to cut, every original image can cut into 800 224 × 224 pixels
Small image.
Step 3: small image is judged and is marked: being NG, region by the small image labeling containing flaw point in region
In flawless fault small image labeling be OK.Due to there is over-fitting when data set lazy weight will lead to trained, therefore walk herein
Increasing data in rapid enhances step, increases data set by way of translation, overturning, adjusting contrast.
(3.1) the file storage for creating entitled NG has the small image of flaw point, and by being overturn, being translated to image
And it adjusts contrast and realizes data enhancing, EDS extended data set；
(3.2) file for creating entitled OK stores the small image of flawless fault, and by being overturn, being translated to image
And it adjusts contrast and realizes data enhancing, EDS extended data set；
(3.3) image in NG file is put into training set file train respectively according to the ratio of 8:2 and verifying collects
NG file under file val；
(3.4) image in OK file is put into training set file train respectively according to the ratio of 8:2 and verifying collects
OK file under file val.
Step 4: according to flaw data obtained, the parameter of regulating networks, to reversed bottleneck structure depth convolutional network
It is trained.
(4.1) pre-training is carried out to reversed bottleneck structure depth convolutional network on ImageNet data set；
(4.2) freeze the feature extraction layer of reversed bottleneck structure depth convolutional network, modification classification layer parameter, due to Home Network
The result that network needs to export is NG or OK, is two classification problems, therefore full articulamentum dimension is revised as (1280,2)；
(4.3) weight parameter that pre-training is carried out on ImageNet is loaded into network, setting optimization algorithm is tool
There is the stochastic gradient descent method of momentum, enable learning rate LR=0.01, momentum Momentum=0.9, BatchSize=256,
Epoch=30 carries out network training to training set；
(4.4) next thaw feature extraction layer, and setting optimization algorithm is the boarding steps decayed with momentum and learning rate
Descent method is spent, learning rate LR=0.001, momentum Momentum=0.9, weight attenuation rate WeightDecay=0.0005 are set,
BatchSize=64, Epoch=50 continue to train, and are collected after the completion of training using verifying and carry out accuracy evaluation, until penalty values
No longer decline, precision is no longer promoted；Otherwise regulating networks parameter continues to train.
Regulating networks parameter in the step (4.4), refers specifically to:
(4.4.1) adjusts the size of learning rate LR, momentum Momentum, weight attenuation rate WeightDecay；
(4.4.2) adjusts the size that every batch of is sent into the training samples number BatchSize of network；
(4.4.3) adjusts the size of the number of iterations Epoch of entire data set；
Wherein, learning rate LR determines the speed of right value update, and momentum Momentum increases gradient fall, from
And penalty values can speed up convergence, the purpose of using weights attenuation rate WeightDecay is to prevent over-fitting.In loss function
In, WeightDecay is placed on a coefficient before regular terms (regularization), and regular terms generally indicates model
Complexity, so the effect of Weight Decay is the influence for adjusting model complexity to loss function, if Weight Decay
Very big, then the value of complicated model loss function is also just big.
Due to the complicated network structure of reversed bottleneck structure depth convolutional network, and the more difficult extraction of training set feature, therefore make
The method of the transfer learning described in (4.1)~(4.4) is trained reversed bottleneck structure depth convolutional network.So not
The speed of trained network is only improved, but also network can preferably extract the feature of flaw point, improves the precision of network
And test accuracy rate.
Loss function used in the present invention is cross entropy loss function, and cross entropy describes between two probability distribution
Distance, the smaller explanation of cross entropy is closer between the two, to intersect entropy loss decline by stochastic gradient descent until restraining,
So that the prediction result of network becomes closer to label value.
Cross entropy loss function C (p, q), q (x) are label value:
C (p, q)=Ep[- ligq]=- ∑ p (x) logq (x)=H (p)+DKL(p||q) (1)
Wherein,
H (p)=- ∑ p (x) logp (x) (2)
In formula (1), C (p, q) is to intersect entropy loss, EpFor expectation, p (x) is the prediction result of network, and q (x) is label
Value, H (p) are the comentropy of p (x), DKL(p | | q) it is K-L divergence.
Step 5: by a series of small image according to the BatchSize of setting size be sent into training complete network into
Row detection.
(5.1) diaphragm detected to needs is taken pictures, and is uploaded in computer；
(5.2) original image of diaphragm is cut into the small image of 800 224 × 224 pixels, and is successively numbered in order
Record；
(5.3) 800 small images are sent into the network of training completion in batches according to the BatchSize size of setting；
(5.4) image is after 16 reversed bottleneck convolution modules, 2 convolutional layers, 1 pond layer and 1 full articulamentum
It exports and saves testing result (NG or OK), until 800 small image all complete by detection.
Step 6: will test number splicing of the small image of completion according to segmentation when becomes original image, will test
The small image of flaw is gone out in the figure of original size with red collimation mark.
(6.1) 800 small image is all after the completion of detection, by small image mosaic is original according to the number of all small images
Region where the small image is gone out (size 224 with collimation mark according to the number for the small image that result is NG by input picture size
× 224), which is flaw point position.
(6.2) the effect picture after checking test, assesses the performance of network.As shown in figure 4, the region gone out in figure with collimation mark
It is as detected diaphragm region defective, the size of frame is the size (224 × 224) of the every small image cut, due to original
Beginning picture size is too big, this figure is only a part of testing result figure, is based on reversed bottleneck structure depth convolution net to show
The validity of the diaphragm spots detection method of network.
The above is only a preferred embodiment of the present invention, it should be pointed out that: for the ordinary skill people of the art
For member, various improvements and modifications may be made without departing from the principle of the present invention, these improvements and modifications are also answered
It is considered as protection scope of the present invention.
Claims (5)
1. a kind of diaphragm spots detection method based on reversed bottleneck structure depth convolutional network, it is characterised in that including walking as follows
It is rapid:
(1) membrane image defective is collected, data set is made as；
(2) collected flaw membrane image is divided into a series of small images；
(3) small image is judged and is marked: is indefectible in region by the small image labeling containing flaw point in region as NG
The small image labeling of point is OK；
(4) according to flaw data obtained, the parameter of regulating networks instructs reversed bottleneck structure depth convolutional network
Practice；
(5) image detected will be needed successively to be cut to a series of small images, and numbered；
(6) the small image of step (5) is inputted into network, loads the weight parameter that training is completed, obtains sentencing every small picture
Disconnected result (NG, OK), and record the number for the small image that result is NG；
(7) it is original input picture size by small image mosaic according to the number of small images all in step (6), is according to result
The number of the small image of NG goes out the region collimation mark where the small image, which is flaw point position.
2. the diaphragm spots detection method based on reversed bottleneck structure depth convolutional network according to claim 1,
It is characterized in that: judgement being carried out to small image in the step (3) and is included the following steps: with mark
(a1) the file storage for creating entitled NG has the small image of flaw point, and by being overturn, being translated and being adjusted to image
It saves contrast and realizes data enhancing, EDS extended data set；
(a2) file for creating entitled OK stores the small image of flawless fault, and by being overturn, being translated and being adjusted to image
It saves contrast and realizes data enhancing, EDS extended data set；
(a3) image in NG file is put into training set file train and verifying collection file respectively according to the ratio of 8:2
Press from both sides the NG file under val；
(a4) image in OK file is put into training set file train and verifying collection file respectively according to the ratio of 8:2
Press from both sides the OK file under val.
3. the diaphragm spots detection method based on reversed bottleneck structure depth convolutional network according to claim 1,
It is characterized in that: the specific steps that the step (4) is trained reversed bottleneck structure depth convolutional network are as follows:
(b1) pre-training is carried out to reversed bottleneck structure depth convolutional network on ImageNet data set；
(b2) freeze the feature extraction layer of reversed bottleneck structure depth convolutional network, modification classification layer parameter；
(b3) to training set carry out network training, training after the completion of using verifying collection progress accuracy evaluation, until penalty values no longer
Decline, precision are no longer promoted；
(b4) defrosting feature extraction layer continues to train, training after the completion of using verifying collection progress accuracy evaluation, until penalty values not
Decline again, precision is no longer promoted；Otherwise regulating networks parameter continues to train.
4. based on the diaphragm spots detection method of reversed bottleneck structure depth convolutional network according to claim 3,
Be characterized in that: regulating networks parameter in step (b4) refers specifically to:
(c1) size of learning rate LR, momentum Momentum, weight attenuation rate WeightDecay are adjusted；
(c2) size of every batch of training samples number BatchSize is adjusted；
(c3) size of the number of iterations Epoch of entire data set is adjusted.
5. the diaphragm spots detection method based on reversed bottleneck structure depth convolutional network according to claim 1, special
Sign is: the reversed bottleneck structure depth convolutional network includes 16 reversed bottleneck convolution modules, 2 convolutional layers, 1 pond layer
With 1 full articulamentum；Wherein each reversed bottleneck convolution module separates convolutional layer by 2 convolutional layers and 1 depth and forms.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910700273.XA CN110415238A (en) | 2019-07-31 | 2019-07-31 | Diaphragm spots detection method based on reversed bottleneck structure depth convolutional network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910700273.XA CN110415238A (en) | 2019-07-31 | 2019-07-31 | Diaphragm spots detection method based on reversed bottleneck structure depth convolutional network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110415238A true CN110415238A (en) | 2019-11-05 |
Family
ID=68364659
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910700273.XA Withdrawn CN110415238A (en) | 2019-07-31 | 2019-07-31 | Diaphragm spots detection method based on reversed bottleneck structure depth convolutional network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110415238A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111104897A (en) * | 2019-12-18 | 2020-05-05 | 深圳市捷顺科技实业股份有限公司 | Training method and device for child face recognition model and storage medium |
| CN111415338A (en) * | 2020-03-16 | 2020-07-14 | 城云科技（中国）有限公司 | Method and system for constructing target detection model |
| CN112101463A (en) * | 2020-09-17 | 2020-12-18 | 成都数之联科技有限公司 | Image semantic segmentation network training method, segmentation device and medium |
| CN112233090A (en) * | 2020-10-15 | 2021-01-15 | 浙江工商大学 | Film flaw detection method based on improved attention mechanism |
| WO2021099938A1 (en) * | 2019-11-22 | 2021-05-27 | International Business Machines Corporation | Generating training data for object detection |
-
2019
- 2019-07-31 CN CN201910700273.XA patent/CN110415238A/en not_active Withdrawn
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2021099938A1 (en) * | 2019-11-22 | 2021-05-27 | International Business Machines Corporation | Generating training data for object detection |
| US11200455B2 (en) | 2019-11-22 | 2021-12-14 | International Business Machines Corporation | Generating training data for object detection |
| GB2606091A (en) * | 2019-11-22 | 2022-10-26 | Ibm | Generating training data for object detection |
| CN111104897A (en) * | 2019-12-18 | 2020-05-05 | 深圳市捷顺科技实业股份有限公司 | Training method and device for child face recognition model and storage medium |
| CN111415338A (en) * | 2020-03-16 | 2020-07-14 | 城云科技（中国）有限公司 | Method and system for constructing target detection model |
| CN112101463A (en) * | 2020-09-17 | 2020-12-18 | 成都数之联科技有限公司 | Image semantic segmentation network training method, segmentation device and medium |
| CN112233090A (en) * | 2020-10-15 | 2021-01-15 | 浙江工商大学 | Film flaw detection method based on improved attention mechanism |
| CN112233090B (en) * | 2020-10-15 | 2023-05-30 | 浙江工商大学 | Film flaw detection method based on improved attention mechanism |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110415238A (en) | Diaphragm spots detection method based on reversed bottleneck structure depth convolutional network | |
| CN110060237B (en) | Fault detection method, device, equipment and system | |
| CN106875373B (en) | Mobile phone screen MURA defect detection method based on convolutional neural network pruning algorithm | |
| CN108686978B (en) | ARM-based fruit category and color sorting method and system | |
| CN111739075A (en) | Deep network lung texture recognition method combining multi-scale attention | |
| CN107203778A (en) | PVR intensity grade detecting system and method | |
| CN109509172A (en) | A kind of liquid crystal display flaw detection method and system based on deep learning | |
| Bo et al. | Particle pollution estimation from images using convolutional neural network and weather features | |
| CN107966454A (en) | A kind of end plug defect detecting device and detection method based on FPGA | |
| CN110782420A (en) | Small target feature representation enhancement method based on deep learning | |
| CN105608446A (en) | Video stream abnormal event detection method and apparatus | |
| CN107220643A (en) | The Traffic Sign Recognition System of deep learning model based on neurological network | |
| CN110826520A (en) | Port grab bucket detection method based on improved YOLOv3-tiny algorithm | |
| CN106355367A (en) | Warehouse monitoring management device | |
| CN104025118A (en) | Object detection using extended surf features | |
| CN109919179A (en) | Aneurysms automatic testing method, device and computer readable storage medium | |
| CN111047565A (en) | Method, storage medium and equipment for forest cloud image segmentation | |
| Yao et al. | AYOLOv3-Tiny: An improved convolutional neural network architecture for real-time defect detection of PAD light guide plates | |
| CN106874913A (en) | A kind of vegetable detection method | |
| CN116385430A (en) | Machine vision flaw detection method, device, medium and equipment | |
| CN115063725B (en) | Aircraft skin defect identification system based on multi-scale self-adaptive SSD algorithm | |
| CN113780423A (en) | Single-stage target detection neural network based on multi-scale fusion and industrial product surface defect detection model | |
| CN109543617A (en) | The detection method of intelligent vehicle movement traffic information based on YOLO target detection technique | |
| CN109816048A (en) | A kind of image composition method based on attribute migration | |
| CN116071315A (en) | Product visual defect detection method and system based on machine vision |
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 | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20191105 |
|
| WW01 | Invention patent application withdrawn after publication |