CN109142374A - Method and system based on the efficient Checking model of extra small sample training - Google Patents
Method and system based on the efficient Checking model of extra small sample training Download PDFInfo
- Publication number
- CN109142374A CN109142374A CN201810929603.8A CN201810929603A CN109142374A CN 109142374 A CN109142374 A CN 109142374A CN 201810929603 A CN201810929603 A CN 201810929603A CN 109142374 A CN109142374 A CN 109142374A
- Authority
- CN
- China
- Prior art keywords
- training
- sample
- test
- flaw
- model
- 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
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/84—Systems specially adapted for particular applications
- G01N21/88—Investigating the presence of flaws or contamination
- G01N21/8851—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges
-
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- 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
-
- 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
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/84—Systems specially adapted for particular applications
- G01N21/88—Investigating the presence of flaws or contamination
- G01N21/8851—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges
- G01N2021/8854—Grading and classifying of flaws
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Evolutionary Biology (AREA)
- Biomedical Technology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Quality & Reliability (AREA)
- Signal Processing (AREA)
- Chemical & Material Sciences (AREA)
- Analytical Chemistry (AREA)
- Biochemistry (AREA)
- Immunology (AREA)
- Pathology (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
- Investigating Materials By The Use Of Optical Means Adapted For Particular Applications (AREA)
Abstract
The present invention provides a kind of methods based on the efficient Checking model of extra small sample training, comprising: determines acquisition type, acquires small sample, small sample simulation is expanded;Random sampling upsets sequence;Squeezenet Pro trains Checking model;Test model result;Model is completed in publication.The present invention guarantees the deep learning model for achieving the effect that the training of hundreds of thousands sample number in the case where sample number is considerably less, to preferably be applied in Industrial Inferential Measurements.
Description
Technical field
The invention belongs to quality testing field, especially relate to a kind of based on the efficient Checking model of extra small sample training
Method and system.
Background technique
Traditional vision algorithm is in the case where detecting slight flaws (1 the percent of flaw size about total size), to light
The requirement of source camera is very harsh, and is directed to surface irregularity, and the product of color-variable needs to write a large amount of recognizer strategy
To meet detection demand, this greatly reduces the development efficiency of quality detection project.
Since internet penetrates into the tempo increase of social every field, the reform and innovation of traditional industries are imperative,
New technology is constantly used, make full use of internet technology, such as big data analysis, and intelligent algorithm is to reach quick, quasi-
Really, easily product quality detects, and provides huge Data Detection analysis of data, accurately finds the multiclass of multi-product quality
Type defect, the detection level of continuous upgrading products quality, upgrading synergy are the trend developed at present.
But in the industrial production, because sample number of faulty materials itself is less, and prepare a large amount of faulty materials and normal specimens
Need to expend a large amount of manpowers, therefore offer large sample is almost impossible for big data analysis, the training of intelligent algorithm, together
This condition of sample also limits the popularization of big data analysis, intelligent algorithm in industrial application.
Summary of the invention
In view of this, the present invention provides a kind of method and system based on the efficient Checking model of extra small sample training, guarantee
In the case where sample number considerably less (more than ten of sample), achieve the effect that the deep learning model of hundreds of thousands sample number training,
To preferably be applied in Industrial Inferential Measurements.
In order to achieve the above objectives, the technical scheme of the present invention is realized as follows:
A method of based on the efficient Checking model of extra small sample training, comprising:
Step 1: determining that acquisition type is accounted for including the faulty materials type and normal kind class to be detected according to flaw size
The percentage of gross sample size is divided into N number of grade, acquires small sample;
Expand Step 2: the small sample of various types of, each grade is simulated;
Step 3: random sampling composition training set, verifying set, test set;Sequence is upset to each set;
Step 4: Squeezenet Pro trains Checking model;
Step 5: test model result;It is jumped in next step according to precision and testing time control;
Step 6: model is completed in publication, the test result of each flaw rank obtained according to step 5 provides every level-one
The accuracy of identification of flaw, and the weight for freezing every level-one Squeezenet Pro is saved, issue corresponding Squeezenet
Pro model.
Further, the particular content of step 1 includes:
1.1, the determination faulty materials type to be detected is K-1 times, >=2 K, and total classification is faulty materials species number+normal product
Type=K class, every class sample collect Mk, k=1,2 ..., K small sample, it collects in totalA sample, wherein Mk
∈ [5,10];
1.2, every class faulty materials are divided into N number of grade according to the percentage that flaw size accounts for gross sample size, wherein
1.3, every class sample is divided into two groups:WithWhereinForIn take out at random
80% sample taken is as training set;ForIn 20% sample randomly selected as test set.
Further, the particular content of step 2 includes:
2.1. manual analog sample;
By the kth class faulty materials of n-th of grade (n is bigger, and identification difficulty is higher) and last a kind of normal productK=1,
2 ..., K extends to the α of raw sample number by made mode1=5 times, i.e.,Faulty materials sum A at this time
=α1M=5M;
2.2. it continuously takes pictures simulation;
By the kth class faulty materials of n-th of grade after manual simulation and last a kind of normal productExisted by unified samples
Same light source Light0And camera Camera0Under the conditions of, with different angle Anglep, direction Directionp, position
PositionpThe mode exptended sample number continuously taken pictures is to original α2=500 times,
Wherein Anglep∈ ANGLE, p=1,2 ..., P, ANGLE are the angles that product is likely to occur on actual production line
Set;
Directionp∈ DIRECTION, p=1,2 ..., P, DIRECTION are that product may go out on actual production line
Existing direction set;
Positionp∈ POSITION, p=1,2 ..., P, POSITION are that product is likely to occur on actual production line
Location sets;
It is simulated by this, there is sample numberEvery class sample is divided into two groups: training setGather with test
To each b0∈B0Sample passes through camera Camera0It takes pictures, so that each b0Sample expands The noise that the photosensitive nuance of camera generates when the step by shooting every time carries out data extending, by continuously taking pictures
The sample number obtained after simulation isIt obtains training setGather with test
2.3. automatic imitation sample;
Algorithm for image enhancement simulates β0Good sample again, image is rotated, is translated, is cut, is filled, brightness, comparison
Degree, color difference operation, carry out sample number expansion for good sample in the case where not influencing product appearance structure and generating flaw, i.e.,
β is carried out to faulty materials using faulty materials simulation algorithm0Expansion again, to every a kind of faulty materialsK=1,2 ...,
K-1 generates the image library of such faulty materials Mainly byInterception, the interconnection of flaw image section in faulty materials image
The online image for searching for similar flaw part and the flaw image simulation based on dimensional gaussian distribution, wherein being based on dimensional Gaussian
The flaw image simulation of distribution uses following formula W, h ∈ [- 2 γn, 2 γn], μ1, μ2∈
[-γn, γn], ρ ∈ (- 1,1), σ1, σ2∈ (0,2 γn],
Wherein γnReduce with the increase of n, indicates that the flaw of simulation becomes smaller, the difficulty of identification gradually rises;
To given flaw type k and given flaw grade n, random selectionFromMiddle random selection product imageAnd random selection coordinate (w0, h0), it willImage is placed in(w0, h0) at simulation generate new faulty materials, weight
Multiple this process β0It is secondary, it obtains
To ownK=1,2 ..., K input generate in confrontation network DCGAN model, regenerate expansion and generate new sample
This
The threedimensional model of product is got, and is based on threedimensional modelRandomly choose coordinate (w1, h1) textures extremely production
In product threedimensional model, and light source Light is added in the three-dimensional model0And camera Camera0With different angle Anglep, direction
Directionp, position (PositionpIt carries out simulation to take pictures, repeats this processIt is secondary, it obtains
The new data for summarizing all generations, obtains Wherein α3=β0+β1+β2,J=0,1,2;
To obtain training setGather with test
Further, the main contents of step 4 include:
4.1 Squeezenet Pro Model Weights copy: by the ownership from Conv1 layers to Conv10 layers when initialization
Weight WI, I=1,2 ..., 10 is by Squeezenet V1.1 Model Weight WI(0) it is replaced;If had existed for
The training pattern of Squeezenet Pro, then all weight WIIt is replaced by the weight of newest Squeezenet Pro;
The replacement of 4.2Squeezenet Pro Model Weight: the hidden neuron σ of the softmax layer of top is replaced with into K
It is a, i.e.,
The W of 4.3 open Squeezenet ProiTraining is provided, frequency of training t is initializedn=0, j initial value are 10, input
Data, using AdamGradient optimization algorithm training W10, and observational learning curve, work as DvalidOn test result reach
Plateau marks leFor training precision, wherein e is epoch cycle of training, and mark ∈ is that training result minimum difference mark E is
The minimum of cycle of training is online, when training precision differenceWhen, deconditioning judges le1It is
It is no to be greater than 0.99, if it is, jumping to 4.4;If not i=i-1 is then arranged, the W of Squeezenet Pro is openediWeight
Training is provided, 4.3 re -trainings are jumped to;
4.4 judge current frequency of training trainnWhether < Q is true, and wherein Q is total training time under n-th grade of flaw
Number;If trainn< Q, then trainn=trainn+ 1, the training of screening again for the three carry out data that go to step;If
trainn> Q, then go to step five.
Further, the main contents of step 5 include: by the test set of step 3 using newest trained
Squeezenet Pro model is tested, if precision is more than or equal to 0.99, n=n+1 and increases flaw identification difficulty, and is jumped
Go to simulation and training that step 2 carries out small sample;If precision less than 0.99, judges current testing time testn< O
Whether true, wherein O is total testing time under n-th grade of flaw;If testn< O, then testn=testn+ 1, it jumps to
Step 2 carries out the simulation and training of small sample, if testn> O, then go to step six.
Another aspect of the present invention additionally provides a kind of system based on the efficient Checking model of extra small sample training, comprising:
Acquisition module: acquisition type is determined, including the faulty materials type and normal kind class to be detected, according to flaw size
The percentage for accounting for gross sample size is divided into N number of grade, acquires small sample;
Enlargement module: the small sample for various types of, each grade, which is simulated, to be expanded;
Decimation blocks: random sampling forms training set, verifying set, test set;Sequence is upset to each set;
Training module: Squeezenet Pro training Checking model;
Test module: test model result;It is jumped in next step according to precision and testing time control;
Release module: model is completed in publication, and the test result of each flaw rank obtained according to test module provides often
The accuracy of identification of level-one flaw, and the weight for freezing every level-one Squeezenet Pro is saved, it issues corresponding
Squeezenet Pro model.
Further, acquisition module includes:
Type unit: the determination faulty materials type to be detected is K-1 time, >=2 K, and total classification is faulty materials species number+just
Normal kind class=K class, every class sample collect Mk, k=1,2 ..., K small sample, it collects in totalA sample,
Wherein Mk∈ [5,10];
Level cells: being divided into N number of grade according to the percentage that flaw size accounts for gross sample size for every class faulty materials, wherein
Grouped element: every class sample is divided into two groups:WithWhereinForIn
80% sample randomly selected is as training set;ForIn 20% sample randomly selected as test
Set.
Further, enlargement module includes:
Manual analogue unit:
By the kth class faulty materials of n-th of grade (n is bigger, and identification difficulty is higher) and last a kind of normal productK=1,
2 ..., K extends to the α of raw sample number by made mode1=5 times, i.e.,Faulty materials sum A at this time
=α1M=5M;
Continuous analogue unit of taking pictures:
By the kth class faulty materials of n-th of grade after manual simulation and last a kind of normal productExisted by unified samples
Same light source Light0And camera Camera0Under the conditions of, with different angle Anglep, direction Directionp, position
PositionpThe mode exptended sample number continuously taken pictures is to original α2=500 times,
Wherein Anglep∈ ANGLE, p=1,2 ..., P, ANGLE are the angles that product is likely to occur on actual production line
Set;
Directionp∈ DIRECTION, p=1,2 ..., P, DIRECTION are that product may go out on actual production line
Existing direction set;
Positionp∈ POSITION, p=1,2 ..., P, POSITION are that product is likely to occur on actual production line
Location sets;
It is simulated by this, there is sample numberEvery class sample is divided into two groups: training setGather with test
To each b0∈B0Sample passes through camera Camera0It takes pictures, so that each b0Sample expands The noise that the photosensitive nuance of camera generates when the step by shooting every time carries out data extending, by continuously taking pictures
The sample number obtained after simulation isIt obtains training setGather with test
Automatic imitation sample unit:
Algorithm for image enhancement simulates β0Good sample again, image is rotated, is translated, is cut, is filled, brightness, comparison
Degree, color difference operation, carry out sample number expansion for good sample in the case where not influencing product appearance structure and generating flaw, i.e.,
β is carried out to faulty materials using faulty materials simulation algorithm0Expansion again, to every a kind of faulty materialsK=1,2 ...,
K-1 generates the image library of such faulty materials Mainly byInterception, the interconnection of flaw image section in faulty materials image
The online image for searching for similar flaw part and the flaw image simulation based on dimensional gaussian distribution, wherein being based on dimensional Gaussian
The flaw image simulation of distribution uses following formula W, h ∈ [- 2 γn, 2 γn], μ1, μ2∈
[-γn, γn], ρ ∈ (- 1,1), σ1, σ2∈ (0,2 γn],
Wherein γnReduce with the increase of n, indicates that the flaw of simulation becomes smaller, the difficulty of identification gradually rises;
To given flaw type k and given flaw grade n, random selectionFromMiddle random selection product imageAnd random selection coordinate (w0, h0), it willImage is placed in(w0, h0) at simulation generate new faulty materials, weight
Multiple this process β0It is secondary, it obtains
To ownK=1,2 ..., K input generates in confrontation network DCGAN model, and it is new to regenerate expansion generation
Sample
The threedimensional model of product is got, and is based on threedimensional modelRandomly choose coordinate (w1, h1) textures extremely production
In product threedimensional model, and light source Light is added in the three-dimensional model0And camera Camera0With different angle Anglep, direction
Directionp, position (PositionpIt carries out simulation to take pictures, repeats this processIt is secondary, it obtains
The new data for summarizing all generations, obtains Wherein α3=β0+β1+β2,J=0,1,2;
To obtain training setGather with test
Further, training module includes:
Squeezenet Pro Model Weight copy cell: by the ownership from Conv1 layers to Conv10 layers when initialization
Weight WI, I=1,2 ..., 10 is by Squeezenet V1.1 Model Weight WI(0) it is replaced;If had existed for
The training pattern of Squeezenet Pro, then all weight WIIt is replaced by the weight of newest Squeezenet Pro;
Squeezenet Pro Model Weight replacement unit: the hidden neuron σ of the softmax layer of top is replaced with
K, i.e.,
Training unit: the W of open Squeezenet ProiTraining is provided, frequency of training t is initializedn=0, j initial value is
10, input data, using AdamGradient optimization algorithm training W10, and observational learning curve, work as DvalidOn test result
Reach Plateau, marks leFor training precision, wherein e is epoch cycle of training, and mark ∈ is training result minimum difference mark
It is online to infuse the minimum that E is cycle of training, when training precision difference When, deconditioning, judgement
le1Whether 0.99 is greater than, if it is, judging unit;If not i=i-1 is then arranged, the W of Squeezenet Pro is openedi
Weight provides training, jumps to this training unit re -training;
Judging unit: judge current frequency of training trainnWhether < Q is true, and wherein Q is total instruction under n-th grade of flaw
Practice number;If trainn< Q, then trainn=trainn+ 1, jump to the training of screening again that decimation blocks carry out data;
If trainn> Q, then jump to test module.
Further, the main contents of test module include:
Precision judging unit: the test set that decimation blocks are obtained uses newest trained Squeezenet Pro model
It is tested, if precision is more than or equal to 0.99, n=n+1 and increases flaw identification difficulty, and it is small to jump to enlargement module progress
The simulation and training of sample;If precision turns frequency judging unit less than 0.99;
Frequency judging unit: judge current testing time testnWhether < O is true, and wherein O is under n-th grade of flaw
Total testing time;If testn< O, then testn=testn+ 1, simulation and training that enlargement module carries out small sample are jumped to,
If testn> O, then jump to release module.
Compared with prior art, the invention has the benefit that
1. use Conventional visual algorithm pretreatment image, prominent faulty materials and normal specimens difference, deep learning algorithm into
Row classification realizes that stability of the variation more only with Conventional visual or deep learning algorithm to light is higher.If in Conventional visual
There is subtle deviation in product polishing, can all impact to detection effect, needs to readjust parameter.And deep learning algorithm
The polishing difference that can occur by model training, adaptive product, without readjusting to parameter, working service is convenient.
2. the top layer using SqueezeNet is revised as being adjusted, the faulty materials classification judged as needed is arranged to top
The mode of the hidden neuron number of layer, it is stronger to the compatibility of multi-product, color.Conventional visual is for every a kind of faulty materials meeting
Need special polishing scheme and criterion, system compatibility very poor with appearance characteristics such as color, sizes.And deep learning is calculated
Method can be compatible with the detection demand of multi-product, different colours by the training to product model, and multi-product generates bumpless transfer.
3. using progressive algorithm, gradually training pattern makes deep learning model to product during training deep learning
The detection of tiny flaw is more stable.Conventional visual scheme processing be less than sample size 1% nuance when, in order to noise
Signal is mutually distinguished, and needs strictly to determine the features such as color, size, and smaller the spent debug time of flaw is longer, stability
It is poor;The accuracy rate of identification is enhanced while reducing the development time using progressive algorithm.
4. making to need a large amount of training originally using Pro editions transfer training algorithms of small sample training algorithm+SqueezeNet
The deep learning model of data only needs more than ten of training sample that training process can be completed, while can reduce out in training process
The case where existing local optimum, accelerates model convergence rate.
Detailed description of the invention
Fig. 1: Conventional visual of embodiment of the present invention algorithm combines schematic diagram with deep learning algorithm;
Fig. 2: Small Sample Database of the embodiment of the present invention expands schematic diagram;
Fig. 3: SqueezeNet of embodiment of the present invention Pro editions application schematic diagrams;
Fig. 4: SqueezeNet of embodiment of the present invention Pro editions adjustable schematic diagrames;
Fig. 5: SqueezeNet of embodiment of the present invention Pro editions faulty materials classification training schematic diagrames;
Fig. 6: gradually training pattern schematic diagram of the embodiment of the present invention.
Specific embodiment
It should be noted that in the absence of conflict, the feature in the embodiment of the present invention and embodiment can be mutual
Combination.
The invention proposes the deep learning model algorithms based on small sample training, pass through this method, it is ensured that
In the case where sample number considerably less (more than ten of sample), achieve the effect that the deep learning model of hundreds of thousands sample number training, from
And preferably it is applied in Industrial Inferential Measurements.
Step 1: acquisition small sample
1.1. the determination faulty materials type to be detected is K-1 times, >=2 K, and total classification is faulty materials species number+normal product
Type=K class, every class sample collect Mk, k=1,2 ..., K small sample, it collects in totalA sample, wherein one
As in the case of Mk∈ [5,10].
1.2. every class faulty materials are divided into N number of grade according to the percentage that flaw size accounts for gross sample size, wherein
1.3. every class sample is divided into two groups:WithWhereinForIn it is random
80% sample extracted is as training set (Training Set);ForIn 20% sample randomly selected
Product are as test set (Testing Set).
Step 2: small sample expands
2.1. manual analog sample
2.1.1. by the kth class faulty materials of n-th of grade (n is bigger, and identification difficulty is higher) and last a kind of normal productk
=1,2 ..., K extends to the α of raw sample number by made mode1=5 times, i.e.,Faulty materials are total at this time
Number A=α1M=5M.
2.2. it continuously takes pictures simulation
2.2.1. by the kth class faulty materials of n-th of grade after manual simulation and last a kind of normal productPass through unification
Sample is in same light source (Light0) and camera (Camera0) under the conditions of, with different angle (Anglep), direction
(Directionp), position (Positionp) the mode exptended sample number continuously taken pictures is to original α2=500 times, wherein
Anglep∈ ANGLE, p=1,2 ..., P, ANGLE are the angle sets that product is likely to occur on actual production line,;Similarly
There is Directionp∈ DIRECTION, p=1,2 ..., P, DIRECTION are the sides that product is likely to occur on actual production line
To set;Positionp∈ POSITION, p=1,2 ..., P, POSITION are that product is likely to occur on actual production line
Location sets.It is simulated by this, there is sample numberLikewise, havingFor based onThe training set (Training Set) of generation;For based onThe test set of generation
(Testing Set)。
2.2.2. to each b0∈B0Sample passes through camera Camera0It takes pictures, so that each b0Sample expands The noise that the photosensitive nuance of camera generates when the step by shooting every time carries out data extending, can be very
The process of effective simulation actual production line detection, so that deep learning be allowed to train the model of stable anti-noise.By continuously clapping
It is according to the sample number obtained after simulationLikewise, havingFor based onIt generates
Training set (Training Set);For based onThe test set (Testing Set) of generation.
2.3. automatic imitation sample
2.3.1. algorithm for image enhancement simulates β0Good sample (β again0It usually takes 10), image is rotated, translate, is cut out
It cuts, fill, brightness, contrast, the operation such as color difference, in the case where not influencing product appearance structure and generating flaw by good sample
Sample number expansion is carried out, i.e.,
2.3.2. β is carried out to faulty materials using faulty materials simulation algorithm0Expansion again, to every a kind of faulty materialsK=1,
2 ..., K-1 generates the image library of such faulty materials Mainly bySection of flaw image section in faulty materials image
It takes, search for the image of similar flaw part and the flaw image simulation based on dimensional gaussian distribution on internet, wherein being based on two
The flaw image simulation for tieing up Gaussian Profile uses following formula W, h ∈ [- 2 γn, 2 γn], μ1, μ2∈
[-γn, γn], ρ ∈ (- 1,1), σ1, σ2∈ (0,2 γn], wherein γnIt can reduce with the increase of n, indicate the flaw of simulation
Become smaller, the difficulty of identification gradually rises.To given flaw type k and given flaw grade n, random selection FromIn
Randomly choose product imageAnd random selection coordinate (w0, h0), it willImage is placed in(w0, h0) at simulate life
The faulty materials of Cheng Xin repeat this process β0It is secondary, therefore obtain
2.3.3. will ownK=1,2 ..., K input generate in confrontation network (DCGAN) model, will regenerate expansion
Fill generation new samples
2.3.4. the threedimensional model of product is got, and provided on threedimensional model based on 2.3.2Random selection is sat
Mark (w1, h1) textures are into product threedimensional model, and addition light source (Light in the three-dimensional model0) and camera (Camera0) with
Different angle (Anglep), direction (Directionp), position (Positionp) simulate and take pictures, repeat this process
It is secondary, therefore obtain
2.3.5. the new data for summarizing all generations, obtains Wherein α3=β0+β1+β2,J=0,1,2.
Note: in this algorithm, αi, i=1,2,3 can be greater than replaced 0 natural number by any.Likewise, havingFor based onThe training set (Training Set) of generation;For based on
The test set (Testing Set) of generation.
Step 3: random sampling forms training set
3.1. to given flaw type k and given flaw grade n, fromIn randomly select 50% data make
For training set Dtrain, 10% data are as verifying set Dvalid;FromIn randomly select 50% data as surveying
Examination collection Dtest.Sequence is upset to each D set, enters step four.
Step 4: Squeezenet Pro trains Checking model
4.1 Squeezenet Pro Model Weights copy: by the ownership from Conv1 layers to Conv10 layers when initialization
Weight WI, I=1,2 ..., 10 is by Squeezenet V1.1 Model Weight WI(0) it is replaced.If had existed for
The training pattern of Squeezenet Pro, then all weight WIIt is replaced by the weight of newest Squeezenet Pro.
The replacement of 4.2 Squeezenet Pro Model Weights: the hidden neuron σ of the softmax layer of top is replaced with
K, i.e.,
The W of 4.3 open Squeezenet ProiTraining is provided, frequency of training t is initializedn=0, j initial value are 10, input number
According to using AdamGradient optimization algorithm training W10, and observational learning curve, work as DvalidOn test result reach Plateau,
Mark leFor training precision, wherein e is cycle of training (epoch), and mark ∈ is training result minimum difference, usually 0.5%, it marks
It is online to infuse the minimum that E is cycle of training, usually 50, have when training precision differenceWhen, stop
Training, judges le1Whether 0.99 is greater than, if it is, jumping to
4.4;If not i=i-1 is then arranged, the W of Squeezenet Pro is openediWeight provides training, jumps to 4.3
Re -training.
4.4 judge current frequency of training trainnWhether < Q is true, and wherein Q is total training time under n-th grade of flaw
Number, general Q=5.If trainn< Q, then trainn=trainn+ 1, jump to the training of screening again of 3.1 carry out data.
If trainn> Q, then go to step five.
Step 5: test model result
5.1. by DtestIt is tested using newest trained Squeezenet Pro model, if precision is more than or equal to
0.99, then n=n+1 increases flaw identification difficulty, and jumps to the simulation and training of 2.1 carry out small samples.If precision is less than
0.99, then judge current testing time testnWhether < O is true, and wherein O is total testing time under n-th grade of flaw, generally
O=3.If testn< O, then testn=testn+ 1, jump to the simulation and training of 2.1 carry out small samples.If testn>
O, then go to step six.
Step 6: model is completed in publication
The test result of each flaw rank obtained according to step 5 provides the accuracy of identification of every level-one flaw, and protects
Deposit the weight for freezing every level-one Squeezenet ProIssue corresponding Squeezenet Pro model.
The features of the present invention is as follows:
It is detected as shown in Figure 1, Conventional visual algorithm combines common product quality of completing with deep learning algorithm.Tradition view
Feel algorithm pretreatment image, prominent faulty materials and normal specimens difference, deep learning algorithm are classified;
As shown in Fig. 2, after Small Sample Database of the invention provides, by continuously taking pictures, simulating bad sample, automatic mold manually
Intend bad sample, threedimensional model generates bad sample, generate confrontation network generate bad sample method carry out data extending, data are expanded
Fill 100,000 times;
To be applied in product quality detection as shown in figure 3, present invention firstly provides SqueezeNet Pro editions, guarantee algorithm
Precision while also faster than the speed of other models;
As shown in figure 4, the top layer of SqueezeNet Pro editions more traditional SqueezeNet is revised as being adjusted in the present invention,
The faulty materials classification judged as needed is arranged to the hidden neuron number of top layer;
As shown in figure 5, SqueezeNet Pro editions of the present invention all layers of the parameters other than top layer by training in advance
The parameter of 1000 class image recognitions is substituted, and is trained by the algorithm of transfer learning to faulty materials classification;
As shown in fig. 6, pass through progressive algorithm gradually training pattern in training process of the present invention, it now will most apparent faulty materials
By being trained after data extending algorithm, it is being stepped up the training for identifying that difficulty finally realizes all faulty materials types.
The foregoing describe the information such as basic principles and main features of the invention and embodiment, but the present invention is not by upper
The limitation for stating implementation process, under the premise of not departing from spirit and range, the present invention can also have various changes and modifications.
Therefore, unless this changes and improvements are departing from the scope of the present invention, they should be counted as comprising in the present invention.
Claims (10)
1. a kind of method based on the efficient Checking model of extra small sample training characterized by comprising
Step 1: determining that acquisition type accounts for gross sample according to flaw size including the faulty materials type and normal kind class to be detected
The percentage of product size is divided into N number of grade, acquires small sample;
Expand Step 2: the small sample of various types of, each grade is simulated;
Step 3: random sampling composition training set, verifying set, test set;Sequence is upset to each set;
Step 4: Squeezenet Pro trains Checking model;
Step 5: test model result;It is jumped in next step according to precision and testing time control;
Step 6: model is completed in publication, the test result of each flaw rank obtained according to step 5 provides every level-one flaw
Accuracy of identification, and save and freeze the weight of every level-one Squeezenet Pro, issue corresponding Squeezenet Pro mould
Type.
2. the method according to claim 1, wherein the particular content of step 1 includes:
1.1, the determination faulty materials type to be detected is K-1 times, >=2 K, and total classification is faulty materials species number+normal kind class
=K class, every class sample collect Mk, k=1,2 ..., K small sample, it collects in totalA sample, wherein Mk∈
[5,10];
1.2, every class faulty materials are divided into N number of grade according to the percentage that flaw size accounts for gross sample size, wherein
1.3, every class sample is divided into two groups:WithWhereinForIn randomly select
80% sample is as training set;ForIn 20% sample randomly selected as test set.
3. the method according to claim 1, wherein the particular content of step 2 includes:
2.1. manual analog sample;
By the kth class faulty materials of n-th of grade (n is bigger, and identification difficulty is higher) and last a kind of normal product
The α of raw sample number is extended to by made mode1=5 times, i.e.,Faulty materials sum A=α at this time1M=
5M;
2.2. it continuously takes pictures simulation;
By the kth class faulty materials of n-th of grade after manual simulation and last a kind of normal productBy unified samples same
Light source Light0And camera Camera0Under the conditions of, with different angle Anglep, direction Directionp, position PositionpEven
The continuous mode exptended sample number taken pictures is to original α2=500 times,
Wherein Anglep∈ ANGLE, p=1,2 ..., P, ANGLE are the angle sets that product is likely to occur on actual production line;
Directionp∈ DIRECTION, p=1,2 ..., P, DIRECTION are the sides that product is likely to occur on actual production line
To set;
Positionp∈ POSITION, p=1,2 ..., P, POSITION are the positions that product is likely to occur on actual production line
Set;
It is simulated by this, there is sample numberEvery class sample is divided into two groups: training setGather with test
To each b0∈B0Sample passes through camera Camera0It takes pictures, so that each b0Sample expands It should
The noise that the photosensitive nuance of camera generates when step by shooting every time carries out data extending, after simulation of continuously taking pictures
Obtained sample number isIt obtains training setGather with test
2.3. automatic imitation sample;
Algorithm for image enhancement simulates β0Image is rotated, is translated, being cut, being filled, brightness, contrast, color difference behaviour by good sample again
Make, good sample is subjected to sample number expansion in the case where not influencing product appearance structure and generating flaw, i.e.,
β is carried out to faulty materials using faulty materials simulation algorithm0Expansion again, to every a kind of faulty materials
Generate the image library of such faulty materials Mainly byIn the interception of flaw image section in faulty materials image, internet
The image of similar flaw part and the flaw image simulation based on dimensional gaussian distribution are searched for, wherein being based on dimensional gaussian distribution
Flaw image simulation use following formula
Wherein γnReduce with the increase of n, indicates that the flaw of simulation becomes smaller, the difficulty of identification gradually rises;
To given flaw type k and given flaw grade n, random selectionFromMiddle random selection product imageAnd random selection coordinate (w0, h0), it willImage is placed in(w0, h0) at simulation generate new faulty materials, weight
Multiple this process β0It is secondary, it obtains
To ownInput generates in confrontation network DCGAN model, regenerates expansion and generates new samples
The threedimensional model of product is got, and is based on threedimensional modelRandomly choose coordinate (w1, h1) textures to product three-dimensional
In model, and light source Light is added in the three-dimensional model0And camera Camera0With different angle Anglep, direction
Directionp, position (PositionpIt carries out simulation to take pictures, repeats this processIt is secondary, it obtains
The new data for summarizing all generations, obtains
Wherein
To obtain training setGather with test
4. the method according to claim 1, wherein the main contents of step 4 include:
4.1 Squeezenet Pro Model Weights copy: by all weight W from Conv1 layers to Conv10 layers when initializationI, I
=1,2 ..., 10 by Squeezenet V1.1 Model Weight WI(0) it is replaced;If having existed for Squeezenet Pro
Training pattern, then all weight WIIt is replaced by the weight of newest Squeezenet Pro;
The replacement of 4.2 Squeezenet Pro Model Weights: replacing with K for the hidden neuron σ of the softmax layer of top,
I.e.
The W of 4.3 open Squeezenet ProiTraining is provided, frequency of training t is initializedn=0, j initial value are 10, input number
According to using AdamGradient optimization algorithm training W10, and observational learning curve, work as DvalidOn test result reach
Plateau marks leFor training precision, wherein e is epoch cycle of training, and mark ∈ is that training result minimum difference mark E is
The minimum of cycle of training is online, when training precision differenceWhen, deconditioning judges le1Whether
Greater than 0.99, if it is, jumping to 4.4;If not i=i-1 is then arranged, the W of Squeezenet Pro is openediWeight mentions
For training, 4.3 re -trainings are jumped to;
4.4 judge current frequency of training trainnWhether < Q is true, and wherein Q is total frequency of training under n-th grade of flaw;If
trainn< Q, then trainn=trainn+ 1, the training of screening again for the three carry out data that go to step;If trainn> Q,
Then go to step five.
5. the method according to claim 1, wherein the main contents of step 5 include: by the test of step 3
Collection is tested using newest trained Squeezenet Pro model, if precision increases more than or equal to 0.99, n=n+1
Flaw is added to identify difficulty, and the simulation and training for the two carry out small samples that go to step;If precision, less than 0.99, judgement is worked as
Preceding testing time testnWhether < O is true, and wherein O is total testing time under n-th grade of flaw;If testn< O, then
testn=testn+ 1, the simulation and training of the two carry out small samples that go to step, if testn> O, then go to step six.
6. a kind of system based on the efficient Checking model of extra small sample training characterized by comprising
Acquisition module: determine that acquisition type accounts for always including the faulty materials type and normal kind class to be detected according to flaw size
The percentage of sample size is divided into N number of grade, acquires small sample;
Enlargement module: the small sample for various types of, each grade, which is simulated, to be expanded;
Decimation blocks: random sampling forms training set, verifying set, test set;Sequence is upset to each set;
Training module: Squeezenet Pro training Checking model;
Test module: test model result;It is jumped in next step according to precision and testing time control;
Release module: model is completed in publication, and the test result of each flaw rank obtained according to test module provides every level-one
The accuracy of identification of flaw, and the weight for freezing every level-one Squeezenet Pro is saved, issue corresponding Squeezenet
Pro model.
7. system according to claim 6, which is characterized in that acquisition module includes:
Type unit: the determination faulty materials type to be detected is K-1 times, >=2 K, and total classification is faulty materials species number+normal product
Type=K class, every class sample collect Mk, k=1,2 ..., K small sample, it collects in totalA sample, wherein Mk
∈ [5,10];
Level cells: being divided into N number of grade according to the percentage that flaw size accounts for gross sample size for every class faulty materials, wherein
Grouped element: every class sample is divided into two groups:WithWhereinForIn it is random
80% sample extracted is as training set;ForIn 20% sample randomly selected as test set.
8. system according to claim 6, which is characterized in that enlargement module includes:
Manual analogue unit:
By the kth class faulty materials of n-th of grade (n is bigger, and identification difficulty is higher) and last a kind of normal product
The α of raw sample number is extended to by made mode1=5 times, i.e.,Faulty materials sum A=α at this time1M=
5M;
Continuous analogue unit of taking pictures:
By the kth class faulty materials of n-th of grade after manual simulation and last a kind of normal productBy unified samples same
Light source Light0And camera Camera0Under the conditions of, with different angle Anglep, direction Directionp, position PositionpEven
The continuous mode exptended sample number taken pictures is to original α2=500 times,
Wherein Anglep∈ ANGLE, p=1,2 ..., P, ANGLE are the angle sets that product is likely to occur on actual production line;
Directionp∈ DIRECTION, p=1,2 ..., P, DIRECTION are the sides that product is likely to occur on actual production line
To set;
Positionp∈ POSITION, p=1,2 ..., P, POSITION are the positions that product is likely to occur on actual production line
Set;
It is simulated by this, there is sample numberEvery class sample is divided into two groups: training setGather with test
To each b0∈B0Sample passes through camera Camera0It takes pictures, so that each b0Sample expands It should
The noise that the photosensitive nuance of camera generates when step by shooting every time carries out data extending, after simulation of continuously taking pictures
Obtained sample number isIt obtains training setGather with test
Automatic imitation sample unit:
Algorithm for image enhancement simulates β0Good sample again, image is rotated, is translated, is cut, is filled, brightness, contrast, color
Difference operation, carries out sample number expansion for good sample in the case where not influencing product appearance structure and generating flaw, i.e.,
β is carried out to faulty materials using faulty materials simulation algorithm0Expansion again, to every a kind of faulty materials
Generate the image library of such faulty materials Mainly byIn the interception of flaw image section in faulty materials image, internet
The image of similar flaw part and the flaw image simulation based on dimensional gaussian distribution are searched for, wherein being based on dimensional gaussian distribution
Flaw image simulation use following formula
Wherein γnReduce with the increase of n, indicates that the flaw of simulation becomes smaller, the difficulty of identification gradually rises;
To given flaw type k and given flaw grade n, random selectionFromMiddle random selection product imageAnd random selection coordinate (w0, h0), it willImage is placed in(w0, h0) at simulation generate new faulty materials, weight
Multiple this process β0It is secondary, it obtains
To ownInput generates in confrontation network DCGAN model, regenerates expansion and generates new samples
The threedimensional model of product is got, and is based on threedimensional modelRandomly choose coordinate (w1, h1) textures to product three-dimensional
In model, and light source Light is added in the three-dimensional model0And camera Camera0With different angle Anglep, direction
Directionp, position (PositionpIt carries out simulation to take pictures, repeats this processIt is secondary, it obtains
The new data for summarizing all generations, obtains
Wherein
To obtain training setGather with test
9. system according to claim 6, which is characterized in that training module includes:
Squeezenet Pro Model Weight copy cell: by all weight W from Conv1 layers to Conv10 layers when initializationI, I
=1,2 ..., 10 by Squeezenet V1.1 Model Weight WI(0) it is replaced;If having existed for Squeezenet Pro
Training pattern, then all weight WIIt is replaced by the weight of newest Squeezenet Pro;
Squeezenet Pro Model Weight replacement unit: replacing with K for the hidden neuron σ of the softmax layer of top,
I.e.
Training unit: the W of open Squeezenet ProiTraining is provided, frequency of training t is initializedn=0, j initial value are 10, defeated
Enter data, using AdamGradient optimization algorithm training W10, and observational learning curve, work as DvalidOn test result reach
Plateau marks leFor training precision, wherein e is epoch cycle of training, and mark ∈ is that training result minimum difference mark E is
The minimum of cycle of training is online, when training precision difference When, deconditioning judges le1Whether
Greater than 0.99, if it is, judging unit;If not i=i-1 is then arranged, the W of Squeezenet Pro is openediWeight mentions
For training, this training unit re -training is jumped to;
Judging unit: judge current frequency of training trainnWhether < Q is true, and wherein Q is total training time under n-th grade of flaw
Number;If trainn< Q, then trainn=trainn+ 1, jump to the training of screening again that decimation blocks carry out data;If
trainn> Q, then jump to test module.
10. system according to claim 6, which is characterized in that the main contents of test module include:
Precision judging unit: the test set that decimation blocks are obtained is carried out using newest trained Squeezenet Pro model
Test if precision is more than or equal to 0.99, n=n+1 and increases flaw identification difficulty, and jumps to enlargement module and carries out small sample
Simulation and training;If precision turns frequency judging unit less than 0.99;
Frequency judging unit: judge current testing time testnWhether < O is true, and wherein O is total test under n-th grade of flaw
Number;If testn< O, then testn=testn+ 1, simulation and training that enlargement module carries out small sample are jumped to, if
testn> O, then jump to release module.
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810929603.8A CN109142374A (en) | 2018-08-15 | 2018-08-15 | Method and system based on the efficient Checking model of extra small sample training |
CN202210197654.2A CN114577812A (en) | 2018-08-15 | 2018-08-15 | Method and system for training efficient quality inspection model based on ultra-small sample |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810929603.8A CN109142374A (en) | 2018-08-15 | 2018-08-15 | Method and system based on the efficient Checking model of extra small sample training |
Related Child Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210197654.2A Division CN114577812A (en) | 2018-08-15 | 2018-08-15 | Method and system for training efficient quality inspection model based on ultra-small sample |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109142374A true CN109142374A (en) | 2019-01-04 |
Family
ID=64789636
Family Applications (2)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210197654.2A Pending CN114577812A (en) | 2018-08-15 | 2018-08-15 | Method and system for training efficient quality inspection model based on ultra-small sample |
CN201810929603.8A Pending CN109142374A (en) | 2018-08-15 | 2018-08-15 | Method and system based on the efficient Checking model of extra small sample training |
Family Applications Before (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210197654.2A Pending CN114577812A (en) | 2018-08-15 | 2018-08-15 | Method and system for training efficient quality inspection model based on ultra-small sample |
Country Status (1)
Country | Link |
---|---|
CN (2) | CN114577812A (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110263934A (en) * | 2019-05-31 | 2019-09-20 | 中国信息通信研究院 | A kind of artificial intelligence data mask method and device |
CN110390717A (en) * | 2019-07-02 | 2019-10-29 | 北京字节跳动网络技术有限公司 | 3D model reconstruction method, device and electronic equipment |
CN110689477A (en) * | 2019-09-07 | 2020-01-14 | 创新奇智(重庆)科技有限公司 | Universal flaw image simulation method |
CN112200217A (en) * | 2020-09-09 | 2021-01-08 | 天津津航技术物理研究所 | Identification algorithm evaluation method and system based on infrared image big data |
CN113920396A (en) * | 2021-10-08 | 2022-01-11 | 中国人民解放军军事科学院军事医学研究院 | Method, system and equipment for quantitatively evaluating visual cognitive ability of special post personnel |
CN114324124A (en) * | 2021-12-21 | 2022-04-12 | 桂林优利特医疗电子有限公司 | Scattered sample data expansion method based on probability density estimation |
TWI770699B (en) * | 2020-10-14 | 2022-07-11 | 台達電子工業股份有限公司 | System of automatically generating training images and method thereof |
US11709913B2 (en) | 2020-10-14 | 2023-07-25 | Delta Electronics, Inc. | Automatic generation system of training image and method thereof |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20170227673A1 (en) * | 2016-02-08 | 2017-08-10 | Goodrich Corporation | Material detection systems |
CN107316007A (en) * | 2017-06-07 | 2017-11-03 | 浙江捷尚视觉科技股份有限公司 | A kind of monitoring image multiclass object detection and recognition methods based on deep learning |
CN105118044B (en) * | 2015-06-16 | 2017-11-07 | 华南理工大学 | A kind of wheel shape cast article defect automatic testing method |
CN107705284A (en) * | 2017-07-17 | 2018-02-16 | 苏州佳赛特智能科技有限公司 | A kind of detection method of surface flaw based on Bayes's small-sample learning |
EP3300002A1 (en) * | 2016-09-22 | 2018-03-28 | Styria medijski servisi d.o.o. | Method for determining the similarity of digital images |
CN107958257A (en) * | 2017-10-11 | 2018-04-24 | 华南理工大学 | A kind of Chinese traditional medicinal materials recognition method based on deep neural network |
CN108038853A (en) * | 2017-12-18 | 2018-05-15 | 浙江工业大学 | A kind of ceramic tile surface defect identification method based on convolutional neural networks and Active Learning |
CN108108768A (en) * | 2017-12-29 | 2018-06-01 | 清华大学 | Photovoltaic glass defect classification method and device based on convolutional neural networks |
-
2018
- 2018-08-15 CN CN202210197654.2A patent/CN114577812A/en active Pending
- 2018-08-15 CN CN201810929603.8A patent/CN109142374A/en active Pending
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105118044B (en) * | 2015-06-16 | 2017-11-07 | 华南理工大学 | A kind of wheel shape cast article defect automatic testing method |
US20170227673A1 (en) * | 2016-02-08 | 2017-08-10 | Goodrich Corporation | Material detection systems |
EP3300002A1 (en) * | 2016-09-22 | 2018-03-28 | Styria medijski servisi d.o.o. | Method for determining the similarity of digital images |
CN107316007A (en) * | 2017-06-07 | 2017-11-03 | 浙江捷尚视觉科技股份有限公司 | A kind of monitoring image multiclass object detection and recognition methods based on deep learning |
CN107705284A (en) * | 2017-07-17 | 2018-02-16 | 苏州佳赛特智能科技有限公司 | A kind of detection method of surface flaw based on Bayes's small-sample learning |
CN107958257A (en) * | 2017-10-11 | 2018-04-24 | 华南理工大学 | A kind of Chinese traditional medicinal materials recognition method based on deep neural network |
CN108038853A (en) * | 2017-12-18 | 2018-05-15 | 浙江工业大学 | A kind of ceramic tile surface defect identification method based on convolutional neural networks and Active Learning |
CN108108768A (en) * | 2017-12-29 | 2018-06-01 | 清华大学 | Photovoltaic glass defect classification method and device based on convolutional neural networks |
Non-Patent Citations (2)
Title |
---|
HALIL DURMUÚ ET AL.,: "Disease Detection on the Leaves of the Tomato Plants by Using Deep Learning", 《2017 6TH INTERNATIONAL CONFERENCE ON AGRO-GEOINFORMATICS》 * |
李腾飞等: "基于迭代深度学习的缺陷检测", 《计算机与数字工程》 * |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110263934A (en) * | 2019-05-31 | 2019-09-20 | 中国信息通信研究院 | A kind of artificial intelligence data mask method and device |
CN110390717A (en) * | 2019-07-02 | 2019-10-29 | 北京字节跳动网络技术有限公司 | 3D model reconstruction method, device and electronic equipment |
CN110390717B (en) * | 2019-07-02 | 2023-03-31 | 北京字节跳动网络技术有限公司 | 3D model reconstruction method and device and electronic equipment |
CN110689477A (en) * | 2019-09-07 | 2020-01-14 | 创新奇智(重庆)科技有限公司 | Universal flaw image simulation method |
CN110689477B (en) * | 2019-09-07 | 2023-01-10 | 创新奇智(重庆)科技有限公司 | Universal flaw image simulation method |
CN112200217A (en) * | 2020-09-09 | 2021-01-08 | 天津津航技术物理研究所 | Identification algorithm evaluation method and system based on infrared image big data |
CN112200217B (en) * | 2020-09-09 | 2023-06-09 | 天津津航技术物理研究所 | Identification algorithm evaluation method and system based on infrared image big data |
TWI770699B (en) * | 2020-10-14 | 2022-07-11 | 台達電子工業股份有限公司 | System of automatically generating training images and method thereof |
US11709913B2 (en) | 2020-10-14 | 2023-07-25 | Delta Electronics, Inc. | Automatic generation system of training image and method thereof |
CN113920396A (en) * | 2021-10-08 | 2022-01-11 | 中国人民解放军军事科学院军事医学研究院 | Method, system and equipment for quantitatively evaluating visual cognitive ability of special post personnel |
CN114324124A (en) * | 2021-12-21 | 2022-04-12 | 桂林优利特医疗电子有限公司 | Scattered sample data expansion method based on probability density estimation |
Also Published As
Publication number | Publication date |
---|---|
CN114577812A (en) | 2022-06-03 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109142374A (en) | Method and system based on the efficient Checking model of extra small sample training | |
CN108960319B (en) | It is a kind of to read the candidate answers screening technique understood in modeling towards global machine | |
CN105930277B (en) | A kind of defect source code localization method based on defect report analysis | |
CN107820620A (en) | Method and system for defect classification | |
CN110992349A (en) | Underground pipeline abnormity automatic positioning and identification method based on deep learning | |
CN110185939A (en) | Gas pipe leakage recognition methods based on convolutional neural networks | |
CN110517259A (en) | A kind of detection method, device, equipment and the medium of product surface state | |
CN104809476B (en) | A kind of multi-target evolution Fuzzy Rule Classification method based on decomposition | |
CN109657799A (en) | A kind of model tuning method and apparatus based on scene adaptation | |
US20220222807A1 (en) | Ai-based new learning model generation system for vision inspection on product production line | |
CN109146873A (en) | A kind of display screen defect intelligent detecting method and device based on study | |
CN111199543A (en) | Refrigerator-freezer surface defect detects based on convolutional neural network | |
CN108763096A (en) | Software Defects Predict Methods based on depth belief network algorithm support vector machines | |
US20220245402A1 (en) | Ai-based pre-training model determination system, and ai-based vision inspection management system using same for product production lines | |
CN109063983A (en) | A kind of natural calamity loss real time evaluating method based on social media data | |
CN108960884A (en) | Information processing method, model building method and device, medium and calculating equipment | |
CN111191935A (en) | Evaluation and assessment system and method | |
CN108090905B (en) | The judgment method and system of producing line exception | |
CN117574201A (en) | Model training method, device, equipment and storage medium based on multi-industry model | |
KR101745874B1 (en) | System and method for a learning course automatic generation | |
CN115240031A (en) | Method and system for generating plate surface defects based on generation countermeasure network | |
CN113779330B (en) | Ultrasonic detection process parameter intelligent computing system based on similarity measurement | |
CN109871906A (en) | A kind of classification method of the coil open defect based on depth convolutional neural networks | |
CN112015830B (en) | Question storage method suitable for adaptive learning | |
CN107885730A (en) | Translation knowledge method for distinguishing validity under more interpreter's patterns |
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 | ||
CB02 | Change of applicant information |
Address after: 518000 4th floor, Zhongdian Difu building, Zhenhua Road, Fuqiang community, Huaqiang North Street, Futian District, Shenzhen City, Guangdong Province Applicant after: Xinjian intelligent control (Shenzhen) Technology Co.,Ltd. Address before: 510000 unit 404, 4th floor, building B3, creative building, 162 science Avenue, Science City, Guangzhou high tech Industrial Development Zone, Guangdong Province Applicant before: GUANGZHOU XINJIAN ZHIKONG TECHNOLOGY Co.,Ltd. |
|
CB02 | Change of applicant information | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190104 |
|
RJ01 | Rejection of invention patent application after publication |