CN109584208A - A kind of method of inspection for industrial structure defect intelligent recognition model - Google Patents
A kind of method of inspection for industrial structure defect intelligent recognition model Download PDFInfo
- Publication number
- CN109584208A CN109584208A CN201811237173.XA CN201811237173A CN109584208A CN 109584208 A CN109584208 A CN 109584208A CN 201811237173 A CN201811237173 A CN 201811237173A CN 109584208 A CN109584208 A CN 109584208A
- Authority
- CN
- China
- Prior art keywords
- threshold value
- inspection
- identification
- frame
- registration
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10004—Still image; Photographic image
-
- 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
- G06T2207/20021—Dividing image into blocks, subimages or windows
-
- 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
- G06T2207/20076—Probabilistic image processing
-
- 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]
-
- 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
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30108—Industrial image inspection
Abstract
The invention discloses a kind of methods of inspection for industrial structure defect intelligent recognition model, comprising steps of 1) preproduction phase, inspection data collection is formed by Image Acquisition, arrangement and screening, object to be checked is marked using rectangle frame to each picture, the image that inspection data is concentrated later inputs identification network to be tested, obtains its recognition result；2) confidence threshold value is selected according to engine request, each confidence level is calculated in single image presets answering marks frame with each higher than the identification indicia framing of threshold value and is overlapped angle value CAr；3) the registration matrix of single image is constructed；4) registration threshold value is selected according to engine request, and judges to identify that frame is corrected errors and statistical magnitude；5) the identification recall rate and accuracy rate of assessment models is calculated.The present invention can overcome traditional mAP method of inspection when calculating multiple dimensioned divisible target identification effect, and identification frame shakes model evaluation result problem of dtmf distortion DTMF caused by overlapping on a large scale.
Description
Technical field
The invention belongs to field of artificial intelligence, and in particular to a kind of for industrial structure defect intelligent recognition model
The method of inspection.
Background technique
Artificial intelligence technology has become the trend, tendency of current global scientific research, industry development, and the progress of this field technology is drawn
The all trades and professions technological innovation tide not seen before into a wheel is led, is given birth to currently, artificial intelligence technology has begun to production
Various aspects infiltration in work.In traditional industry field, safety monitoring needs to expend huge manpower and material resources, artificial intelligence skill
Art, intelligent recognition model especially therein have become a kind of new efficient solutions of the problem.
Paper " the faster r-cnn:towards real time published by Ren Shaoqing et al. for 2015
Object detection with region proposal networks " [1], providing one kind can be directly from picture
High speed, the high-accuracy method of positioning and identification multiclass object, structure diagram are shown in attached drawing 1.The model is rolled up using depth first
Product neural network extracts the feature in image with information, forms characteristic pattern, the concept of anchor point and candidate frame is subsequently introduced
And two points are carried out to all candidate frames using Regionproposalnetwork (RPN), finally it is with the boundary for choosing candidate frame
Foundation, the feature for extracting corresponding position from the characteristic pattern of extraction carry out Classification and Identification using Softmax classifier.Except this it
Outside, it is suggested there are also several functionally similar intelligent recognition networks, their common trait is: being identified for each
Object uses that fully defining rectangular box marks in the picture by four parameters, while an object can generate multiple small models
Enclose the identification frame of shake.In actual scientific research and engineer application, often come using non-maxima suppression (NMS) Processing Algorithm
Overcome the problems, such as that single object repeats to identify.
It is the class model in application and industrial defect that whether intelligent recognition model, which can effectively and reliably complete identification mission,
Most concerned problem when identification application.Therefore, model testing index algorithm development, scientific research, in terms of all
Particularly significant, the objective method of inspection is the premise of model optimization and industrial products evaluation.Currently, universal in intelligent recognition field
Mean accuracy (mAP) method used is a kind of objective effective inspection gimmick, but is found in practice, differs nothing in visual effect
In the case where different, level of the mAP index obtained in industrial defect recognition inspection far below identical algorithms to general Object identifying.
By multiple network training and Inspection Research, it is found that this problem is prevalent in a kind of special objective: multiple dimensioned divisible
Target, i.e., any one part in such object images may be considered an independent target, and attached drawing 2 is to this feature
It shows.This makes the shake identification frame coverage area for an object very big, non-maxima suppression (NMS) Processing Algorithm
No longer valid, the model testing index obtained accordingly is also no longer objective credible.
Bibliography:
[1]Ren S,He K,Girshick R,et al.Faster R-CNN:towards real-time object
detection with region proposal networks[C]//International Conference on
Neural Information Processing Systems.MIT Press,2015:91-99.
Summary of the invention
The purpose of the present invention is to provide a kind of succinct, efficient, Practical significances to be explicitly directed to industrial structure defect intelligence
The method of inspection of identification model, this method can overcome traditional mAP method of inspection calculating multiple dimensioned divisible target identification effect
When fruit, identification frame shakes model evaluation result problem of dtmf distortion DTMF caused by overlapping on a large scale.
What the present invention was achieved through the following technical solutions:
A kind of method of inspection for industrial structure defect intelligent recognition model, comprising the following steps:
1) preproduction phase forms inspection data collection by Image Acquisition, arrangement and screening, uses rectangle to each picture
Collimation mark remembers object to be checked, and the model evaluation constituted in form for several label coordinates presets answer, later concentrates inspection data
Image input identification network to be tested, obtain its recognition result, be in form several indicia framing coordinates and its confidence level；
2) confidence threshold value is selected according to engine request, the knowledge that each confidence level is higher than threshold value is calculated in single image
Other indicia framing is overlapped angle value CAr with each default answering marks frame；
3) the registration matrix of single image is constructed；
4) registration threshold value is selected according to engine request, and judges to identify that frame is corrected errors and statistical magnitude；
5) the identification recall rate and accuracy rate of assessment models are calculated according to the following formula:
A further improvement of the present invention lies in that for general task, confidence threshold value takes 0.5, for high standard in step 2)
Quasi- task, confidence threshold value take 0.1.
A further improvement of the present invention lies in that the calculation formula used in step 2) is as follows:
In formula, Ai, Dj indicate that any default answering marks frame and confidence level are higher than the identification indicia framing of threshold value, SAi、SDjTable
Show that its area, S (Ai ∩ Dj) indicate the area of the two intersecting area.
A further improvement of the present invention lies in that in step 3), according to the registration matrix of following format building single image:
The every a line of matrix indicates that a confidence level is higher than the identification indicia framing of threshold value, and each column indicate a default answer mark
Remember that frame, each element representation current line are overlapped angle value with when forefront indicia framing.
A further improvement of the present invention lies in that in step 4), judge that identification frame is corrected errors and statistical magnitude according to following standard:
401) for each row element in matrix, if at least one value is more than or equal to registration threshold value, it is believed that should
The corresponding default answering marks of row are correctly detected；
402) for each column element in matrix, if at least one value is more than or equal to registration threshold value, it is believed that should
It is correct higher than the identification label of threshold value to arrange corresponding confidence level, does not fall through, accordingly the correct detection quantity of statistics and accurate detection
Quantity.
A further improvement of the present invention lies in that for general task, registration threshold value takes 0.5, for high standard task, weight
Right threshold value takes 0.8.
The present invention has following beneficial technical effect:
The present invention is universal according to industrial structure defect for the purpose of the practical effect of objective evaluation intelligent recognition model
Multiple dimensioned divisible feature, propose to can be widely applied to the new algorithm of industrial structure defect recognition model testing.It has such as
Lower advantage:
First: algorithm is concise, and computer program easy to use is realized, matrix variation and calculating wherein included
Method design is suitble to using multithreading acceleration or high-performance GPU operation, high-efficient.
Second: algorithm is overlapped angle value by definition, weakens the scale requirement to identification frame and indicia framing, this change symbol
The common multiple dimensioned divisible feature of industrial structure defect is closed, therefore overcomes traditional mAP method of inspection registration algorithm application
In this numerical value conversion problem of dtmf distortion DTMF.
Third: for algorithm by defining registration matrix, concise practical significance and difficulty in computation weaken all kinds of classical moulds
Preset answer is required correspondingly with recognition result in type, this change meets common multiple dimensioned of industrial structure defect
Divisible feature, numerical value conversion problem of dtmf distortion DTMF caused by overcoming all kinds of methods of inspection for these reasons.
4th: two key indexes --- recall rate and the accuracy rate that algorithm proposes meet all kinds of warps in similar problems
Allusion quotation index definition mode, therefore have fabulous practical value.
In conclusion the method for inspection provided by the invention for industrial structure defect intelligent recognition model, passes through reality
Verifying, it is ensured that the feasibility and validity of its algorithm.Using algorithm propose two key indexes, can easily with tradition
Algorithm combines, and forms the evaluation of monodrome, it means that expanding space using the algorithm as major part is big.In addition, this hair
Bright use scope is not limited to the Object identifying model testing for having multiple dimensioned divisible feature, and it is right to can be equally used for tradition
As examining, and provide the inspection result similar with classical method of inspection trend.The present invention has passed through actual verification, it is ensured that it can
By property.
Detailed description of the invention
Fig. 1 is the structural schematic diagram that Advanced target identifies network fasterR-CNN.
Fig. 2 is the multiple dimensioned divisible schematic diagram of industrial injury object, wherein Fig. 2 (a) is that crackle Object identifying is (typical more
The divisible industrial injury to be checked of scale) example, Fig. 2 (b) is that animal identification (the general object of no multiple dimensioned divisible feature) shows
Example.
Fig. 3 is registration index contrast schematic diagram used in the registration index redefined and traditional algorithm, wherein
Fig. 3 (a) is that registration the index IoU, Fig. 3 (b) that traditional algorithm defines are the registration index CAr that the present invention newly defines.
Fig. 4 is the schematic diagram that the present invention carries out recall rate and accuracy calculating process by defining registration matrix.
Fig. 5 is three authentication images prepared in embodiment, visualizes box comprising original image and its default answering marks,
The object coordinates listed in box and embodiment are consistent.
Fig. 6 is the recognition result for carrying out automatic identification acquisition in embodiment by network, and box and its confidence level are in figure
Mark.
Specific embodiment
The present invention is made further instructions below in conjunction with drawings and examples.
A kind of method of inspection for industrial defect intelligent recognition model provided by the invention, specific steps include:
Step 1, the preproduction phase forms inspection data collection by Image Acquisition, arrangement and screening, uses each picture
Rectangle frame marks object to be checked, and the model evaluation constituted in form for several label coordinates presets answer (Groundtruth).It
The image that inspection data is concentrated afterwards inputs identification network to be tested, obtains its recognition result, is in form several indicia framings seat
Mark and its confidence level (Confidence).One group of default answer for obtaining the secondary inspection through the above steps is sat with corresponding label
After mark and confidence level, complete to examine according to following 5 step.
Step 2, suitable confidence threshold value is selected according to engine request, generally takes 0.5 (general task) or 0.1 (high standard
Quasi- task), calculate identification indicia framing of each confidence level higher than threshold value and each according to the following formula in single image
Registration (CAr) value of default answering marks frame:
In formula, Ai, Dj indicate that any default answering marks frame and confidence level are higher than the identification indicia framing of threshold value, SAi、SDjTable
Show that its area, S (Ai ∩ Dj) indicate the area of the two intersecting area.
Step 3, according to the registration matrix (CArMatrix) of following format building single image:
The every a line of matrix indicates that a confidence level is higher than the identification indicia framing of threshold value, and each column indicate a default answer mark
Remember that frame, each element representation current line are overlapped angle value with when forefront indicia framing.
Step 4, it is selected suitable registration threshold value (indicating the accuracy requirement to identification object) according to engine request, one
As take 0.5 (general task) or 0.8 (high standard task).And judge that identification frame is corrected errors and statistical magnitude according to following standard:
401) for each row element in matrix, if at least one value is more than or equal to registration threshold value, it is believed that should
The corresponding default answering marks of row are correctly detected；
402) for each column element in matrix, if at least one value is more than or equal to registration threshold value, it is believed that should
It is correct higher than the identification label of threshold value to arrange corresponding confidence level, does not fall through, accordingly the correct detection quantity of statistics and accurate detection
Quantity.
Step 5, the identification recall rate and accuracy rate of assessment models are calculated according to the following formula:
Step 6, according to concrete engineering task and requirement, using the recall rate of calculating and accuracy rate as foundation, comprehensive analysis section
Learn research, technological development, the target identification effect that network is used in engineer application, and accordingly progress model optimization, parameter tuning,
The work such as model structure adjustment, engineering model examination.Specific example is as follows:
It works for model foundation, can be assessed according to These parameters and establish model level.
Recall rate and accuracy rate index need to be set according to actual conditions for engineering development, using with examination, and with above-mentioned
The inspection recall rate that method obtains is compared with accuracy rate, to complete intelligent detection equipment exploitation and checking and accepting.
Fig. 3 is registration index contrast schematic diagram used in the registration index redefined and traditional algorithm, wherein
Fig. 3 (a) is that registration the index IoU, Fig. 3 (b) that traditional algorithm defines are the registration index CAr that the present invention newly defines.Fig. 4 is
The present invention carries out the schematic diagram of recall rate and accuracy calculating process by defining registration matrix.
Embodiment:
It tests to the Identification of Cracks intelligent network based on fasterR-CNN for having completed training
The default answering marks for preparing the training set comprising 88 images first with them are as follows:
Example images are shown in attached drawing 5, for three authentication images prepared in embodiment, include original image and its default answer mark
Note visualizes box, and the object coordinates listed in box and embodiment are consistent.
Checking image is inputted into network later, obtains form recognition effect consistent with the above, an identification frame includes 5
Parameter (4 positional parameters and confidence level), image viewing effect is shown in attached drawing.
The recall rate and accuracy rate value of every image then are obtained according to above-mentioned calculating step, average 88 image results obtain
To the general performance of whole network, final result such as following table.
| Average recall rate | Average Accuracy | Traditional mAP is examined | |
| Examine point (%) | 96.63 | 90.46 | 41.6 |
On the basis of attached image recognition visual effect shown in fig. 6, it is more of the invention in evaluation method score and tradition
MAP examine score, can obviously discover traditional mAP method of inspection to model capability underestimate and the present invention in evaluation method it is superior
Property.
Claims (6)
1. a kind of method of inspection for industrial structure defect intelligent recognition model, which comprises the following steps:
1) preproduction phase forms inspection data collection by Image Acquisition, arrangement and screening, uses rectangle collimation mark to each picture
Remember object to be checked, the model evaluation constituted in form for several label coordinates presets answer, the figure for later concentrating inspection data
As inputting identification network to be tested, its recognition result is obtained, is in form several indicia framing coordinates and its confidence level；
2) confidence threshold value is selected according to engine request, the identification mark that each confidence level is higher than threshold value is calculated in single image
Note frame is overlapped angle value CAr with each default answering marks frame；
3) the registration matrix of single image is constructed；
4) registration threshold value is selected according to engine request, and judges to identify that frame is corrected errors and statistical magnitude；
5) the identification recall rate and accuracy rate of assessment models are calculated according to the following formula:
2. a kind of method of inspection for industrial structure defect intelligent recognition model according to claim 1, feature exist
In in step 2), for general task, confidence threshold value takes 0.5, and for high standard task, confidence threshold value takes 0.1.
3. a kind of method of inspection for industrial structure defect intelligent recognition model according to claim 1, feature exist
In the calculation formula used in step 2) is as follows:
In formula, Ai, Dj indicate that any default answering marks frame and confidence level are higher than the identification indicia framing of threshold value, SAi、SDjIndicate it
Area, S (Ai ∩ Dj) indicate the area of the two intersecting area.
4. a kind of method of inspection for industrial structure defect intelligent recognition model according to claim 3, feature exist
In in step 3), according to the registration matrix of following format building single image:
The every a line of matrix indicates that a confidence level is higher than the identification indicia framing of threshold value, and each column indicate a default answering marks
Frame, each element representation current line are overlapped angle value with when forefront indicia framing.
5. a kind of method of inspection for industrial structure defect intelligent recognition model according to claim 4, feature exist
In judging that identification frame is corrected errors and statistical magnitude according to following standard in step 4):
401) for each row element in matrix, if at least one value is more than or equal to registration threshold value, it is believed that the row pair
The default answering marks answered correctly are detected；
402) for each column element in matrix, if at least one value is more than or equal to registration threshold value, it is believed that the column pair
The identification label that the confidence level answered is higher than threshold value is correct, does not fall through, accordingly the correct detection quantity of statistics and accurate detection quantity.
6. a kind of method of inspection for industrial structure defect intelligent recognition model according to claim 5, feature exist
In for general task, registration threshold value takes 0.5, and for high standard task, registration threshold value takes 0.8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811237173.XA CN109584208B (en) | 2018-10-23 | 2018-10-23 | Inspection method for intelligent identification model of industrial structure defects |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811237173.XA CN109584208B (en) | 2018-10-23 | 2018-10-23 | Inspection method for intelligent identification model of industrial structure defects |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109584208A true CN109584208A (en) | 2019-04-05 |
| CN109584208B CN109584208B (en) | 2021-02-02 |
Family
ID=65920383
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811237173.XA Active CN109584208B (en) | 2018-10-23 | 2018-10-23 | Inspection method for intelligent identification model of industrial structure defects |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109584208B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110689538A (en) * | 2019-10-12 | 2020-01-14 | 太原科技大学 | Tunnel lining crack image detection method |
| CN111710412A (en) * | 2020-05-29 | 2020-09-25 | 北京百度网讯科技有限公司 | Diagnostic result checking method and device and electronic equipment |
| CN113034498A (en) * | 2021-04-28 | 2021-06-25 | 江苏欧密格光电科技股份有限公司 | LED lamp bead defect detection and assessment method and device, computer equipment and medium |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2016023946A (en) * | 2014-07-16 | 2016-02-08 | 株式会社デンソー | Target detector |
| CN105488478A (en) * | 2015-12-02 | 2016-04-13 | 深圳市商汤科技有限公司 | Face recognition system and method |
| CN108108887A (en) * | 2017-12-18 | 2018-06-01 | 广东广业开元科技有限公司 | A kind of Internet of Things based on multidimensional data is traveled out the intelligent evaluation model of row index |
| CN108460341A (en) * | 2018-02-05 | 2018-08-28 | 西安电子科技大学 | Remote sensing image object detection method based on integrated depth convolutional network |
| CN108681693A (en) * | 2018-04-12 | 2018-10-19 | 南昌大学 | Licence plate recognition method based on trusted area |
-
2018
- 2018-10-23 CN CN201811237173.XA patent/CN109584208B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2016023946A (en) * | 2014-07-16 | 2016-02-08 | 株式会社デンソー | Target detector |
| CN105488478A (en) * | 2015-12-02 | 2016-04-13 | 深圳市商汤科技有限公司 | Face recognition system and method |
| CN108108887A (en) * | 2017-12-18 | 2018-06-01 | 广东广业开元科技有限公司 | A kind of Internet of Things based on multidimensional data is traveled out the intelligent evaluation model of row index |
| CN108460341A (en) * | 2018-02-05 | 2018-08-28 | 西安电子科技大学 | Remote sensing image object detection method based on integrated depth convolutional network |
| CN108681693A (en) * | 2018-04-12 | 2018-10-19 | 南昌大学 | Licence plate recognition method based on trusted area |
Non-Patent Citations (2)
| Title |
|---|
| SEYYED HAMED NAGHAVI 等: "Integrated real-time object detection for self-driving vehicles", 《2017 10TH IRANIAN CONFERENCE ON MACHINE VISION AND IMAGE PROCESSING (MVIP)》 * |
| 李策 等: "一种高分辨率遥感图像视感知目标检测算法", 《西安交通大学学报》 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110689538A (en) * | 2019-10-12 | 2020-01-14 | 太原科技大学 | Tunnel lining crack image detection method |
| CN110689538B (en) * | 2019-10-12 | 2022-03-29 | 太原科技大学 | Tunnel lining crack image detection method |
| CN111710412A (en) * | 2020-05-29 | 2020-09-25 | 北京百度网讯科技有限公司 | Diagnostic result checking method and device and electronic equipment |
| CN113034498A (en) * | 2021-04-28 | 2021-06-25 | 江苏欧密格光电科技股份有限公司 | LED lamp bead defect detection and assessment method and device, computer equipment and medium |
| CN113034498B (en) * | 2021-04-28 | 2023-11-28 | 江苏欧密格光电科技股份有限公司 | LED lamp bead defect detection and assessment method, device, computer equipment and medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109584208B (en) | 2021-02-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108257114A (en) | A kind of transmission facility defect inspection method based on deep learning | |
| US20180253836A1 (en) | Method for automated detection of defects in cast wheel products | |
| CN110473173A (en) | A kind of defect inspection method based on deep learning semantic segmentation | |
| CN108734283B (en) | Neural network system | |
| CN110232379A (en) | A kind of vehicle attitude detection method and system | |
| CN109584208A (en) | A kind of method of inspection for industrial structure defect intelligent recognition model | |
| CN105787482A (en) | Specific target outline image segmentation method based on depth convolution neural network | |
| CN108898085A (en) | A kind of road disease intelligent detecting method based on mobile video | |
| CN105426870A (en) | Face key point positioning method and device | |
| CN106228161A (en) | A kind of pointer-type dial plate automatic reading method | |
| CN108802041B (en) | Method for rapidly changing small sample set of screen detection | |
| CN103745461B (en) | A kind of printing image defect detection method based on areas combine feature | |
| CN107507194A (en) | A kind of insulator chain fault detection method based on infrared image temperature distributing rule and BP neural network | |
| CN109241901B (en) | A kind of detection and recognition methods to the three-dimensional point cloud with hole | |
| CN111080622A (en) | Neural network training method, workpiece surface defect classification and detection method and device | |
| CN111401419A (en) | Improved RetinaNet-based employee dressing specification detection method | |
| CN111985499B (en) | High-precision bridge apparent disease identification method based on computer vision | |
| CN104990926A (en) | TR element locating and defect detecting method based on vision | |
| CN110490842A (en) | A kind of steel strip surface defect detection method based on deep learning | |
| CN107818321A (en) | A kind of watermark date recognition method for vehicle annual test | |
| CN112419299A (en) | Bolt loss detection method, device, equipment and storage medium | |
| CN105957059A (en) | Electronic component missing detection method and system | |
| CN107424150A (en) | A kind of road damage testing method and device based on convolutional neural networks | |
| CN104915678A (en) | Detection method and apparatus of target object in power transmission line | |
| CN105005798A (en) | Target recognition method based on collecting and matching local similar structure |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |