CN110111331A - Honeycomb paper core defect inspection method based on machine vision - Google Patents
Honeycomb paper core defect inspection method based on machine vision Download PDFInfo
- Publication number
- CN110111331A CN110111331A CN201910418796.5A CN201910418796A CN110111331A CN 110111331 A CN110111331 A CN 110111331A CN 201910418796 A CN201910418796 A CN 201910418796A CN 110111331 A CN110111331 A CN 110111331A
- Authority
- CN
- China
- Prior art keywords
- defect
- picture
- frame
- sample
- coordinate
- 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
- 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
- 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
- G01N21/95—Investigating the presence of flaws or contamination characterised by the material or shape of the object to be examined
- G01N21/9515—Objects of complex shape, e.g. examined with use of a surface follower device
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/51—Indexing; Data structures therefor; Storage structures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/587—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using geographical or spatial information, e.g. location
-
- 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
- G06T7/00—Image analysis
- G06T7/60—Analysis of geometric attributes
- G06T7/62—Analysis of geometric attributes of area, perimeter, diameter or volume
-
- 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
- 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
- G01N2021/8861—Determining coordinates of flaws
-
- 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
- G01N2021/8887—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 based on image processing techniques
-
- 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
- G06T2207/20081—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
- G06T2207/20112—Image segmentation details
- G06T2207/20164—Salient point detection; Corner detection
-
- 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
-
- 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)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Life Sciences & Earth Sciences (AREA)
- Pathology (AREA)
- Databases & Information Systems (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Immunology (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Chemical & Material Sciences (AREA)
- Analytical Chemistry (AREA)
- Biochemistry (AREA)
- Software Systems (AREA)
- Geometry (AREA)
- Library & Information Science (AREA)
- Quality & Reliability (AREA)
- Signal Processing (AREA)
- Image Processing (AREA)
- Investigating Materials By The Use Of Optical Means Adapted For Particular Applications (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of honeycomb paper core defect inspection method based on machine vision, for the various defect problems generated in honeycomb paper core production process, by the honeycomb paper core picture for acquiring production scene, using the defects of SSD deep neural network detection honeycomb paper core, its defect classification is determined and exports its specific location, then quickly rechecking is carried out with machine vision algorithm, prevent erroneous detection, acquired results are finally passed into honeycomb paper core defect mending system, correct feedback signal is provided, to realize the auto-mending to honeycomb paper core defect.The present invention is measured in real time honeycomb paper core defect with deep learning model and machine vision algorithm, feedback information can be provided for honeycomb paper core production process automatic defect patch system, have the advantages that accurate identification, accurate positioning and recognition speed are fast, the requirement of honeycomb cardboard production automation can be met.
Description
Technical field
The present invention relates to field of image detection, especially a kind of honeycomb paper core defect inspection method based on machine vision.
Background technique
Honeycomb paper is made according to nature honeycomb structure principle, it is the body paper being stacked for being cut into strip
Numerous hollow three-dimensional regular hexagon is connected into gluing knot method, forms the stressed member of an entirety, and bond on its two sides
A kind of environmental protection and energy saving material of new midsole structure made of facial tissue.With very high mechanical strength, it is able to take handling process
In various collisions and falling fall, be chiefly used in the packaging and transport of various accurate devices, fragile device or even military project device, have very strong
Industrial applicibility.And in honeycomb cardboard production process, because of the industrial fact of the improper and various complexity stretched, hold very much
It is also easy to produce defect, such as the cavity of large area, irregular empty structure, continuous fracture etc., this affects honeycomb cardboard receiving and touches
The ability hit is the problem of not can bypass in honeycomb cardboard production process.On existing production line, artificial detection is often used
Defect and the method repaired fill up the defect generated in real time in production process, need to expend a large amount of human resources.Separately
On the one hand, in the case where long-term high load capacity works, artificial detection defect is easy to produce fault, causes quality to decline, defect rate
The various problems such as raising.
Summary of the invention
The technical problem to be solved by the present invention is in view of the shortcomings of the prior art, provide a kind of bee based on machine vision
Nest core defect inspection method, real-time detection honeycomb paper core defect produced on the production line enhance the robust performance of system, prevention
The erroneous detection problem being likely to occur.
In order to solve the above technical problems, the technical scheme adopted by the invention is that: a kind of honeycomb paper based on machine vision
Core defect inspection method, includes the following steps:
1) production equipment is transferred in honeycomb paper core industrial site and generate great amount of samples, build vision platform and it is carried out
Acquisition, image capture position are production line exit, the intake historical information few as far as possible for only acquisition exit real time information,
Acquisition image be set as strip (and acquire on the width with the size of production line equivalent width, it is small as far as possible in length).By bee
Nest core strip image is equidistantly divided, and it is [300,300] that gained, which divides picture size, carries out Artificial Cognition to every piece of picture,
Judge that it, with the presence or absence of defect (may a width figure in there are multiple defects), filters out the picture containing defect, to defective locations into
Rower is fixed, saves the centre coordinate of lower calibration frame and the height and width of framework, and judge the classification information of the position defect, whole
The true value that above- mentioned information are closed as picture marks (ground truth), and building database 1 is used as tranining database, by above-mentioned letter
Breath deposit is wherein.It establishes vision operation sample database simultaneously with spare, includes zero defect sample table and typical defect sample table, wherein
Defect sample has also carried out handmarking.
2) data extending is carried out for data in database 1, concrete operations include interception, rotation, adjustment luminance saturation
Deng, will newly generated sample be stored in database 1 in.
3) SSD (Single Shot MultiBox Detector) network is constructed, with picture in database 1 and its right
Answer ground turth (label and bounding box location information) training pattern.Suitable priori frame number is selected, in SSD
Suitable convolutional layer is added after basic network to generate the characteristic layer of different scale, and according to the general shape of defect in each ruler
The priori frame of multiple aspect ratios is chosen on degree, constitutes priori frame set.The matching of priori frame is carried out when training first, setting is handed over and compared
(IoU) threshold value (being set as 0.5 herein) judges priori frame whether the ground truth good with image calibration in database matches,
Then it is judged as positive example more than the threshold value, is otherwise counter-example, the picture of counter-example is directly identified as background, and every width figure can obtain largely just
Negative data, but counter-example number often substantially exceeds the positive example number (which results in the imbalances of data set) of successful match, therefore
Operational difficulties negative sample excavates (hard negative mining) algorithm and is sampled to negative sample, lower negative sample specific gravity.
The loss function of network is such as given a definition:
Wherein x is input, and c is confidence level predicted value, and l is first
The position prediction value on the correspondence boundary of frame is tested, and g is the location parameter of ground truth.Its LconfThe confidence generated for classification
Spend error, LlocCalculation formula for location error, two errors is as follows:
Confidence level error:
Wherein first item is positive sample cross entropy, and Section 2 negative sample is directly calculated as 0 probability.
Location error:
Wherein use smooth L1 costing bio disturbance position gap, in order that make loss coordinate origin accessory change more
Smoothly.Wherein bounding box coding function are as follows:
Subscript x, y, w, h respectively represent the centre coordinate and Width x Height of bounding box,For bounding box
Predicted value,To correspond to bounding box actual position,For priori frame position.
Optimize loss function gradient descent procedures with optimization method, constantly reduces its comprehensive loss until its satisfaction is wanted
It asks.Its Model Weight is preserved.
4) same vision platform in step 1 is built, production line advance step is adjusted, controls its each advance step and adopt
Collection image at (i.e. between each group of picture non-overlapping and connect) in the same size of production line length direction, adopt in real time by each step
Collection image is analyzed, and every group of image is carried out cutting operation same as step 1, the picture being partitioned into is saved in database
2 are used as production scene volatile data base, while writing down the location information of every width picture in the production line, and the database root is according to production
Actual conditions carry out real-time update.Coordinate system is established according to production line actual size, coordinate origin is production line entrance one end, with
Production line width direction establishes x-axis, establishes y-axis with length direction, set a scale just include a width segmentation picture, i.e. x,
Mono- scale of y includes 300 pixel sizes, is defects detection final detection result and image coordinate result COMPREHENSIVE CALCULATING,
Obtain defect more specific location information.The formula of location position is carried out for x-axis variation to defect:I=1,
2...n；J={ upper left, lower-left, upper right, bottom right }；Wherein i is required sample, and j is defects detection frame vertex position,
For the x coordinate on i-th of vertex sample defects detection frame j, x is samples pictures relative to the relative coordinate on the direction x of production line
Size, αxIt is herein 300 for every width segmented image x-axis direction absolute size, gainedLine coordinates is produced for defect frame is opposite
It is absolute x coordinate.Y-axis variation:I=1,2...n；J={ upper left, lower-left, upper right, bottom right }；Wherein i
For required sample, j is defects detection frame vertex position,For the y-coordinate on i-th of vertex sample defects detection frame j, y
It is samples pictures relative to the relative coordinate size on the direction y of production line, αyIt is absolutely big for every width segmented image y-axis direction
It is small, it is herein 300, gainedIt is defect frame with respect to the absolute y-coordinate of production line coordinate system.
5) the SSD detection network model that reconstruct is kept, the defect picture location information provided according to step 7 is from database
Picture is taken out in 3, is input to SSD detection network model and is tested and analyzed, forward process first checks classification confidence level, filters out
It is determined as background priori frame, then IoU threshold value is set and filters off below standard positive example priori frame, finally with smooth non-maximum suppression algorithm
(Soft-NMS) overlapping frame is got rid of, the defect for exporting final defective locations and the position is classified, by network output defect inspection
4 apex coordinates that frame location information is converted into framework are surveyed, are stored in database 2 in real time, and carry out according to its defective locations information
Further integration: situation defective for picture edge checks its neighboring picture with the presence or absence of defect, for being cut by cutting
Disconnected defect carries out the fusion completion of defect, and defect merges completion algorithm are as follows:
Repairing up and down: setting picture i defect and be located at below picture, and picture i+1 is defective to be located above picture, above and below two pictures
It is adjacent to connect, judge in picture iWithWhether α is equal toyIf condition meets, search pictures i+1 checks no presence
Defect frame, there is no then without carrying out defect frame fusion, exist, judge in picture i+1WithWhether 0 is equal to,
And judgeWhether set up simultaneously, if above-mentioned condition is all satisfied, to defect frame into
Row fusion, the new defects detection frame of fusion gained are WhereinY-coordinate or x coordinate for picture i defect frame in corresponding vertex position, αyFor every width point
Image y-axis direction absolute size is cut, is herein 300,Exist for picture i+1 defect frame
The y-coordinate or x coordinate of corresponding vertex position,For defect in picture i or picture i+1
The absolute xy coordinate set on frame vertex.
Left and right repairing: picture j defect is located at picture right, and picture j+1 is defective to be located at picture left, two pictures or so phase
Neighbour connects, and judges in picture jWithWhether α is equal toxIf condition meets, search pictures j+1 checks no presence
Defect frame, there is no then without carrying out defect frame fusion, exist, judge in picture j+1WithWhether 0 is equal to,
And judgeWhether set up simultaneously, if above-mentioned condition is all satisfied, to defect frame into
Row fusion, the new defects detection frame of fusion gained are WhereinY-coordinate or x coordinate for picture j defect frame in corresponding vertex position, αxFor every width point
Image x-axis direction absolute size is cut, is herein 300,For picture j+1 defect frame
Y-coordinate or x coordinate in corresponding vertex position,To be lacked in picture j or picture j+1
Fall into the absolute xy coordinate set on frame vertex.
6) SSD network query function application of results machine vision algorithm is rechecked:
Carry out off-line operation to vision operation sample database: operation obtains reinspection feature backed-up value and is directed to typical defect sample
Image is become gray level image, carries out histogram manipulation by table, and it is stand-by that the histogram of acquisition is transcoded into array deposit database；
For zero defect sample table: gray processing, then OtsU algorithm operation is carried out to it, seek binarization threshold, according to obtained by this algorithm
Threshold value carries out binaryzation operation, carries out further closed operation to binary image.It synchronizes and is handled as follows: on the one hand refining
Gained closed operation image, and Corner Detection (angle point takes two and two intersection points with top edge at this) is carried out, it obtains in image
Angle point number obtains angle point number contained by unit number of contours, in table with angle point number divided by the core polygonal profile number of institute's altimetric image
All pictures do all same processing as above, angle point number contained by final averaging of income unit number of contours are as follows:And it is spare to seek per unit number of contours simultaneouslyWherein biIt is total for sample angle point,
liFor sample profile sum, siFor the sample gross area, N is sample size；On the other hand, each core polygonal wheel is directly calculated
Wide area and perimeter, and the mean value of the average single contour area perimeter of its whole picture is sought, all pictures carry out in sample database
It states operation and finally obtains its COMMON MEAN:Wherein N is sample size, and M is each
Corresponding number of contours, s in a sampleijFor the area of i sample jth profile, cijFor the perimeter of i sample jth profile.It is all offline
Operation acquired results are stored in stand-by in memory.
Online reinspection operation: following processing: gray processing is done to image in gained honeycomb paper core defect frame region, carries out histogram
Figure operation compared with carrying out one by one with the histogram of corresponding defect in typical defect sample table, calculates its maximum Pasteur's distance:Wherein Hi(I) ash in i-th of typical defect sample histogram is indicated
The pixel number that grade is I is spent, H (I) is to recheck the pixel number that gray level is I in histogram,N
For the total number of pixels counted in histogram.Defect picture is carried out such as identical pretreatment in off-line operation, calculating online
Angle point number, acquired results are denoted as BtestIf gained defect frame size is Stest, take the mean unit number of contours kept before
Contained angle point number B and per unit number of contours L, seeks angle point number relative difference:It is average to calculate it
The mean value S of single contour area perimetertest、Ctest, off-line calculation averaging of income list contour area and perimeter S and C are taken, face is calculated
Product, perimeter relative differenceIt is rechecked because having outlined defective locations, therefore framework is smaller,
Line arithmetic speed is exceedingly fast.
Required variable calculates from weighting on-line operation result:Obtain Comprehensive Assessment
Index δ, wherein α, β, γ, ε are calculating parameter, dmax、It is derived from online reinspection operation acquired results.The bigger expression inspection of δ
Mapping piece and zero defect sample difference are bigger, are that the confidence level of defect picture is higher.Many experiments calculate zero defect and defective
The value difference is away from selection overall merit threshold value, is not determined as erroneous detection up to threshold value as the case may be, deletes interim produce under situation
The defect pictorial information saved in database.
7) mode with patch system cooperation: vision platform is located at production line arrival end, and patch system is located at production line tail
End saves data with production line volatile data base, and the corresponding part of obtained picture containing defect to be detected is along production line advance
Repair operation is carried out again in into the patch system repairing visual field, that is, carries out the picture containing defect of history while detection in the other end
The patch work of corresponding part, the image data after repairing treatment are put into that the backing sheet deleted in real time is spare, and main table is according to coordinate
It is that (the real-time coordinates information of picture preservation carries out real-time increasing certainly according to the stepping of production line, exceeds and sits for increasing update at any time certainly for stepping
Mark fastens limit deposit backing sheet, deletes from main table).Database defect frame increases rule certainly:WhereinFor
Defect frame is with respect to the absolute y-coordinate of production line coordinate system, αyFor every width segmented image y-axis direction absolute size, it is for 300, n herein
The stepping number experienced from after entering production line, stepping simultaneously have not been changed defect frame with respect to the absolute x coordinate of production line coordinate system.
Compared with prior art, the advantageous effect of present invention is that:
1. realizing honeycomb paper core defect with machine vision technique to detect automatically, the class of defect is provided for patch system
Other information and accurate location information provide algorithm condition for industry honeycomb paper core production system automatic defect repairing work.
2.SSD algorithm has training convenient in detection algorithm, detects accurately to height, detects fireballing advantage.This hair
The bright honeycomb paper core for adjusting rear existing defects to stretched condition with the network carries out defects detection, orients different classes of
Defect provides defect information to the patch system of next stage, and network positions and classification are accurate, and speed is exceedingly fast, and meet real-time
The requirement of detection.Particularly, assembly line cardboard is cut into fritter and handled by the present invention, will not direct plunge into fortune integrally
It calculates, but establishes coordinate system, pass through the exhausted of picture production line coordinates and SSD network output position information comprehensive judgement defect relatively
To coordinate.It does so and efficiently reduces the influence that SSD network detects upper ineffective problem in small-sized object, prevent defect
Missing inspection problem.
3. rechecking with machine vision algorithm to deep learning object detection results, traditional algorithm is efficiently utilized
More stable advantage is used in industry spot, and realizes high speed in conjunction with honeycomb paper core concrete property and rechecks.It both utilizes in this way
The advantages of both sides, and the shortcomings that evaded both sides, it efficiently solves the problems, such as the erroneous detection of defect, further improves industrial production
Stability and product certified product rate.
Detailed description of the invention
Fig. 1 is the method for the present invention flow chart；
Fig. 2 is system structure of the invention figure；
Fig. 3 is SSD defects detection network；
Fig. 4 is effect picture after vision operation reinspection pretreatment.
Specific embodiment
1. transferring production equipment in honeycomb paper core industrial site generates a large amount of honeycomb paper core samples, taken in industry spot
Vision platform is built, camera is fixed, and gives a certain amount of constant daylighting, acquisition is dimensioned to strip picture, is schemed
Piece width 2400, height 300 adjust height, camera width direction are made just to take in the overall width of production line, i.e. intake paper
The full width information of tape entry position, picture height part take 300 as followed by analysis when production line one before
The linear module that progress is adjusted.The strip picture being collected equidistantly is divided and (is divided into 8 pieces), that is, generate one group [300,
300] image manually checks and whether there is defect in picture, and defect then uses labelImg to carry out defective locations if it exists
The picture demarcated is stored in sql server database, and leaves the defect classification of the picture by mark, and defect classification includes small
Type cavity, large-scale cavity, irregular notch is continuous to be broken, and pays attention to retention class background, uses so that counter-example marks, this database
Entitled tranining database abandons other pictures.Pay attention to needing to collect the sample of all situations when acquisition, and guarantees various lack
Sunken number of samples, which all reaches certain value, can just stop, and for defective locations, except figure center situation is located at, also because of collecting part
The defect separated is demarcated, so that it can also detected for being just divided the defect of operation partition, reinforces model
Robust performance adapts to various different situation.
2. carrying out following every kind of data extending simultaneously for the picture in tranining database with 50% variation probability: random
It cuts out, cutting out length-width ratio is [3/4,4/3], and stochastical sampling area is [20%, 100%]；To data set with 50% probability into
Row overturning；Scale tone, saturation degree, brightness, zoom factor uniform sampling from [0.6,1.4].By the figure after data extending
Piece is stored in corresponding database.
3. building SSD network under tensorflow frame, feature is carried out to picture as basic network with VGG network
It extracts, and is loaded into its pre-training model on ILSVRC CLS-LOC data set.Respectively by the full articulamentum fc6 of VGG16 and
Fc7 is converted into 3*3 convolution conv6 and 1*1 convolution conv7, and tetra- pieces of convolutional layers of conv8-11 are added after basic network and (contain 1*
1 and 3*3, two kinds of convolutional layers), conv7, conv8_2, conv9_2 are extracted, in conv10_2, conv11_2 and basic network
Conv4_3 has six convolutional layers altogether as characteristic pattern, extracts priori frame according to pixel number on each characteristic pattern, then to every
The priori frame generated on a characteristic pattern carries out length-width ratio variation, and the ratio of generation is long four kinds of block diagrams than wide [2,1/2,3,1/3]
In addition original priori frame, generates priori frame collection.When training, progress priori frame matching first, for the ground of investment
Truth, calculates whether all priori frames match with ground truth, and it is more than threshold value that it is 0.5 that setting IoU, which (hands over and compare) threshold value,
Priori frame be positive example, the priori frame for being less than threshold value is counter-example, and counter-example is labeled as the context marker that retains of front, by acquisition
Positive negative data deposit memory is stand-by.Due to being not up to the priori frame of threshold value often far more than the priori frame for reaching threshold value, this meeting
It is unbalanced to lead to the problem of data set, therefore uses Hard Negative Mining technology, negative sample is sampled, when sampling
Descending arrangement is carried out according to confidence level error (confidence level of projected background is smaller, and error is bigger), chooses the biggish of error
Negative sample of the top-k as training, positive and negative sample proportion is controlled in 1:3.Iterate training, and wherein gradient optimizing method selects
Select momentum optimization method: its gradient updating formula are as follows:Wherein vtDecline step for the subgradient,
vt-1Declining step for previous subgradient, γ is attenuation rate,For the gradient direction of this decline, α is learning rate, is set as referring to
(0.005 is initial learning rate to number decaying learning rate, and every 5 epoch decline 5%), until test set feed forward process final result
IoU reaches 70% between true groud truth, and classification accuracy reaches 98%, optimal weight information under preservation.
4. build same vision platform in 1, control production line every time in advance step and 1 unanimously, i.e., each stepping 300
A pixel, carries out Image Acquisition and analysis between matched each stepping, and one group of picture of acquisition carries out it as identical in 1
Dividing processing, using the picture of acquisition deposit SQL server database as production scene volatile data base, and be stored in every width
Picture corresponds to the location information of production line.Coordinate system, the corresponding production line width side of x-axis are established by origin of production line entrance one end
To y-axis corresponds to production line direction of advance, is that represent every width picture opposite in reference axis for a unit with 300 pixels
Location information.The database root is according to whether there is or not finishing last repairing treatment to be updated data, i.e., reference axis y-axis direction is established and arrived
Position more two positions (surplus prevents from losing information) again are repaired, storing data is beyond the y-axis upper limit then from database in database
Middle deletion.
5. the SSD defects detection network built in reconstruct 4 is loaded into the obtained weight information of training in 4, will adopt in 5
The picture of collection inputs network 4 in real time, carries out real-time monitoring to honeycomb paper core defect, then exports progress to it for defective picture
It saves, is stored in production scene volatile data base, for zero defect picture, network is then without the whole whole quilts of corresponding defect classification output
It is detected as background image, so without further operating.The defect being partitioned from defect blending algorithm completion updates
Production line volatile data base, entire detection process are step with production line stepping, and it is interim that real-time detection goes out result deposit production line
Database.It is wherein specific further include:
1) it defects detection network feed forward process: is exported according to softmax layers of output of classification and determines its classification, and filter out category
In the prediction block of background.Then below standard positive example frame is filtered off according to the rule that IoU threshold value is more than 0.65, then with smooth non-maximum
Restrainable algorithms (Soft-NMS) carry out selection deletion to the priori frame of the overlapping chosen, the final remaining elder generation of iteration optimization
Frame is tested, the score of Soft-NMS resets function are as follows:Wherein M is that high score is first
Frame is tested, b is wait select overlapping frame, when the two IoU is more than threshold value Nt, just reduce and choose score si.Best obtain is chosen in operation repeatedly
The priori frame divided is as final result.
2) the 4 apex coordinates deposit production scene for converting framework for network output defects detection frame location information is interim
Database is stand-by, its conversion formula by taking the x coordinate of upper left as an example are as follows:WhereinFor the defect frame upper left corner
Vertex x coordinate, dcxDefect frame center x coordinate, d are exported for networkcwDefect width of frame size is exported for network.
6. reinspection process: establish sql server vision operation sample database, according to the same manner in 1 acquire 30 it is intact
Sample image is fallen into, segmentation size is similarly [300,300], is stored in zero defect sample table, chooses from training sample database various types of
Each 5, typical defect sample, the wherein region bounding box picture is extracted, is stored in typical defect sample table, is carried out
The off-line operation as described in summary of the invention: to defective picture, the intensity histogram of calculating picture after gray processing is carried out to sample
It is stand-by to convert the histogram to array data deposit typical defect sample table corresponding position for figure.For zero defect picture, by picture
After gray processing, its binarization threshold is calculated with OtsU adaptive thresholding algorithm, binaryzation is carried out to picture with gained threshold value,
And closed operation processing is carried out to picture, and wherein disconnected crosspoint is eliminated, keeps cellular cavity continuous and complete, final process effect
The visible attached drawing 3 of fruit.Open two lines journey and carry out multithreading operation: thread 1 refines image, extracts the skeleton of image, fortune
Its angle point number (two and two intersection points with top edge) is counted with the skeleton image after refinement, while being counted in picture
Profile number obtains angle point number under sample unit's profile divided by number of contours with angle point number, to all intact in sample database
Sunken sample all carries out such operation, and it is standby that angle point number deposit memory under zero defect image averaging list number of contours is obtained after average
With, while unit of account area bottom profiled number (equally carry out and be averaged on all samples again), deposit memory are spare；Line
Journey 2 directly calculates the area and perimeter of each core polygonal profile in image after closed operation, and asks its whole picture average single
Data all in table are equally all done once-through operation by the mean value of contour area perimeter, its average result of statistical calculation is saved to interior
In depositing.Image is calculated as above in the detection frame region that line computation hour hands export SSD network, calculates ash first
Histogram is spent, corresponding set of histograms (saving histogram in memory in reconstruct typical defect table) is selected according to output defect classification
It is continuously compared therewith, calculates its Pasteur's distance, obtain the maximum value of 5 calculating.Then according to defect frame areal calculation its
The online result obtained under due profile average area perimeter and angle point number, with same algorithm in the case of corresponding zero defect
Subtract each other and seek absolute value, and there should be value to make comparisons with obtained by off-line operation, obtains its relative difference, 4 data are finally obtained after integration:
dmax、Correspond to maximum histogram Pasteur distance, average angle point number difference, average area difference, average perimeter
Difference.For this 4 value generally between [0,1], a small amount of value is slightly larger than 1, is not required to be normalized, therefore is added acquisition synthesis and comments
Valence index:Wherein α, β, γ, ε parameter are respectively set to 2,2,0.5,1.It carries out at the scene more
Secondary experiment, frame take lot of examples frame as sample (containing various defects and without each half of defect), calculate two kinds of different samples institutes
δ difference size, acquisition overall merit threshold value 2.5 are obtained, it is more than that threshold value 2.5 then thinks that its detects standard that input defect block diagram piece, which calculates δ,
Really, erroneous detection does not occur.
7. patch system traverses production line volatile data base in real time, check that is wherein saved (repairs view in end-position
In open country) picture group data and its testing result, carry out patch work, for complete patch work picture group remained into
It in the backing sheet separately built in database, is deleted from main table, backing sheet is more than that 5 groups of historical datas are updated with regard to delete in real time.
The embodiment of the present invention carries out honeycomb paper core defect by establishing deep learning SSD network integration machine vision algorithm
Real-time detection, matching defect repairing equipment carry out automation repairing, and accurate positioning is quick, helps to improve the production of honeycomb paper core
Quality.
Claims (6)
1. a kind of honeycomb paper core defect inspection method based on machine vision, which comprises the following steps:
1) the honeycomb paper core picture sample containing defect is obtained, and is equidistantly divided according to the ratio of width to height, the picture after screening segmentation
Collection saves the picture containing defect, classifies to defect therein, and carries out location position to defect, establishes honeycomb paper core and lacks
Detection tranining database is fallen into, different degrees of data extending is carried out to the tranining database and is handled；
2) SSD target detection model is established, model training optimization is carried out with the sample in tranining database, after saving optimization
Weight information obtains honeycomb paper core defects detection network model；
3) guarantee image acquisition process is reached synchronous with production line stepping, and every stepping once carries out a picture collection and in stepping
Interval carries out the analysis of this acquisition picture, and gained honeycomb paper core picture is bar chart；Equidistant segmentation bar chart, it is existing to establish production
Field volatile data base, it is interim to save gained picture, establish production line coordinate system, recordable picture relative position, according to production procedure
Real-time update is carried out to Picture Coordinate；
4) picture for obtaining step 3) successively puts into the honeycomb paper core defects detection network model established in step 2), to figure
The defects of piece is positioned and is classified, and gained defect information is stored in production volatile data base and is temporarily saved, in real time
The defects detection frame for being divided operation truncation is checked for, fusion completion is carried out to the defect being truncated；
5) defect is rechecked, establishes vision operation sample database, with machine vision algorithm, off-line calculation sample database picture
Area, perimeter, histogram and corner feature obtain comprehensive judgment threshold；In the inspection of line computation SSD target detection model output
Four features for surveying defect picture in frame, are compared with off-line calculation threshold value obtained, judge whether there is erroneous detection situation,
Delete the location information of erroneous detection defect in the volatile data base of production scene；
6) database information is passed into tail end defect repair unit, to carry out real-time defect mending operation.
2. the honeycomb paper core defect inspection method according to claim 1 based on machine vision, which is characterized in that establish life
Producing line coordinate system, establishes x-axis along production line width direction, establishes y-axis along production line length direction, sets a scale and just wraps
Divide picture containing a width, the formula of location position carried out to defect are as follows: x-axis variation:I=1,2...n；j
={ upper left, lower-left, upper right, bottom right }；Wherein i is required sample, and j is defects detection frame vertex position,It is i-th
The x coordinate on the vertex sample defects detection frame j, x are samples pictures relative to the relative coordinate size on the direction x of production line, αx
For every width segmented image x-axis direction absolute size, gainedIt is defect frame with respect to the absolute x coordinate of production line coordinate system；Y-axis becomes
Change:I=1,2...n；J={ upper left, lower-left, upper right, bottom right }；Wherein i is required sample, and j is defect
Detection block vertex position,For the y-coordinate on i-th of vertex sample defects detection frame j, y is samples pictures relative to production
Relative coordinate size on the direction y of line, αyFor every width segmented image y-axis direction absolute size, gainedIt is opposite for defect frame
The absolute y-coordinate of production line coordinate system.
3. the honeycomb paper core defect inspection method according to claim 1 based on machine vision, which is characterized in that defect
The specific implementation process rechecked includes:
1) establish vision operation sample database, it includes storage table are as follows: zero defect sample table, typical defect sample table；
2) gray level image is become to image in typical defect sample table, carries out histogram manipulation, the histogram of acquisition is transcoded into
It is stand-by that array is stored in database；
3) to image gray processing in zero defect sample table, then OtsU algorithm operation is carried out to it, seek binarization threshold, according to this
Threshold value obtained by algorithm carries out binaryzation operation, carries out further closed operation to binary image；It synchronizes and is handled as follows: one
Aspect refinement gained closed operation image, and Corner Detection is carried out, image interior angle points are obtained, with angle point number divided by institute's altimetric image
Core polygonal profile number obtains angle point number contained by unit number of contours, does all same processing as above to pictures all in table,
Angle point number contained by final averaging of income unit number of contours are as follows:And it is standby to seek per unit number of contours simultaneously
WithWherein biFor sample angle point sum, liFor sample profile sum, siFor the sample gross area, N is sample size；
On the other hand, the area and perimeter of each core polygonal profile are directly calculated, and seeks the average single contour area of its whole picture
The mean value of perimeter, all pictures carry out above-mentioned operation and finally obtain its COMMON MEAN in sample database:Wherein N is sample size, and M is corresponding number of contours in each sample, sijFor
The area of i sample jth profile, cijFor the perimeter of i sample jth profile；
4) following processing: gray processing is done to image in gained honeycomb paper core defect frame region, carries out histogram manipulation, lacked with typical case
The histogram for falling into corresponding defect in sample table is compared one by one, calculates its maximum Pasteur's distance:Wherein Hi(I) gray scale in i-th of typical defect sample histogram is indicated
Grade is the pixel number of I, and H (I) is to recheck the pixel number that gray level is I in histogram,N is
The total number of pixels counted in histogram；Defect picture is carried out online to calculate angle point number, institute such as identical pretreatment in 3)
It obtains result and is denoted as BtestIf gained defect frame size is Stest, take angle point number contained by the mean unit number of contours kept before
B and per unit number of contours L, seeks angle point number relative difference:Calculate its average single contour area
The mean value S of perimetertest、Ctest, take off-line calculation averaging of income list contour area and perimeter S and C, reference area, perimeter be opposite
Difference
5) formula is utilizedComprehensive evaluation index δ is obtained, wherein α, β, γ, ε are calculating parameter；δ
Bigger expression confidence level of the picture containing defect is higher；Many experiments calculate the gap of δ under zero defect and defective situation, select comprehensive
Evaluation threshold is closed, δ does not reach threshold value and is determined as erroneous detection, deletes the defect pictorial information saved in interim Production database.
4. the honeycomb paper core defect inspection method according to claim 1 based on machine vision, which is characterized in that be truncated
The specific implementation process of the fusion completion of defect includes:
Repairing up and down: setting picture i defect and be located at below picture, and picture i+1 is defective to be located above picture, and two pictures are neighbouring
Connect, judges in picture iWithWhether α is equal toyIf be equal to, search pictures i+1 checks no existing defects frame,
There is no then without carrying out defect frame fusion, exist, judge in picture i+1WithWhether it is equal to 0, and sentences
It is disconnectedWhether set up, if above-mentioned condition is all satisfied, defect frame is melted simultaneously
It closes, the new defects detection frame of fusion gained is WhereinY-coordinate or x coordinate for picture i defect frame in corresponding vertex position, αyFor the segmentation of every width
Image y-axis direction absolute size,It is picture i+1 defect frame in corresponding vertex position
Y-coordinate or x coordinate,For in picture i or picture i+1 defect frame vertex it is absolute
Xy coordinate set；
Left and right repairing: picture j defect is located at picture right, and picture j+1 is defective to be located at picture left, the adjacent phase of two pictures or so
It connects, judges in picture jWithWhether α is equal toxIf condition meets, search pictures j+1 checks no existing defects
Frame, there is no then without carrying out defect frame fusion, exist, judge in picture j+1WithWhether equal to 0, and
JudgementWhether set up, if above-mentioned condition is all satisfied, defect frame is melted simultaneously
It closes, the new defects detection frame of fusion gained is WhereinY-coordinate or x coordinate for picture j defect frame in corresponding vertex position, αxFor the segmentation of every width
Image x-axis direction absolute size,It is picture j+1 defect frame in corresponding vertex position
Y-coordinate or x coordinate,For in picture j or picture j+1 defect frame vertex it is absolute
Xy coordinate set.
5. the honeycomb paper core defect inspection method according to claim 1 based on machine vision, which is characterized in that defect is repaired
The specific implementation process of benefit includes: setting vision platform in production line arrival end, repairs unit and is located at production line tail end, with life
Producing line volatile data base saves data, and the corresponding part of obtained picture containing defect to be detected advances into repairing along production line and is
System repairing carries out repair operation in the visual field again, that is, carries out the corresponding part of picture containing defect of history while detection in the other end
Patch work, it is spare that the image data after repairing treatment is put into the backing sheet deleted in real time, and main table is according to coordinate system stepping from increasing
It updates at any time, i.e., the real-time coordinates information that picture saves carries out exceeding the coordinate system upper limit from increasing in real time according to the stepping of production line
It is stored in backing sheet, is deleted from main table.
6. the honeycomb paper core defect inspection method according to claim 5 based on machine vision, which is characterized in that picture is protected
The real-time coordinates information deposited carries out in real time according to the stepping of production line from the rule increased are as follows:WhereinFor
Defect frame is with respect to the absolute y-coordinate of production line coordinate system, αyFor every width segmented image y-axis direction absolute size, n is to enter production certainly
Stepping number experienced after line, stepping simultaneously have not been changed defect frame with respect to the absolute x coordinate of production line coordinate system.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910418796.5A CN110111331B (en) | 2019-05-20 | 2019-05-20 | Honeycomb paper core defect detection method based on machine vision |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910418796.5A CN110111331B (en) | 2019-05-20 | 2019-05-20 | Honeycomb paper core defect detection method based on machine vision |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110111331A true CN110111331A (en) | 2019-08-09 |
| CN110111331B CN110111331B (en) | 2023-06-06 |
Family
ID=67491182
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910418796.5A Active CN110111331B (en) | 2019-05-20 | 2019-05-20 | Honeycomb paper core defect detection method based on machine vision |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110111331B (en) |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110910372A (en) * | 2019-11-23 | 2020-03-24 | 郑州智利信信息技术有限公司 | Deep convolutional neural network-based uniform light plate defect detection method |
| CN111144475A (en) * | 2019-12-22 | 2020-05-12 | 上海眼控科技股份有限公司 | Method and device for determining car seat, electronic equipment and readable storage medium |
| CN111179241A (en) * | 2019-12-25 | 2020-05-19 | 成都数之联科技有限公司 | Panel defect detection and classification method and system |
| CN111487250A (en) * | 2020-04-02 | 2020-08-04 | 苏州奥创智能科技有限公司 | Intelligent visual detection method and system applied to injection molding defective product detection |
| CN111583238A (en) * | 2020-05-09 | 2020-08-25 | 中南大学 | Method for extracting vertex of included angle of contour line of honeycomb, method for detecting vertex of included angle of contour line of honeycomb and device for detecting vertex of included angle of contour line of honeycomb |
| CN111627020A (en) * | 2020-06-03 | 2020-09-04 | 山东贝特建筑项目管理咨询有限公司 | Detection method and system for anchor bolt in heat insulation board and computer storage medium |
| CN111808367A (en) * | 2020-07-09 | 2020-10-23 | 浙江七色鹿色母粒有限公司 | Improvement method for plastic PPR silver grain whitening defect |
| CN112232215A (en) * | 2020-10-16 | 2021-01-15 | 哈尔滨市科佳通用机电股份有限公司 | Railway wagon coupler yoke key joist falling fault detection method |
| CN112329896A (en) * | 2021-01-05 | 2021-02-05 | 武汉精测电子集团股份有限公司 | Model training method and device |
| CN112485260A (en) * | 2020-11-26 | 2021-03-12 | 常州微亿智造科技有限公司 | Workpiece defect detection method and device |
| CN112991344A (en) * | 2021-05-11 | 2021-06-18 | 苏州天准科技股份有限公司 | Detection method, storage medium and detection system based on deep transfer learning |
| CN113011057A (en) * | 2021-02-22 | 2021-06-22 | 河南农业大学 | Method and system for predicting performance of aged bonding structure based on gradient degradation of adhesive layer |
| WO2021227288A1 (en) * | 2020-05-09 | 2021-11-18 | 中南大学 | Method for evaluating geometrical morphology of honeycomb product, and detection and evaluation system |
| CN114429433A (en) * | 2022-04-01 | 2022-05-03 | 武汉市宏伟纸箱包装有限公司 | Honeycomb paper feed amount control method based on image enhancement |
| CN114782286A (en) * | 2022-06-21 | 2022-07-22 | 奥蒂玛光学科技（深圳）有限公司 | Defect repairing method, optical repairing device, electronic device and storage medium |
| CN115063725A (en) * | 2022-06-23 | 2022-09-16 | 中国民航大学 | Airplane skin defect identification system based on multi-scale self-adaptive SSD algorithm |
| CN115546203A (en) * | 2022-11-23 | 2022-12-30 | 常熟理工学院 | Production monitoring and analyzing method based on image data algorithm |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2018000731A1 (en) * | 2016-06-28 | 2018-01-04 | 华南理工大学 | Method for automatically detecting curved surface defect and device thereof |
| CN107833220A (en) * | 2017-11-28 | 2018-03-23 | 河海大学常州校区 | Fabric defect detection method based on depth convolutional neural networks and vision significance |
| CN108257122A (en) * | 2018-01-09 | 2018-07-06 | 北京百度网讯科技有限公司 | Paper sheet defect detection method, device and server based on machine vision |
-
2019
- 2019-05-20 CN CN201910418796.5A patent/CN110111331B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2018000731A1 (en) * | 2016-06-28 | 2018-01-04 | 华南理工大学 | Method for automatically detecting curved surface defect and device thereof |
| CN107833220A (en) * | 2017-11-28 | 2018-03-23 | 河海大学常州校区 | Fabric defect detection method based on depth convolutional neural networks and vision significance |
| CN108257122A (en) * | 2018-01-09 | 2018-07-06 | 北京百度网讯科技有限公司 | Paper sheet defect detection method, device and server based on machine vision |
Cited By (21)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110910372A (en) * | 2019-11-23 | 2020-03-24 | 郑州智利信信息技术有限公司 | Deep convolutional neural network-based uniform light plate defect detection method |
| CN111144475A (en) * | 2019-12-22 | 2020-05-12 | 上海眼控科技股份有限公司 | Method and device for determining car seat, electronic equipment and readable storage medium |
| CN111179241A (en) * | 2019-12-25 | 2020-05-19 | 成都数之联科技有限公司 | Panel defect detection and classification method and system |
| CN111487250A (en) * | 2020-04-02 | 2020-08-04 | 苏州奥创智能科技有限公司 | Intelligent visual detection method and system applied to injection molding defective product detection |
| CN111583238A (en) * | 2020-05-09 | 2020-08-25 | 中南大学 | Method for extracting vertex of included angle of contour line of honeycomb, method for detecting vertex of included angle of contour line of honeycomb and device for detecting vertex of included angle of contour line of honeycomb |
| WO2021227288A1 (en) * | 2020-05-09 | 2021-11-18 | 中南大学 | Method for evaluating geometrical morphology of honeycomb product, and detection and evaluation system |
| CN111627020A (en) * | 2020-06-03 | 2020-09-04 | 山东贝特建筑项目管理咨询有限公司 | Detection method and system for anchor bolt in heat insulation board and computer storage medium |
| CN111808367A (en) * | 2020-07-09 | 2020-10-23 | 浙江七色鹿色母粒有限公司 | Improvement method for plastic PPR silver grain whitening defect |
| CN111808367B (en) * | 2020-07-09 | 2023-05-30 | 浙江七色鹿色母粒有限公司 | Improvement method for plastic PPR silver grain whitening defect |
| CN112232215A (en) * | 2020-10-16 | 2021-01-15 | 哈尔滨市科佳通用机电股份有限公司 | Railway wagon coupler yoke key joist falling fault detection method |
| CN112485260A (en) * | 2020-11-26 | 2021-03-12 | 常州微亿智造科技有限公司 | Workpiece defect detection method and device |
| CN112485260B (en) * | 2020-11-26 | 2023-01-03 | 常州微亿智造科技有限公司 | Workpiece defect detection method and device |
| CN112329896A (en) * | 2021-01-05 | 2021-02-05 | 武汉精测电子集团股份有限公司 | Model training method and device |
| CN113011057A (en) * | 2021-02-22 | 2021-06-22 | 河南农业大学 | Method and system for predicting performance of aged bonding structure based on gradient degradation of adhesive layer |
| CN112991344A (en) * | 2021-05-11 | 2021-06-18 | 苏州天准科技股份有限公司 | Detection method, storage medium and detection system based on deep transfer learning |
| CN114429433A (en) * | 2022-04-01 | 2022-05-03 | 武汉市宏伟纸箱包装有限公司 | Honeycomb paper feed amount control method based on image enhancement |
| CN114782286A (en) * | 2022-06-21 | 2022-07-22 | 奥蒂玛光学科技（深圳）有限公司 | Defect repairing method, optical repairing device, electronic device and storage medium |
| CN115063725A (en) * | 2022-06-23 | 2022-09-16 | 中国民航大学 | Airplane skin defect identification system based on multi-scale self-adaptive SSD algorithm |
| CN115063725B (en) * | 2022-06-23 | 2024-04-26 | 中国民航大学 | Aircraft skin defect identification system based on multi-scale self-adaptive SSD algorithm |
| CN115546203A (en) * | 2022-11-23 | 2022-12-30 | 常熟理工学院 | Production monitoring and analyzing method based on image data algorithm |
| CN115546203B (en) * | 2022-11-23 | 2023-03-10 | 常熟理工学院 | Production monitoring and analyzing method based on image data algorithm |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110111331B (en) | 2023-06-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110111331A (en) | Honeycomb paper core defect inspection method based on machine vision | |
| CN110378909B (en) | Single wood segmentation method for laser point cloud based on Faster R-CNN | |
| CN109166094B (en) | Insulator fault positioning and identifying method based on deep learning | |
| CN109613002B (en) | Glass defect detection method and device and storage medium | |
| CN110779937B (en) | Casting product internal defect intelligent detection device | |
| CN106127204B (en) | A kind of multi-direction meter reading Region detection algorithms of full convolutional neural networks | |
| CN112083017A (en) | Intelligent detection device, method and medium for internal defects of welding seam | |
| CN105844621A (en) | Method for detecting quality of printed matter | |
| CN108037770A (en) | Unmanned plane power transmission line polling system and method based on artificial intelligence | |
| CN110276386A (en) | A kind of apple grading method and system based on machine vision | |
| CN108711148B (en) | Tire defect intelligent detection method based on deep learning | |
| CN112200225B (en) | Steel rail damage B display image identification method based on deep convolution neural network | |
| CN103927534A (en) | Sprayed character online visual detection method based on convolutional neural network | |
| CN112634237A (en) | Long bamboo strip surface defect detection method and system based on YOLOv3 improved network | |
| CN109675827B (en) | Building rubbish identification and sorting equipment, identification method and grabbing method thereof | |
| CN112232263B (en) | Tomato identification method based on deep learning | |
| CN108764012A (en) | The urban road shaft recognizer of mobile lidar data based on multi-frame joint | |
| CN111523511B (en) | Video image Chinese wolfberry branch detection method for Chinese wolfberry harvesting and clamping device | |
| CN113160062A (en) | Infrared image target detection method, device, equipment and storage medium | |
| CN110929746A (en) | Electronic file title positioning, extracting and classifying method based on deep neural network | |
| CN112560675A (en) | Bird visual target detection method combining YOLO and rotation-fusion strategy | |
| CN113392849A (en) | R-CNN-based complex pavement crack identification method | |
| CN113313107A (en) | Intelligent detection and identification method for multiple types of diseases on cable surface of cable-stayed bridge | |
| CN109446964A (en) | Face detection analysis method and device based on end-to-end single-stage multiple scale detecting device | |
| CN115049662A (en) | Sprayer flow control method based on image processing |
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 |