CN110310259A - It is a kind of that flaw detection method is tied based on the wood for improving YOLOv3 algorithm - Google Patents
It is a kind of that flaw detection method is tied based on the wood for improving YOLOv3 algorithm Download PDFInfo
- Publication number
- CN110310259A CN110310259A CN201910529463.XA CN201910529463A CN110310259A CN 110310259 A CN110310259 A CN 110310259A CN 201910529463 A CN201910529463 A CN 201910529463A CN 110310259 A CN110310259 A CN 110310259A
- Authority
- CN
- China
- Prior art keywords
- image
- flaw
- wood
- knot
- defect detection
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2200/00—Indexing scheme for image data processing or generation, in general
- G06T2200/28—Indexing scheme for image data processing or generation, in general involving image processing hardware
-
- 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
- G06T2207/30161—Wood; Lumber
Abstract
The invention discloses a kind of based on the wood knot flaw detection method for improving YOLOv3 algorithm, belongs to technical field of machine vision.There are wood knot flaw and indefectible image as the progress data amplification of batten surface image data set by choosing largely；Dimension cluster is carried out to flaw target frame using k-means++ algorithm；3 scales in YOLOv3 network are reduced to 2 size measurements, while improving loss loss function according to contrast and normalization thought；YOLOv3 algorithm after being improved using training sample set off-line training, different flaws in high-precision on-line checking batten surface.This method takes full advantage of the enhancing sample multiplicity sexuality of data amplification, k-means++ makes initial candidate frame number and size be more suitable for wood knot Defect Detection simultaneously, multiple scale detecting improves the identification accuracy and detection speed significantly improved to different size objectives from loss function, meets industrial requirement of real-time.
Description
Technical field
The present invention relates to a kind of based on the wood knot flaw detection method for improving YOLOv3 algorithm, belongs to machine vision technique neck
Domain.
Background technique
How reasonably to make full use of forest tree resource is China's forestry Scientific Research Workers significant problem urgently to be solved,
In to carry out Defect Detection to timber be a kind of effective means that China's wood comprehensive utilization rate can be improved, and need to study
Emphasis.Batten has highly textured natural material as a kind of be prevalent in nature, and flaw shows as being embedded in
In batten surface texture, the apparent region of color difference.The significant challenge of wood knot flaw identification is from possessing complex characteristic
Natural texture in it is carried out precisely detect and positioning, this to come into the artificial detection of high investment, inefficient stage
Method is difficult to be competent at.So studying a kind of intelligent measurement identification wood knot flaw method that accuracy is high, detection speed is fast, at low cost
There are particularly important theory significance and very urgent current demand to the processing of high quality batten.
In recent years, machine vision technique is developed rapidly, and in the industrial production, people have utilized technology success
Various products appearance is detected and realizes automatic operation.But batten surface image reflects that one kind becomes independent of color or brightness
The complex texture of change acquires batten surface image by NI Vision Builder for Automated Inspection, and detects wood using traditional image processing algorithm
Flaw is tied, huge challenge can be faced.
As convolutional neural networks are in commonly used and GPU the fast development of image domains, people get over attention
Come it is more be transferred to deep learning in terms of, be widely used to recognition of face, vehicle detection, object detection and know in video
Not, many fields such as Navigation of Pilotless Aircraft become current mainstream algorithm.Since the shape of natural texture is most irregular, distribution
Position uncertainty it is larger, based on the algorithm of target detection of deep learning by feat of speed faster, the higher advantage of accuracy,
Just gradually substitute traditional image processing algorithm.
Algorithm of target detection based on deep learning is divided into based on candidate region and based on recurrence.It is main based on candidate region
There are the methods of R-CNN, Fast R-CNN, Faster R-CNN, generates network (RPN) constantly repetition training mould with candidate region
Type.There is researcher to classify using support vector machines (SVM), total algorithm is cumbersome, operation is slow, and timber Defect Detection speed is
660s is unable to satisfy industrial requirement of real-time.Mainly there is the side such as SSD and YOLO, YOLOv2, YOLOv3 based on returning
Method.SSD can solve the dimension disaster problem generated in extracting characteristic procedure, but to small size and fuzzy flaw target
Detection accuracy it is lower.YOLOv3 is improved on the basis of YOLO and YOLOv2, has taken into account network complexity and detection
Accuracy, detects fast 3 times of speed ratio SSD on COCO data set, and accuracy of the mean (MAP) is imitated up to 57.9% optimum detection
Fruit.Therefore it attempts using the improved YOLOv3 algorithm of target detection of a large amount of batten surface image set of data samples training, by more
Layer Fusion Features sufficiently extract wood knot unwanted visual characteristic, recycle trained algorithm model on-line checking wood knot flaw, this is one
The effective solution of kind, has high generalization.
Therefore operand can be reduced, to mention while guaranteeing wood knot flaw identification with setting accuracy by providing one kind
The accuracy of high detection and the detection method of real-time are necessary.
Summary of the invention
The present invention is being based on aiming at the problem that traditional images algorithm is difficult to meet the requirement of real-time in industrial production
In the wood knot Defect Detection system of YOLOv3, the identification of batten surface blemish and location difficulty propose one kind and are based on improving
The wood knot flaw detection method of YOLOv3 algorithm.Improved method is used for Defect Detection in batten surface image, is utilized
The fireballing advantage of the detection of YOLOv3 first acquires the progress data amplification of batten surface image data set and reaches training data sample
Collection carries out dimension cluster to the wood knot flaw target frame manually marked using k-means++ clustering algorithm, obtains suitable initial
3 scales are reduced to 2 size measurements, while according to contrast and returning then in YOLOv3 network by candidate frame number and size
One, which changes thought, improves loss loss function, utilizes YOLOv3 algorithm on-line checking batten surface image after the good improvement of off-line training
In different flaws, can effectively enhance sample multiplicity sexuality in this way, improve the accuracy and detection speed of system, can be with
Meet the requirement of real-time in industrial production.
For technical purpose more than realization, the present invention provides a kind of based on the wood knot Defect Detection side for improving YOLOv3 algorithm
Method, which comprises
Step 1: the acquisition of original long stuff surface image is carried out using optical imaging system；
Step 2: the collected original long stuff surface image of step 1 is divided into the identical image of high wide size；
Step 3: having randomly selected wood knot flaw image and indefectible image as trained and test data sample set；
Step 4: at host computer PC end, YOLOv3 algorithm model after being improved using training data sample set off-line training；
Step 5: utilizing wood knot in YOLOv3 algorithm model on-line checking batten surface image after the trained improvement of step 4
Defect classification and position location；
Step 6: upper computer detection result being exported and shown, to drive slave computer movement executing mechanism to detected
Long stuff carry out in real time cut processing, construct wood knot Defect Detection system.
Optionally, the step 3 further includes data amplification and mark being carried out to training data sample set, to increase sample
The diversity of illumination, the picture quality of training for promotion set of data samples, comprises the following steps:
Step 3.1: acquisition fast knot flaw and slip-knot flaw two major classes image and indefectible image as training sample set with
Test sample collection, wherein having fast knot flaw flaw, thering is slip-knot flaw flaw and the identical image of indefectible high width size respectively to choose
5000；
Step 3.2: image data sample set is expanded using horizontal and vertical mirror image；
Step 3.3: artificial mark wood knot flaw being carried out to set of data samples after expansion, wherein tag along sort is fast knot and work
Two class wood knot flaw of knot；
Step 3.4: image enhancement is carried out to image data sample set using adaptive histogram equalization；
Step 3.5 step: in the training process, to image carry out saturation degree and exposure adjustment, random cropping and ±
15 °, ± 30 ° of rotations.
Optionally, in the step 4, YOLOv3 algorithm model after being improved using training data sample set off-line training,
The development of YOLOv3 algorithm model is specific as follows after middle improvement:
Step 4.1: dimension being carried out to the fast knot and slip-knot flaw target frame manually marked using k-means++ clustering algorithm
Cluster obtains K initial candidate frame for being suitble to wood knot flaw；Select improved k-means++ algorithm] to batten surface data
Collection is that point apart from each other has more maximum probability to be chosen as next cluster centre as clustering, participates in Europe using friendship and than IOU
Formula distance calculates, to improve the setting accuracy of wood knot flaw.Wherein Euclidean distance function is Di(xj), expression is
Di(xj)=1-IOU (xj,ki)
X in formulaj∈ X={ x1,x2... } and it is real goal frame；ki∈{k1,k2,…,kKIt is cluster centre；K is anchor
The number of box；Clustering objective function J (K) indicates sum of the distance minimum value of each sample to its cluster centre, expression
For
Step 4.2: construction deep layer convolutional neural networks Darknet-53, specific structure are as follows: the 0th layer of input for pixel be
416 × 416, the image that port number is 3,32 layers of convolution kernel, each convolution kernel size is 3 × 3, and output is 32 channels
The characteristic pattern of 416 × 416 sizes；From the 0th layer up to 74 layers, one shares 53 convolutional layers, remaining is res layers, uses a system
3 × 3 and 1 × 1 convolution of column constitutes residual error module, effectively avoids occurring the problems such as gradient disappears or explodes in training process；
Under in network training process, the output of convolutional layer is the input as next layer of pond layer, and the output of pond layer is used as
The input of one layer of convolutional layer；In j-th of convolution kernel, i-th of convolution operation output matrix of first of convolutional layer can be obtained, is expressed
Formula is
F () is activation primitive in formula, and b is biasing, and x is the convolution kernel of n × n, and m is the matrix of input picture m × m；
Using the Relu function of Leakey function replacement YOLOv3；When inputting x less than 0, export as 0.01x, specific table
It is up to formula
F ()=Leakey (x)=max (0.01x, x)
It selects the operation of average value pondization to carry out down-sampling to the output of convolutional layer, it is extensive to further decrease calculation amount, raising
Property；In the region of j-th of pond, i-th of pondization operation output matrix of first of pond layer can be obtained, expression formula is
N is the neuron number of (l-1) a convolutional layer in formula,For this layer of convolution operation output matrix it
It is pond step-length with, N；
Step 4.3: loss loss function is improved according to contrast and normalization thought；The loss loss function is used for
When carrying out prediction target category, optimize wood knot flaw classification and bbox position coordinates, while introducing scale factor λcoordWith
λnoobj, IOU error and error of coordinate are corrected, the susceptibility of nontarget area, the stability of lift scheme are inhibited；In view of image
Middle wood knot flaw target sizes have an impact network undated parameter and small size Defect Detection result, utilize contrast and normalizing
Change thought, in error of coordinate w and h improve, improved error of coordinate can effectively improve wood knot flaw positioning
Accuracy, expression are
B is the number of each grid forecasting bounding box in formula；X and y for predicted boundary frame center relative to its institute
The position of the net boundary of category；W and h is width and height of the predicted boundary frame relative to whole image；S2For input picture quilt
It is divided into S × S grid；ErrcoordError of coordinate between prediction data and nominal data；ErrIOUFor IOU error；
ErrclassFor error in classification；λcoordFor 2-wi·hi, can be according to wide height adjustment weight coefficient, to reduce bbox size to Errcoord
The influence of generation；For the predicted value of i-th of grid, xi,yi,wi,hiFor its true value；Indicate mesh
Mark is with the presence or absence of in j-th of bbox of i-th of grid；pi(c) forUnder the premise of i-th grid include to belong to c
Classification target true conditional probability；For its predicted condition probability；Class is general objective classification number；Lcross-entropyIt is two
First cross entropy (logistic) function, expression are
Lcross-entropy(p, q)=plog (q)+(1-p) log (1-q)
In formulaFor the predicted value for identifying target category number in i-th of grid；CiFor its true calibration value；λnoobjFor weight
0.5；Indicate whether target is not present in j-th of bbox of i-th of grid；
Step 4.4: 3 kinds of scales being reduced to 2 kinds of scales, multiple scale detecting network is constructed, passes through convolution kernel (3 × 3 and 1
× 1) mode realizes the local feature between characteristic pattern, is the feature interaction layer of YOLOv3 network from the 75 to 95th layer, is divided into
Two scales, specific input and output are as follows:
The characteristic pattern that the input of smallest dimension layer is 13 × 13, altogether 1024 channels export the characteristic pattern for 13 × 13, and one
Totally 75 channels, carry out classification on this basis and position returns；
The input of mesoscale layer is that the characteristic pattern in the 79th layer of 13 × 13,512 channels is carried out to convolution operation, generation 13 ×
13, the characteristic pattern in 256 channels, is then up-sampled, and the characteristic pattern in 26 × 26,256 channels, while 26 with the 61st layer are generated
The characteristic pattern of the mesoscale in × 26,512 channels merges, and carries out convolution operation, exports as the characteristic pattern of 26 × 26 sizes, and altogether 75
A channel, then carries out classification herein and position returns；
Step 4.5: YOLOv3 algorithm model after being improved using training data sample set off-line training.
Optionally, in the step 6, wood knot Defect Detection system includes the following:
Image capture module, including lighting system and CCD optical imaging system, sweep camera using lighting device and CCD line
Acquire long stuff surface image；
Image procossing and identification module, the image procossing and identification module are host computer PC end, including MATLAB schemes
As processing software, Python composing software, database and with the communication module that slave computer is communicated, compiled using Python
Software to after processing wood knot flaw image detect, and Defect Detection as the result is shown and export give slave computer Motor execution machine
Structure；The Python composing software is to utilize open source in the Ubuntu system of host computer PC end for realizing YOLOv3 algorithm
After frame TensorFlow realizes improved YOLOv3 algorithm, using training data sample set off-line training algorithm model, and benefit
Accelerate training process up to GTX 1080Ti video card with tall and handsome；The database be used for batten surface image set of data samples and
Wood knot Defect Detection result carries out storage management；The communication module uses Ethernet ICP/IP protocol, allows host computer PC end
Software is communicated with slave computer movement executing mechanism；
Slave computer Motor execution module, the slave computer movement executing mechanism, according to host computer PC end transmit come wood
The classification and positioning position information for tying flaw, using programmable logic controller (PLC) PLC control wood processing equipment to batten surface
Flaw carries out cutting processing.
Optionally, in the step 1, the optical imaging system is that CCD industrial camera is swept using Basier-12k line, is used
In the original long stuff surface image of acquisition.
It is another object of the present invention to provide a kind of wood knot Defect Detection system, shown wood knot Defect Detection system packet
It includes:
Image capture module, for acquiring long stuff surface image；
Image procossing and identification module, the image procossing are used for image capture module is collected with identification module
Long stuff surface image is divided into the identical image of high wide size and carries out data and expands to obtain training data sample set, to expansion
Increase obtained training data sample set and carry out artificial mark wood knot flaw target frame, wherein tag along sort is two class of fast knot and slip-knot
Wood knot flaw；Dimension cluster is carried out to the wood knot flaw target frame manually marked using k-means++ clustering algorithm, it is suitable to obtain
Initial candidate frame number and size 3 scales are reduced to 2 size measurements, while according to comparison then in YOLOv3 network
Degree and normalization thought improve loss loss function, utilize YOLOv3 algorithm on-line checking batten table after the good improvement of off-line training
Different flaws in the image of face；The Defect Detection result that on-line checking goes out is exported and gives slave computer movement executing mechanism, flaw inspection
Survey the classification and positioning position information that result includes wood knot flaw；
Slave computer Motor execution module, the slave computer movement executing mechanism are passed according to image procossing and identification module
Defeated next Defect Detection is as a result, carry out batten surface blemish using programmable logic controller (PLC) PLC control wood processing equipment
Cut processing.
Optionally, described image processing is realized with identification module using host computer PC end, including MATLAB image procossing is soft
Part, Python composing software, database and the communication module communicated with slave computer；MATLAB image processing software will scheme
After being handled as the collected long stuff surface image of acquisition module, by Python composing software to wood knot flaw figure after processing
It as the result is shown and exports as being detected, and Defect Detection and gives slave computer movement executing mechanism；The Python compiling is soft
Part is to be changed in the Ubuntu system of host computer PC end using Open Framework TensorFlow realization for realizing YOLOv3 algorithm
Into YOLOv3 algorithm after, using training data sample set off-line training algorithm model, and using tall and handsome aobvious up to GTX 1080Ti
Card accelerates training process；The database is used to carry out batten surface image set of data samples and wood knot Defect Detection result
Storage management；The communication module uses Ethernet ICP/IP protocol, allows host computer PC end software and slave computer Motor execution
Mechanism is communicated.
Optionally, described image acquisition module includes lighting system and CCD optical imaging system.
Optionally, the CCD optical imaging system includes that Basier-12k line sweeps CCD industrial camera.
Third object of the present invention be to provide it is above-mentioned based on the wood knot flaw detection method for improving YOLOv3 algorithm and/
Or above-mentioned wood knot Defect Detection system utilizes the application in technical field in treating of wood.
According to above technical solution, may be implemented it is below the utility model has the advantages that
Compared with the prior art, the present invention has the following advantages:
1) compared with k-means dimension cluster in YOLOv3, using k-means++ algorithm to self-control batten surface image number
According to concentrating real goal frame to carry out dimension cluster, identified initial candidate frame number and size are more suitable for wood knot Defect Detection,
It is more stable compared to before improving.
2) present invention introduces contrast and normalization thought, improves while all advantages of succession YOLOv3 algorithm
Error of coordinate in loss loss function avoids the situation that wood knot flaw is not of uniform size in image, is conducive to improve small size flaw
The setting accuracy of target；It is 2 kinds by original 3 kinds of size compressions simultaneously by improving the multiple scale detecting of YOLOv3 algorithm
Size detection, improves the accuracy of classification with fixation and recognition large scale flaw target, while reducing operand.
3) present invention is calculated by the batten surface image set of data samples training deep learning after different data amplification method
Method recycles YOLOv3 algorithm detection after trained improvement to identify wood knot flaw, solve traditional images Processing Algorithm without
Method detects the problem for identifying complicated natural texture flaw, can detect wood knot flaw in batten high-precision surface, and know
Then the classification of other flaw is realized using Open Framework TensorFlow, and add using using tall and handsome up to GTX 1080Ti video card
Fast training process meets industrial requirement of real-time.
4) present invention passes through image capture module, image procossing and identification module and the Motor execution module building wood knot flaw
Defect detection system, the cooperation between module can efficiently realize the detection to different complex texture batten surface images, can not only
Fast knot and slip-knot flaw are enough accurately detected out, adaptability is extensive, and has preferable robustness and faster detection speed.
Detailed description of the invention
To describe the technical solutions in the embodiments of the present invention more clearly, make required in being described below to embodiment
Attached drawing is briefly described, it should be apparent that, drawings in the following description are only some embodiments of the invention, for
For those of ordinary skill in the art, without creative efforts, it can also be obtained according to these attached drawings other
Attached drawing.
A kind of Fig. 1 system structure signal based on the wood knot flaw detection method for improving YOLOv3 algorithm of the present invention
Figure.
A kind of Fig. 2 working-flow based on the wood knot flaw detection method for improving YOLOv3 algorithm of the present invention
Figure.
A kind of Fig. 3 multiple scale detecting net based on the wood knot flaw detection method for improving YOLOv3 algorithm of the present invention
Network.
A kind of Fig. 4 k-means++ calculation based on the wood knot flaw detection method for improving YOLOv3 algorithm of the present invention
Method flow chart.
A kind of Fig. 5 algorithm network structure based on the wood knot flaw detection method for improving YOLOv3 algorithm of the present invention
Schematic diagram.
Specific embodiment
To make the object, technical solutions and advantages of the present invention clearer, below in conjunction with attached drawing to embodiment party of the present invention
Formula is described in further detail.
Embodiment one:
The present embodiment provides a kind of wood knot Defect Detection system, shown wood knot Defect Detection system includes:
Image capture module, for acquiring long stuff surface image；
Image procossing and identification module, the image procossing are used for image capture module is collected with identification module
Long stuff surface image is divided into the identical image of high wide size and carries out data and expands to obtain training data sample set, to expansion
Increase obtained training data sample set and carry out artificial mark wood knot flaw target frame, wherein tag along sort is two class of fast knot and slip-knot
Wood knot flaw；Dimension cluster is carried out to the wood knot flaw target frame manually marked using k-means++ clustering algorithm, it is suitable to obtain
Initial candidate frame number and size 3 scales are reduced to 2 size measurements, while according to comparison then in YOLOv3 network
Degree and normalization thought improve loss loss function, utilize YOLOv3 algorithm on-line checking batten table after the good improvement of off-line training
Different flaws in the image of face；The Defect Detection result that on-line checking goes out is exported and gives slave computer movement executing mechanism, flaw inspection
Survey the classification and positioning position information that result includes wood knot flaw；
Slave computer Motor execution module, the slave computer movement executing mechanism are passed according to image procossing and identification module
Defeated next Defect Detection is as a result, carry out batten surface blemish using programmable logic controller (PLC) PLC control wood processing equipment
Cut processing.
Described image, which is handled, to be realized with identification module using host computer PC end, including MATLAB image processing software,
Python composing software, database and the communication module communicated with slave computer；MATLAB image processing software is by image
After the collected long stuff surface image of acquisition module is handled, by Python composing software to wood knot flaw image after processing
Detected, and Defect Detection as the result is shown and export give slave computer movement executing mechanism；The Python composing software
It for realizing YOLOv3 algorithm, is realized and is improved using Open Framework TensorFlow in the Ubuntu system of host computer PC end
YOLOv3 algorithm after, add using training data sample set off-line training algorithm model, and using tall and handsome up to GTX1080Ti video card
Fast training process；The database is used to store batten surface image set of data samples and wood knot Defect Detection result
Management；The communication module uses Ethernet ICP/IP protocol, allows host computer PC end software and slave computer movement executing mechanism
It is communicated.
Described image acquisition module includes lighting system and CCD optical imaging system, and the CCD optical imaging system uses
Basier-12k line sweeps CCD industrial camera.
As shown in Figure 1, image capture module includes light source and CCD camera in Fig. 1, show figure for acquiring timber
Picture；
Timber is sent in CCD camera field range by transport mechanism, CCD camera is collected after timber shows image and passed
It send to the end PC, Defect Detection is carried out to timber by the end PC, and Defect Detection result is sent to slave computer Motor execution module, with
Just slave computer Motor execution module according to the end PC transmission come Defect Detection as a result, using programmable logic controller (PLC) PLC control
Wood processing equipment carries out cutting processing to batten surface blemish, since the application does not limit slave computer Motor execution module
It is fixed, so not showing that slave computer Motor execution module in Fig. 1.
Wood knot Defect Detection system provided in this embodiment, specific detection process please refer to following embodiments two.
Embodiment two:
The present embodiment provides a kind of based on the wood knot flaw detection method for improving YOLOv3 algorithm.
It is provided by the invention that flaw detection method is tied based on the wood for improving YOLOv3 algorithm, pass through and acquires batten surface image
Set of data samples carries out dimension cluster using k-means++ clustering algorithm, provides suitable initial candidate frame number and size,
Then reduce multiple scale detecting network in YOLOv3 network, improve loss loss function, after the good improvement of off-line training
The different wood knot flaws of model on-line checking, entire wood knot Defect Detection system is by image capture module, image procossing and identification mould
Block and slave computer Motor execution module composition.
As shown in figure 3, specific implementation step are as follows:
Step 1: carrying out the acquisition of original long stuff surface image using optical imaging system, that is, use Basier-12k line
Sweep CCD industrial camera acquisition long stuff surface image；
Step 2: original long stuff surface image is divided into the identical image of high wide size；
Step 3: having randomly selected wood knot flaw image and indefectible image as trained and test data sample set；Its
In, trained and test data sample set is subjected to data amplification and mark, to increase the diversity of sample light photograph, promotes sample set
Picture quality, the specific steps are as follows:
Step 3.1: acquisition fast knot flaw and slip-knot flaw two major classes image and indefectible image as training sample set with
Test sample collection, wherein having fast knot flaw flaw, thering is slip-knot flaw flaw and the identical image of indefectible high width size respectively to choose
5000；
Step 3.2: image data sample set is expanded using horizontal and vertical mirror image；
Step 3.3: artificial mark wood knot flaw being carried out to set of data samples after expansion, wherein tag along sort is fast knot and work
Two class wood knot flaw of knot；
Step 3.4: image enhancement is carried out to set of data samples using adaptive histogram equalization；
Step 3.5 step: in the training process, to image carry out saturation degree and exposure adjustment, random cropping and ±
15 °, ± 30 ° of rotations.
Step 4: at host computer PC end, YOLOv3 algorithm model after being improved using training data sample set off-line training；
Wherein using training data sample set off-line training improve after YOLOv3 algorithm model, wherein to YOLOv3 algorithm into
Row improves, and development is specific as follows:
Step 4.1: dimension being carried out to the fast knot and slip-knot flaw target frame manually marked using k-means++ clustering algorithm
Cluster obtains K initial candidate frame for being suitble to wood knot flaw；
Wherein, in former YOLOv3 algorithm, K initial cluster center is selected at random from sample set using k-means, is caused
Different cluster results can be obtained by executing algorithm every time, and accuracy rate is relatively low, and anchor box setting quality also will affect detection
Speed and accuracy.As shown in figure 4, the application selects improved k-means++ algorithm to cluster batten surface data collection
Analysis, main thought is that point apart from each other has more maximum probability to be chosen as next cluster centre, participates in Europe using friendship and than (IOU)
Formula distance calculates, to improve the setting accuracy of wood knot flaw.Wherein Euclidean distance function is Di(xj), expression is
Di(xj)=1-IOU (xj,ki)
X in formulaj∈ X={ x1,x2... } and it is real goal frame；ki∈{k1,k2,…,kKIt is cluster centre；K is anchor
The number of box.Clustering objective function J (K) indicates sum of the distance minimum value of each sample to its cluster centre, expression
For
Step 4.2: construction deep layer convolutional neural networks Darknet-53；
As shown in figure 3, constructing deep layer convolutional neural networks Darknet-53, specific structure are as follows: the 0th in the step 4.2
It is 416 × 416 that layer input, which is pixel, the image that port number is 3,32 layers of convolution kernel, and each convolution kernel size is 3 × 3, defeated
It is out the characteristic pattern (feature map) of 416 × 416 sizes in 32 channels；From the 0th layer up to 74 layers, one shares 53
Convolutional layer, remaining is res layers, constitutes residual error module using a series of 3 × 3 and 1 × 1 convolution, effectively avoids in training process
There is the problems such as gradient disappears or explodes.In network training process, the output of convolutional layer is as next layer of pond layer
Input, and input of the output of pond layer as next layer of convolutional layer.In j-th of convolution kernel, first of convolutional layer can be obtained
I-th of convolution operation output matrix, expression are as follows:
F () is activation primitive in formula, and b is biasing, and x is the convolution kernel of n × n, and m is the matrix of input picture m × m
It selects the operation of average value pondization to carry out down-sampling (Up-sampling) to the output of convolutional layer, further decreases meter
Calculation amount improves generalization.In the region of j-th of pond, i-th of pondization operation output matrix of first of pond layer can be obtained, specifically
Expression formula is
N is the neuron number of (l-1) a convolutional layer in formula,For this layer of convolution operation output matrix
The sum of, N is pond step-length.
Leaky function is better than Relu function as the model of convolutional layer activation primitive on identifying and positioning accuracy,
Therefore with the Relu function of Leakey function replacement YOLOv3.When inputting x less than 0, output is no longer 0, but one small
Digital 0.01x, expression is
F ()=Leakey (x)=max (0.01x, x)
Step 4.3: loss loss function is improved according to contrast and normalization thought；
Wherein, when carrying out prediction target category, in YOLOv3 loss loss function for optimize wood knot flaw classification with
Bbox position coordinates introduce scale factor λcoordAnd λnoobj, correct error E rrIOUAnd Errcoord, while inhibiting nontarget area
Susceptibility, improve the stability of model.Expression are as follows:
B is the number of each grid forecasting bounding box in formula；X and y for predicted boundary frame center relative to its institute
The position of the net boundary of category；W and h is width and height of the predicted boundary frame relative to whole image；S2For input picture quilt
It is divided into S × S grid；ErrcoordError of coordinate between prediction data and nominal data；ErrIOUFor IOU error；
ErrclassFor error in classification；λcoordFor 2-wi·hi, can be according to wide height adjustment weight coefficient, to reduce bbox size pair
ErrcoordThe influence of generation；For the predicted value of i-th of grid, xi,yi,wi,hiFor its true value；
Indicate that target whether there is in j-th of bbox of i-th of grid；pi(c) forUnder the premise of i-th of grid include
Belong to c classification target true conditional probability；For its predicted condition probability；Class is general objective classification number.
Lcross-entropyFor binary cross entropy (logistic) function, expression is
Lcross-entropy(p, q)=plog (q)+(1-p) log (1-q)
In formulaFor the predicted value for identifying target category number in i-th of grid；CiFor its true calibration value；λnoobjFor weight
0.5；Indicate whether target is not present in j-th of bbox of i-th of grid.
In view of wood knot flaw target sizes generate shadow to network undated parameter and small size Defect Detection result in image
It rings, YOLOv3 directly carries out w and h to seek variance, cannot handle the problem well, right using contrast and normalization thought
Error of coordinate ErrcoordIn w and h improve, improved error of coordinate ErrcoordIt can effectively improve wood knot flaw
Setting accuracy, expression are
Step 4.4: 3 kinds of scales being reduced to 2 kinds of scales, construct multiple scale detecting network；
As shown in figure 3, multiple scale detecting network is constructed in the step 4.4, wherein passing through convolution kernel (3 × 3 and 1 × 1)
Mode realize the local feature between characteristic pattern, from the 75 to 95th layer be YOLOv3 network feature interaction layer, be divided into two
Scale, specifically input and output are as follows: the input of smallest dimension layer is 13 × 13 characteristic pattern, altogether 1024 channels, and exporting is 13
× 13 characteristic pattern, 75 channels, carry out classification on this basis and position return altogether；The input of mesoscale layer is by the 79th layer
13 × 13,512 channels characteristic pattern carry out convolution operation, generate 13 × 13,256 channels characteristic pattern, then carry out adopting
Sample generates the characteristic pattern in 26 × 26,256 channels, while the characteristic pattern of the mesoscale in 26 × 26,512 channels with the 61st layer closes
And a series of convolution operations are carried out, it exports as the characteristic pattern of 26 × 26 sizes, 75 channels, then classify herein altogether
It is returned with position.
Step 4.5: algorithm model after being improved using training data sample set off-line training.
Step 5: using YOLOv3 algorithm model on-line checking batten surface image after trained improvement, carrying out mathematically
Description and identification wood knot defect classification and position location.
Step 6: upper computer detection result being exported and shown, to drive slave computer movement executing mechanism to detected
Long stuff carry out in real time cut processing, construct wood knot Defect Detection system.Wherein wood knot Defect Detection system specifically includes:
Image capture module, including lighting system and CCD optical imaging system, sweep camera using lighting device and CCD line,
For acquiring long stuff surface image.
Image procossing and identification module, the image procossing and identification module are host computer PC end, including MATLAB schemes
As processing software, Python composing software, database and the communication module communicated with slave computer；Image capture module will
Collected long stuff surface image is uploaded to the end PC, and the end PC is using Python composing software to through MATLAB image processing software
After processing wood knot flaw image detected, and Defect Detection as the result is shown and export give slave computer movement executing mechanism；The flaw
Defect testing result includes the classification and positioning position information of wood knot flaw.The Python composing software is for realizing YOLOv3
Algorithm, be in the Ubuntu system of host computer PC end, after realizing improved YOLOv3 algorithm using Open Framework TensorFlow,
Accelerate training process up to GTX 1080Ti video card using training data sample set off-line training algorithm model, and using tall and handsome；Number
It is used to carry out storage management to batten surface image set of data samples and wood knot Defect Detection result according to library；Communication module use with
Too net ICP/IP protocol allows host computer PC end software and slave computer movement executing mechanism communicate.
Motor execution module, the slave computer movement executing mechanism, according to host computer PC end transmission come wood tie flaw
Classification and positioning position information, using programmable logic controller (PLC) PLC control wood processing equipment to batten surface blemish into
Row cuts processing.
The present invention passes through image capture module, image procossing and identification module and Motor execution module building wood knot flaw
Detection system, the cooperation between module can efficiently realize the detection to different complex texture batten surface images, can not only
Fast knot and slip-knot flaw are accurately detected out, adaptability is extensive, and has preferable robustness and faster detection speed.
Part steps in the embodiment of the present invention, can use software realization, and corresponding software program can store can
In the storage medium of reading, such as CD or hard disk.
The foregoing is merely presently preferred embodiments of the present invention, is not intended to limit the invention, it is all in spirit of the invention and
Within principle, any modification, equivalent replacement, improvement and so on be should all be included in the protection scope of the present invention.
Claims (10)
1. a kind of based on the wood knot flaw detection method for improving YOLOv3 algorithm, which is characterized in that the described method includes:
Step 1: the acquisition of original long stuff surface image is carried out using optical imaging system；
Step 2: the collected original long stuff surface image of step 1 is divided into the identical image of high wide size；
Step 3: having randomly selected wood knot flaw image and indefectible image as trained and test data sample set；
Step 4: at host computer PC end, YOLOv3 algorithm model after being improved using training data sample set off-line training；
Step 5: utilizing wood knot flaw in YOLOv3 algorithm model on-line checking batten surface image after the trained improvement of step 4
Classification and position location；
Step 6: upper computer detection result being exported and shown, to drive slave computer movement executing mechanism to detected length
Batten carries out cutting processing in real time, constructs wood knot Defect Detection system.
2. being carried out the method according to claim 1, wherein the step 3 further includes to training data sample set
Data amplification and mark, to increase the diversity of sample light photograph, the picture quality of training for promotion set of data samples includes following step
It is rapid:
Step 3.1: acquisition fast knot flaw and slip-knot flaw two major classes image and indefectible image are as training sample set and test
Sample set, wherein having fast knot flaw flaw, thering is slip-knot flaw flaw and the identical image of the indefectible wide size of height respectively to choose 5000
?；
Step 3.2: image data sample set is expanded using horizontal and vertical mirror image；
Step 3.3: wood knot flaw being labeled to set of data samples after expansion, wherein tag along sort is fast knot and slip-knot two classes wood
Tie flaw；
Step 3.4: image enhancement is carried out to set of data samples using adaptive histogram equalization；
Step 3.5 step: in the training process, to input picture carry out saturation degree and exposure adjustment, random cropping and ±
15 °, ± 30 ° of rotations.
3. according to the method described in claim 2, it is characterized in that, being instructed offline in the step 4 using training data sample set
Practice YOLOv3 algorithm model after improving, wherein the development of YOLOv3 algorithm model is specific as follows after improving:
Step 4.1: it is poly- that dimension being carried out to the fast knot and slip-knot flaw target frame manually marked using k-means++ clustering algorithm
Class obtains K initial candidate frame for being suitble to wood knot flaw；Select improved k-means++ algorithm to batten surface data collection
It is that point apart from each other has more maximum probability to be chosen as next cluster centre as clustering, is participated in using friendship and than IOU European
Distance calculates, to improve the setting accuracy of wood knot flaw；Wherein Euclidean distance function is Di(xj), expression is
Di(xj)=1-IOU (xj,ki)
X in formulaj∈ X={ x1,x2... } and it is real goal frame；ki∈{k1,k2,…,kKIt is cluster centre；K is anchor box
Number；Clustering objective function J (K) indicates each sample to the sum of the distance minimum value of its cluster centre, and expression is
Step 4.2: construction deep layer convolutional neural networks Darknet-53, specific structure are as follows: the 0th layer of input be pixel be 416 ×
416, port number be 3 image, 32 layers of convolution kernel, each convolution kernel size be 3 × 3, output be 32 channels 416 ×
The characteristic pattern of 416 sizes；From the 0th layer up to 74 layers, one shares 53 convolutional layers, remaining is res layers, uses a series of 3
× 3 and 1 × 1 convolution constitutes residual error module, effectively avoids occurring the problems such as gradient disappears or explodes in training process；In network
In training process, the output of convolutional layer is the input as next layer of pond layer, and the output of pond layer is as next layer of volume
The input of lamination；In j-th of convolution kernel, i-th of convolution operation output matrix of first of convolutional layer can be obtained, expression formula is
F () is activation primitive in formula, and b is biasing, and x is the convolution kernel of n × n, and m is the matrix of input picture m × m；
Using the Relu function of Leakey function replacement YOLOv3；When inputting x less than 0, export as 0.01x, expression
For
F ()=Leakey (x)=max (0.01x, x)
It selects the operation of average value pondization to carry out down-sampling to the output of convolutional layer, further decreases calculation amount, improves generalization；?
In the region of j-th of pond, i-th of pondization operation output matrix of first of pond layer can be obtained, expression formula is
N is the neuron number of (l-1) a convolutional layer in formula,For the sum of this layer of convolution operation output matrix, N
For pond step-length；
Step 4.3: loss loss function is improved according to contrast and normalization thought；The loss loss function is for carrying out
When predicting target category, optimize wood knot flaw classification and bbox position coordinates, while introducing scale factor λcoordAnd λnoobj, repair
Positive IOU error and error of coordinate, inhibit the susceptibility of nontarget area, the stability of lift scheme；In view of wood knot in image
Flaw target sizes have an impact network undated parameter and small size Defect Detection result, are thought using contrast and normalization
Think, in error of coordinate w and h improve, improved error of coordinate can effectively improve wood knot flaw registration
Degree, expression are
B is the number of each grid forecasting bounding box in formula；X and y is the center of predicted boundary frame relative to belonging to it
The position of net boundary；W and h is width and height of the predicted boundary frame relative to whole image；S2It is divided for input picture
At S × S grid；ErrcoordError of coordinate between prediction data and nominal data；ErrIOUFor IOU error；Errclass
For error in classification；λcoordFor 2-wi·hi, can be according to wide height adjustment weight coefficient, to reduce bbox size to ErrcoordIt generates
It influences；For the predicted value of i-th of grid, xi,yi,wi,hiFor its true value；Whether indicate target
It is present in j-th of bbox of i-th of grid；pi(c) forUnder the premise of i-th grid include to belong to c class target
True conditional probability；For its predicted condition probability；Class is general objective classification number；Lcross-entropyFor binary intersection
Entropy (logistic) function, expression are
Lcross-entropy(p, q)=plog (q)+(1-p) log (1-q)
In formulaFor the predicted value for identifying target category number in i-th of grid；CiFor its true calibration value；λnoobjFor weight 0.5；Indicate whether target is not present in j-th of bbox of i-th of grid；
Step 4.4: 3 kinds of scales being reduced to 2 kinds of scales, multiple scale detecting network is constructed, passes through convolution kernel (3 × 3 and 1 × 1)
Mode realize the local feature between characteristic pattern, from the 75 to 95th layer be YOLOv3 network feature interaction layer, be divided into two
Scale, specific input and output are as follows:
The characteristic pattern that the input of smallest dimension layer is 13 × 13,1024 channels, export the characteristic pattern for 13 × 13 altogether, and altogether 75
A channel, carries out classification on this basis and position returns；
The input of mesoscale layer is that the characteristic pattern in the 79th layer of 13 × 13,512 channels is carried out convolution operation, generates 13 × 13,256
The characteristic pattern in channel, is then up-sampled, and generates the characteristic pattern in 26 × 26,256 channels, while with the 26 × 26 of the 61st layer,
The characteristic pattern of the mesoscale in 512 channels merges, and carries out convolution operation, exports as the characteristic pattern of 26 × 26 sizes, altogether 75 it is logical
Road, then carries out classification herein and position returns；
Step 4.5: YOLOv3 algorithm model after being improved using training data sample set off-line training.
4. the method according to claim 1, wherein wood ties Defect Detection system, including such as in the step 6
Under:
Image capture module, including lighting system and CCD optical imaging system are swept camera using lighting device and CCD line and are acquired
Long stuff surface image；
Image procossing and identification module, the image procossing and identification module are at host computer PC end, including MATLAB image
Software, Python composing software, database and the communication module communicated with slave computer are managed, Python composing software is utilized
To after processing wood knot flaw image detect, and Defect Detection as the result is shown and export give slave computer movement executing mechanism；
The Python composing software is to utilize open source frame in the Ubuntu system of host computer PC end for realizing YOLOv3 algorithm
After frame TensorFlow realizes improved YOLOv3 algorithm, using training data sample set off-line training algorithm model, and utilize
It is tall and handsome to accelerate training process up to GTX 1080Ti video card；The database is used for batten surface image set of data samples and wood
It ties Defect Detection result and carries out storage management；The communication module uses Ethernet ICP/IP protocol, makes host computer PC end soft
Part is communicated with slave computer movement executing mechanism；
Slave computer Motor execution module, the slave computer movement executing mechanism, according to host computer PC end transmission come wood tie the flaw
The classification and positioning position information of defect, using programmable logic controller (PLC) PLC control wood processing equipment to batten surface blemish
Carry out cutting processing.
5. the method according to claim 1, wherein in the step 1, the optical imaging system be using
Basier-12k line sweeps CCD industrial camera, for acquiring original long stuff surface image.
6. a kind of wood knot Defect Detection system, which is characterized in that shown wood ties Defect Detection system and includes:
Image capture module, for acquiring long stuff surface image；
Image procossing and identification module, the image procossing and identification module are used for the collected long wood of image capture module
Surface image is divided into the identical image of high wide size and carries out data and expands to obtain training data sample set, to expanding
The training data sample set arrived carries out artificial mark wood knot flaw target frame, and wherein tag along sort is two class wood knot of fast knot and slip-knot
Flaw；Dimension cluster is carried out to the wood knot flaw target frame manually marked using k-means++ clustering algorithm, is obtained suitable first
3 scales are reduced to 2 size measurements then in YOLOv3 network by beginning candidate frame number and size, at the same according to contrast and
It normalizes thought and improves loss loss function, utilize YOLOv3 algorithm on-line checking batten exterior view after the good improvement of off-line training
Different flaws as in；The Defect Detection result that on-line checking goes out is exported and gives slave computer movement executing mechanism, Defect Detection knot
Fruit includes the classification and positioning position information of wood knot flaw；
Slave computer Motor execution module, the slave computer movement executing mechanism, according to image procossing and identification module transmit come
Defect Detection as a result, using programmable logic controller (PLC) PLC control wood processing equipment batten surface blemish is cut
Processing.
7. wood knot Defect Detection system according to claim 6, which is characterized in that described image processing is adopted with identification module
With host computer PC end realize, including MATLAB image processing software, Python composing software, database and with slave computer carry out
The communication module of communication；MATLAB image processing software handles the collected long stuff surface image of image capture module
Afterwards, by Python composing software to after processing wood knot flaw image detect, and Defect Detection as the result is shown and export give
Slave computer movement executing mechanism；The Python composing software is at host computer PC end for realizing YOLOv3 algorithm
In Ubuntu system, after realizing improved YOLOv3 algorithm using Open Framework TensorFlow, training data sample set is used
Off-line training algorithm model, and accelerate training process up to GTX 1080Ti video card using tall and handsome；The database is used for wood
Surface image set of data samples and wood knot Defect Detection result carry out storage management；The communication module uses Ethernet
ICP/IP protocol allows host computer PC end software to be communicated with slave computer movement executing mechanism.
8. wood knot Defect Detection system according to claim 7, which is characterized in that described image acquisition module includes illumination
System and CCD optical imaging system.
9. wood knot Defect Detection system according to claim 8, which is characterized in that the CCD optical imaging system includes
Basier-12k line sweeps CCD industrial camera.
10. any wood knot Defect Detection system of claim the 1-5 any method and/or claim 6-9 exists
Treating of wood utilizes the application in technical field.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910529463.XA CN110310259B (en) | 2019-06-19 | 2019-06-19 | Improved YOLOv3 algorithm-based knot defect detection method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910529463.XA CN110310259B (en) | 2019-06-19 | 2019-06-19 | Improved YOLOv3 algorithm-based knot defect detection method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110310259A true CN110310259A (en) | 2019-10-08 |
| CN110310259B CN110310259B (en) | 2021-07-27 |
Family
ID=68077740
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910529463.XA Active CN110310259B (en) | 2019-06-19 | 2019-06-19 | Improved YOLOv3 algorithm-based knot defect detection method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110310259B (en) |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111062915A (en) * | 2019-12-03 | 2020-04-24 | 浙江工业大学 | Real-time steel pipe defect detection method based on improved YOLOv3 model |
| CN111091550A (en) * | 2019-12-12 | 2020-05-01 | 创新奇智(北京)科技有限公司 | Multi-size self-adaptive PCB solder paste area detection system and detection method |
| CN111160440A (en) * | 2019-12-24 | 2020-05-15 | 广东省智能制造研究所 | Helmet wearing detection method and device based on deep learning |
| CN111222474A (en) * | 2020-01-09 | 2020-06-02 | 电子科技大学 | Method for detecting small target of high-resolution image with any scale |
| CN111652853A (en) * | 2020-05-11 | 2020-09-11 | 西南科技大学 | Magnetic powder flaw detection method based on deep convolutional neural network |
| CN111738072A (en) * | 2020-05-15 | 2020-10-02 | 北京百度网讯科技有限公司 | Training method and device of target detection model and electronic equipment |
| CN111899232A (en) * | 2020-07-20 | 2020-11-06 | 广西大学 | Method for nondestructive testing of bamboo-wood composite container bottom plate by utilizing image processing |
| CN112070158A (en) * | 2020-09-08 | 2020-12-11 | 哈尔滨工业大学（威海） | Facial flaw detection method based on convolutional neural network and bilateral filtering |
| CN112184679A (en) * | 2020-09-30 | 2021-01-05 | 佛山市南海区广工大数控装备协同创新研究院 | YOLOv 3-based wine bottle flaw automatic detection method |
| CN112233090A (en) * | 2020-10-15 | 2021-01-15 | 浙江工商大学 | Film flaw detection method based on improved attention mechanism |
| CN112525919A (en) * | 2020-12-21 | 2021-03-19 | 福建新大陆软件工程有限公司 | Wood board defect detection system and method based on deep learning |
| CN112634237A (en) * | 2020-12-25 | 2021-04-09 | 福州大学 | Long bamboo strip surface defect detection method and system based on YOLOv3 improved network |
| CN112633308A (en) * | 2020-09-15 | 2021-04-09 | 北京华电天仁电力控制技术有限公司 | Detection method and detection system for whether power plant operating personnel wear safety belts |
| CN112785574A (en) * | 2021-01-25 | 2021-05-11 | 金陵科技学院 | Scarf pattern defect detection method based on improved YOLOv3 |
| CN112907597A (en) * | 2021-02-04 | 2021-06-04 | 西华师范大学 | Railway track line detection method based on deep convolutional neural network |
| CN113298809A (en) * | 2021-06-25 | 2021-08-24 | 成都飞机工业（集团）有限责任公司 | Composite material ultrasonic image defect detection method based on deep learning and superpixel segmentation |
| CN113449699A (en) * | 2021-08-30 | 2021-09-28 | 上海兴容信息技术有限公司 | Energy efficiency analysis method and system for target object |
| CN113724223A (en) * | 2021-08-27 | 2021-11-30 | 江南大学 | YOLOv3 data set production method and system based on optical microscope |
| CN114240822A (en) * | 2021-12-28 | 2022-03-25 | 中山小池科技有限公司 | Cotton cloth flaw detection method based on YOLOv3 and multi-scale feature fusion |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106596575A (en) * | 2016-12-15 | 2017-04-26 | 南通维新自动化科技有限公司 | Wood plate surface defect detection device |
| CN107392896A (en) * | 2017-07-14 | 2017-11-24 | 佛山市南海区广工大数控装备协同创新研究院 | A kind of Wood Defects Testing method and system based on deep learning |
| CN108520114A (en) * | 2018-03-21 | 2018-09-11 | 华中科技大学 | A kind of textile cloth defect detection model and its training method and application |
-
2019
- 2019-06-19 CN CN201910529463.XA patent/CN110310259B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106596575A (en) * | 2016-12-15 | 2017-04-26 | 南通维新自动化科技有限公司 | Wood plate surface defect detection device |
| CN107392896A (en) * | 2017-07-14 | 2017-11-24 | 佛山市南海区广工大数控装备协同创新研究院 | A kind of Wood Defects Testing method and system based on deep learning |
| CN108520114A (en) * | 2018-03-21 | 2018-09-11 | 华中科技大学 | A kind of textile cloth defect detection model and its training method and application |
Non-Patent Citations (3)
| Title |
|---|
| JOSEPH REDMON等: "YOLOv3: An Incremental Improvement", 《HTTPS://ARXIV.ORG/ABS/1804.02767》 * |
| 司徒伟明 等: "基于OpenCV的木材表面缺陷检测系统", 《现代电子技术》 * |
| 吕铄 等: "基于改进损失函数的YOLOv3网络", 《计算机系统应用》 * |
Cited By (28)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111062915A (en) * | 2019-12-03 | 2020-04-24 | 浙江工业大学 | Real-time steel pipe defect detection method based on improved YOLOv3 model |
| CN111062915B (en) * | 2019-12-03 | 2023-10-24 | 浙江工业大学 | Real-time steel pipe defect detection method based on improved YOLOv3 model |
| CN111091550A (en) * | 2019-12-12 | 2020-05-01 | 创新奇智(北京)科技有限公司 | Multi-size self-adaptive PCB solder paste area detection system and detection method |
| CN111160440A (en) * | 2019-12-24 | 2020-05-15 | 广东省智能制造研究所 | Helmet wearing detection method and device based on deep learning |
| CN111160440B (en) * | 2019-12-24 | 2023-11-21 | 广东省智能制造研究所 | Deep learning-based safety helmet wearing detection method and device |
| CN111222474B (en) * | 2020-01-09 | 2022-11-04 | 电子科技大学 | Method for detecting small target of high-resolution image with any scale |
| CN111222474A (en) * | 2020-01-09 | 2020-06-02 | 电子科技大学 | Method for detecting small target of high-resolution image with any scale |
| CN111652853A (en) * | 2020-05-11 | 2020-09-11 | 西南科技大学 | Magnetic powder flaw detection method based on deep convolutional neural network |
| CN111738072A (en) * | 2020-05-15 | 2020-10-02 | 北京百度网讯科技有限公司 | Training method and device of target detection model and electronic equipment |
| CN111899232A (en) * | 2020-07-20 | 2020-11-06 | 广西大学 | Method for nondestructive testing of bamboo-wood composite container bottom plate by utilizing image processing |
| CN111899232B (en) * | 2020-07-20 | 2023-07-04 | 广西大学 | Method for nondestructive detection of bamboo-wood composite container bottom plate by image processing |
| CN112070158A (en) * | 2020-09-08 | 2020-12-11 | 哈尔滨工业大学（威海） | Facial flaw detection method based on convolutional neural network and bilateral filtering |
| CN112070158B (en) * | 2020-09-08 | 2022-11-15 | 哈尔滨工业大学（威海） | Facial flaw detection method based on convolutional neural network and bilateral filtering |
| CN112633308A (en) * | 2020-09-15 | 2021-04-09 | 北京华电天仁电力控制技术有限公司 | Detection method and detection system for whether power plant operating personnel wear safety belts |
| CN112184679A (en) * | 2020-09-30 | 2021-01-05 | 佛山市南海区广工大数控装备协同创新研究院 | YOLOv 3-based wine bottle flaw automatic detection method |
| CN112233090A (en) * | 2020-10-15 | 2021-01-15 | 浙江工商大学 | Film flaw detection method based on improved attention mechanism |
| CN112233090B (en) * | 2020-10-15 | 2023-05-30 | 浙江工商大学 | Film flaw detection method based on improved attention mechanism |
| CN112525919A (en) * | 2020-12-21 | 2021-03-19 | 福建新大陆软件工程有限公司 | Wood board defect detection system and method based on deep learning |
| CN112634237A (en) * | 2020-12-25 | 2021-04-09 | 福州大学 | Long bamboo strip surface defect detection method and system based on YOLOv3 improved network |
| CN112785574A (en) * | 2021-01-25 | 2021-05-11 | 金陵科技学院 | Scarf pattern defect detection method based on improved YOLOv3 |
| CN112785574B (en) * | 2021-01-25 | 2023-06-06 | 金陵科技学院 | Scarf pattern defect detection method based on improved YOLOv3 |
| CN112907597A (en) * | 2021-02-04 | 2021-06-04 | 西华师范大学 | Railway track line detection method based on deep convolutional neural network |
| CN113298809B (en) * | 2021-06-25 | 2022-04-08 | 成都飞机工业（集团）有限责任公司 | Composite material ultrasonic image defect detection method based on deep learning and superpixel segmentation |
| CN113298809A (en) * | 2021-06-25 | 2021-08-24 | 成都飞机工业（集团）有限责任公司 | Composite material ultrasonic image defect detection method based on deep learning and superpixel segmentation |
| CN113724223A (en) * | 2021-08-27 | 2021-11-30 | 江南大学 | YOLOv3 data set production method and system based on optical microscope |
| CN113449699B (en) * | 2021-08-30 | 2021-12-03 | 上海兴容信息技术有限公司 | Energy efficiency analysis method and system for target object |
| CN113449699A (en) * | 2021-08-30 | 2021-09-28 | 上海兴容信息技术有限公司 | Energy efficiency analysis method and system for target object |
| CN114240822A (en) * | 2021-12-28 | 2022-03-25 | 中山小池科技有限公司 | Cotton cloth flaw detection method based on YOLOv3 and multi-scale feature fusion |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110310259B (en) | 2021-07-27 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110310259A (en) | It is a kind of that flaw detection method is tied based on the wood for improving YOLOv3 algorithm | |
| CN110543878B (en) | Pointer instrument reading identification method based on neural network | |
| CN113160192B (en) | Visual sense-based snow pressing vehicle appearance defect detection method and device under complex background | |
| CN106127204B (en) | A kind of multi-direction meter reading Region detection algorithms of full convolutional neural networks | |
| CN108562589B (en) | Method for detecting surface defects of magnetic circuit material | |
| CN109409365A (en) | It is a kind of that method is identified and positioned to fruit-picking based on depth targets detection | |
| CN109919934B (en) | Liquid crystal panel defect detection method based on multi-source domain deep transfer learning | |
| CN109509187B (en) | Efficient inspection algorithm for small defects in large-resolution cloth images | |
| CN108961235A (en) | A kind of disordered insulator recognition methods based on YOLOv3 network and particle filter algorithm | |
| CN110210362A (en) | A kind of method for traffic sign detection based on convolutional neural networks | |
| CN107316295A (en) | A kind of fabric defects detection method based on deep neural network | |
| CN111540006B (en) | Plant stomata intelligent detection and identification method and system based on deep migration learning | |
| CN111401419A (en) | Improved RetinaNet-based employee dressing specification detection method | |
| CN109063706A (en) | Verbal model training method, character recognition method, device, equipment and medium | |
| CN110135430A (en) | A kind of aluminium mold ID automatic recognition system based on deep neural network | |
| Hu et al. | Defect identification method for poplar veneer based on progressive growing generated adversarial network and MASK R-CNN Model | |
| CN109886947A (en) | The high-tension bus-bar defect inspection method of convolutional neural networks based on region | |
| CN110378239A (en) | A kind of real-time traffic marker detection method based on deep learning | |
| CN110197259A (en) | Wafer defect detection method based on small lot data set Yu deep learning algorithm | |
| CN112634237A (en) | Long bamboo strip surface defect detection method and system based on YOLOv3 improved network | |
| CN110390673A (en) | Cigarette automatic testing method based on deep learning under a kind of monitoring scene | |
| CN109543585A (en) | Underwater optics object detection and recognition method based on convolutional neural networks | |
| CN115829991A (en) | Steel surface defect detection method based on improved YOLOv5s | |
| Hussain et al. | A simple and efficient deep learning-based framework for automatic fruit recognition | |
| CN115272204A (en) | Bearing surface scratch detection method based on machine vision |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |