CN108305242A - A kind of intelligent visual detection method, system and device for industrial production line - Google Patents
A kind of intelligent visual detection method, system and device for industrial production line Download PDFInfo
- Publication number
- CN108305242A CN108305242A CN201710908716.5A CN201710908716A CN108305242A CN 108305242 A CN108305242 A CN 108305242A CN 201710908716 A CN201710908716 A CN 201710908716A CN 108305242 A CN108305242 A CN 108305242A
- Authority
- CN
- China
- Prior art keywords
- parts
- detected
- image block
- image
- sample
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- 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]
Abstract
Include sample image acquisition being carried out to the parts on various types of production line, and be labeled to sample image the invention discloses a kind of intelligent visual detection method, system and device for industrial production line；According to sample image, is handled by RPN networks and SVM classifier, obtain parts detection model；According to obtained parts detection model, parts to be detected are detected, obtain the detection data of parts to be detected.The present invention is by being combined RPN networks and SVM classifier, can also be trained by convolutional neural networks even for the less industrial production line parts of textural characteristics and the smooth packaging bag in surface can indicate the parts detection model of object features, and parts detection model all has good adaptability for illumination, target occlusion and deformation, effectively improve the accuracy rate of detection, meet the requirement detected in real time in industrial production, can be widely applied in industrial production line.
Description
Technical field
The present invention relates to technical field of vision detection more particularly to a kind of intelligent visual detection sides for industrial production line
Method, system and device.
Background technology
In the big production of industry of modernization, more stringent requirements are proposed to context of detection for many emerging industries.For example,
Spare and accessory parts detect in industrial production line, the detection of product packaging printing, image real time monitoring, semiconductor chip packaging detection, print
Printed circuit board positioning etc..In such applications, it has been difficult to meet the need of production and living using the method for traditional artificial detection
It wants, constrains the development and raising of productivity level.This aspect is the person's character due to the mankind, and working long hours, it is tired to easy to produce
Labor can not ensure very high detection accuracy.On the other hand, due to the physical endurance of human eye, it is difficult in speed, precision etc.
It is improved.Therefore, modern industrial production is regarded there is an urgent need to a kind of new intelligent visual detection technology appearance to substitute the mankind
Feel.At the same time, as computer technology, Electromechanical Control technology, intelligent testing technology and digital image processing techniques are constantly sent out
It opens up and perfect, people start the high speed, high-precision, high reliability phase of the intelligent abstracting power of human vision and processor
In conjunction with the production efficiency in raising industrial production line.
Detecting system in existing industrial production line is typically all the method based on shape matching and template matches；This method
It is divided into off-line phase and on-line stage.
Off-line phase mainly pre-processes image, that is, user chooses the ROI containing target object from reference picture
Its template is generated, using Canny filter process template and search image, and the direction vector put on edge is calculated, builds
The image pyramid of template and search image, is used in combination mean filter smothing filtering.
On-line stage be exactly carry out similarity measurements it is flux matched and by least square adjust return pose parameter.Use image
Pyramid technology search strategy.
It is above-mentioned disadvantage of the prior art is that, some matching algorithms based on provincial characteristics are substantially that local window is utilized
Between half-tone information degree of correlation, its flat in topography and texture-rich place can reach higher precision, and can obtain
Fine and close optical parallax field is obtained, and there is very strong noise inhibiting ability.But the matching algorithm be directed to industrial detection in textural characteristics compared with
Few parts and product packaging bags are helpless, due to before matching, first have to template carry out characteristic point extraction and
Pretreatment, and it is difficult to extract effective characteristic point for the parts of local smoother and some packaging bags, and
The gray-scale watermark in a certain size window is chosen in matching process as Matching unit, this just determines the algorithm to scenery
Surface texture and illumination reflection are more sensitive, and when illumination changes, match window size is difficult to select.
Detection based on provincial characteristics matching process on general industry production line mostly selects special label form assembly
On the parts for needing to detect, and increases label template on the parts of required detection and not only add the inspection on production line
Process is surveyed, certain cost is also taken, after marking template to be capped or block due to artificial origin, it is difficult to ensure that detection
Accuracy rate, and the environment detected in industrial production line compares evil and omits, when the parts of required detection are due to illumination, extruding etc.
Influence will also result in prodigious false drop rate.
And the detection of feature based matching process take it is longer, due to the size of given matching template be it is fixed,
During detection, entire search graph is needed to be traversed for, it is best to find out with the maximum image block of Prototype drawing related coefficient
With point, when search graph is bigger, the used time is longer, cannot meet the real-time testing requirements in industrial production line.
Invention content
In order to solve the above-mentioned technical problem, the object of the present invention is to provide it is a kind of can provide Detection accuracy for industry
The intelligent visual detection method, system and device of production line.
The technical solution used in the present invention is：
A kind of intelligent visual detection method for industrial production line includes the following steps：
Sample image acquisition is carried out to the parts on various types of production line, and sample image is labeled；
According to sample image, is handled by RPN networks and SVM classifier, obtain parts detection model；
According to obtained parts detection model, parts to be detected are detected, obtain the inspection of parts to be detected
Measured data.
It is described to sample as a kind of intelligent visual detection further improvements in methods for industrial production line
This image is labeled, the step for specifically include：
Image interception is carried out to the parts of required detection in every sample image, obtains the image block of parts；
According to the obtained image block of interception, obtain the coordinate information of each image block and to the image block in sample image into
Rower is noted, the image block marked.
As a kind of intelligent visual detection further improvements in methods for industrial production line, the basis
Sample image is handled by RPN networks and SVM classifier, obtains parts detection model, the step for include：
RPN convolutional network operation processings are carried out to sample image, obtain the characteristic pattern of last layer；
Preset convolution kernel and obtained characteristic pattern are subjected to convolution, obtain feature vector；
According to the length-width ratio type of the size category of sample image and image block, the image predicted by convolution kernel
Block；
By the last volume base access classification layer of RPN convolutional networks and layer progress probabilistic forecasting is returned, obtains each prediction
Image block prediction probability；
Overlapping ratio calculation will be carried out in the image block of each prediction with the image block marked in sample image, selects overlap ratio
The positive sample that the image block of prediction of the rate more than 0.8 is trained as SVM, the image block of prediction of the selection overlapping ratio less than 0.3
Cast out as trained negative sample, and by the image block for being overlapped prediction of the ratio between 0.3 and 0.8, obtains training sample
Collection；
According to the prediction probability of image block, divide from the highest n image block of middle selection prediction probability of positive sample as SVM
The input of class device carries out the classification based training of variety classes parts, and the figure that training sample is concentrated is obtained by RPN networks when training
As block, the roughing region of parts to be detected is obtained, roughing region, which is then sent to SVM classifier, differentiates, passes through spy
Whether the roughing region for levying vector determination parts to be detected is best region, and position is arrived if so, indicating to be accurately positioned, various
Parts detection model is obtained after the completion of type component training, wherein n is preset value.
As a kind of intelligent visual detection further improvements in methods for industrial production line, the basis
Obtained parts detection model, is detected parts to be detected, obtains the detection data of parts to be detected, this step
Suddenly include：
Image Acquisition is carried out to parts to be detected, obtains part diagram picture to be detected；
Part diagram picture to be detected is loaded into parts detection model, and net is accelerated by GPU parallel processing manners
The propagation of network obtains score and the position data of parts to be detected to get to the detection data of parts to be detected.
Another technical solution of the present invention is：
A kind of intelligent visual detection system for industrial production line, including：
Image acquisition units, for carrying out sample image acquisition to the parts on various types of production line, and to sample graph
As being labeled；
Training unit is modeled, for according to sample image, being handled by RPN networks and SVM classifier, obtains parts inspection
Survey model；
Object detection unit, for according to obtained parts detection model, being detected, obtaining to parts to be detected
The detection data of parts to be detected.
As a kind of being further improved for intelligent visual detection system for industrial production line, described image is adopted
Collection unit specifically includes：
Interception unit carries out image interception for the parts to required detection in every sample image, obtains parts
Image block；
Unit is marked, the image block for being obtained according to interception obtains the coordinate information of each image block and to sample graph
Image block as in is labeled, the image block marked.
As a kind of being further improved for intelligent visual detection system for industrial production line, the modeling instruction
Practicing unit includes：
RPN processing units obtain the feature of last layer for carrying out RPN convolutional network operation processings to sample image
Figure；
Convolution unit obtains feature vector for preset convolution kernel and obtained characteristic pattern to be carried out convolution；
Image block predicting unit, for according to the size category of sample image and the length-width ratio type of image block, passing through volume
The image block that product core is predicted；
Probability prediction unit, for the last volume base access classification layer and recurrence layer of RPN convolutional networks to be carried out probability
Prediction, obtains the prediction probability of the image block of each prediction；
Sample selecting unit, for overlap ratio will to be carried out with the image block marked in sample image in the image block of each prediction
Rate calculates, the positive sample for selecting the image block of prediction of the overlapping ratio more than 0.8 to be trained as SVM, and selection overlapping ratio is less than
Negative sample of the image block of 0.3 prediction as training, and the image block that prediction of the ratio between 0.3 and 0.8 will be overlapped
Cast out, obtains training sample set；
Model training unit, for the prediction probability according to image block, from the middle highest n of selection prediction probability of positive sample
A image block carries out the classification based training of variety classes parts as the input of SVM classifier, is obtained by RPN networks when training
The image block that training sample is concentrated, obtains the roughing region of parts to be detected, roughing region is then sent to SVM classifier
Differentiated, judge whether the roughing region of parts to be detected is best region by feature vector, if so, indicating accurate
Position is navigated to, parts detection model is obtained after the completion of various type component training, wherein n is preset value.
As a kind of being further improved for intelligent visual detection system for industrial production line, the target inspection
Surveying unit includes：
Collecting unit obtains part diagram picture to be detected for carrying out Image Acquisition to parts to be detected；
Detection unit for part diagram picture to be detected to be loaded into parts detection model, and is located parallel by GPU
Reason mode accelerates the propagation of network, obtains score and the position data of parts to be detected to get to the inspection of parts to be detected
Measured data.
Another technical solution of the present invention is：
A kind of intelligent visual detection device for industrial production line, including：
Memory, for storing program；
Processor, for executing described program, described program to be used for industrial production described in the processor execution
The intelligent visual detection method of line.
The beneficial effects of the invention are as follows：
A kind of intelligent visual detection method, system and device for industrial production line of the invention by RPN networks with
SVM classifier is combined, can also even for the less industrial production line parts of textural characteristics and the smooth packaging bag in surface
The parts detection model of object features can be indicated by being trained by convolutional neural networks, be used in combination parts detection model pre-
The exact position of target to be detected is measured, and parts detection model all has very well illumination, target occlusion and deformation
Adaptability, effectively improve the accuracy rate of detection, meet the requirement detected in real time in industrial production.
Description of the drawings
Fig. 1 is a kind of step flow chart of intelligent visual detection method for industrial production line of the invention；
Fig. 2 is a kind of block diagram of intelligent visual detection system for industrial production line of the invention.
Specific implementation mode
The specific implementation mode of the present invention is described further below in conjunction with the accompanying drawings：
With reference to figure 1, a kind of intelligent visual detection method for industrial production line of the invention includes the following steps：
Sample image acquisition is carried out to the parts on various types of production line, and sample image is labeled；
According to sample image, is handled by RPN networks and SVM classifier, obtain parts detection model；
According to obtained parts detection model, parts to be detected are detected, obtain the inspection of parts to be detected
Measured data.
Be further used as preferred embodiment, it is described that sample image is labeled, the step for specifically include：
Image interception is carried out to the parts of required detection in every sample image, obtains the image block of parts；
According to the obtained image block of interception, obtain the coordinate information of each image block and to the image block in sample image into
Rower is noted, the image block marked.
It is further used as preferred embodiment, it is described according to sample image, at RPN networks and SVM classifier
Reason, obtains parts detection model, the step for include：
RPN convolutional network operation processings are carried out to sample image, obtain the characteristic pattern of last layer；
Preset convolution kernel and obtained characteristic pattern are subjected to convolution, obtain feature vector；
According to the length-width ratio type of the size category of sample image and image block, the image predicted by convolution kernel
Block；
By the last volume base access classification layer of RPN convolutional networks and layer progress probabilistic forecasting is returned, obtains each prediction
Image block prediction probability；
Overlapping ratio calculation will be carried out in the image block of each prediction with the image block marked in sample image, selects overlap ratio
The positive sample that the image block of prediction of the rate more than 0.8 is trained as SVM, the image block of prediction of the selection overlapping ratio less than 0.3
Cast out as trained negative sample, and by the image block for being overlapped prediction of the ratio between 0.3 and 0.8, obtains training sample
Collection；
According to the prediction probability of image block, divide from the highest n image block of middle selection prediction probability of positive sample as SVM
The input of class device carries out the classification based training of variety classes parts, and the figure that training sample is concentrated is obtained by RPN networks when training
As block, the roughing region of parts to be detected is obtained, roughing region, which is then sent to SVM classifier, differentiates, passes through spy
Whether the roughing region for levying vector determination parts to be detected is best region, and position is arrived if so, indicating to be accurately positioned, various
Parts detection model is obtained after the completion of type component training, wherein n is preset value.
It is further used as preferred embodiment, the parts detection model that the basis obtains, to zero to be detected
Part is detected, and obtains the detection data of parts to be detected, the step for include：
Image Acquisition is carried out to parts to be detected, obtains part diagram picture to be detected；
Part diagram picture to be detected is loaded into parts detection model, and net is accelerated by GPU parallel processing manners
The propagation of network obtains score and the position data of parts to be detected to get to the detection data of parts to be detected.
The present invention can expand sample image, and in the present embodiment, it is 640*480 only to acquire a small amount of (10) resolution ratio
Sample image, slide into lower right corner interception image block from the upper left corner of parts to be detected on every image, and mark for it
Label name.It obtains the coordinate information of each image block and the image block in sample image is labeled, the image marked
Block can have multiple and different parts on one image, it is only necessary to the image block of detection parts, every image needed for mark
After mark interception, an XML file is generated to record all information of interception image block.Wherein, expand in sample set
In the process, pixel operation transformation increase image block sample set is carried out using to truncated picture block on every image, by figure
As the inverse transform of physics, affine transformation expands picture sample collection, is only needed to image labeling before expanding picture sample collection
Once.
In the present embodiment, zero is obtained using full convolutional neural networks RPN and the method for support vector machines joint training
Part detection model.Assuming that the input picture of given 600*1000, the convolution characteristic pattern of last layer is obtained by full convolution operation
(size is about 40*60), last layer of convolutional layer share 256 characteristic patterns.The volume of 3*3 is used on these obtained characteristic patterns
Product core (sliding window) carries out convolution with characteristic pattern, then can obtain the feature of one 256 dimension after the region convolution of this 3*3
Vector.Because on the region of this 3*3,1 dimensional vector is obtained on each characteristic pattern, 256 performance plots can be obtained 256
Dimensional feature vector.3*3 sliding window center positions, corresponding prediction input part diagram are 3 kinds long as 3 kinds of scales (128,256,512)
Width is than (1:1,1:2,2:1) image block produces the candidate frame of k=9 prediction.I.e. each regions 3*3 can generate 9 in advance
The image block of survey.So for the characteristic pattern of this 40*60, a total of about 20000 (40*60*9) a candidate frame, that is, in advance
Survey 20000 image blocks.Two full articulamentums are linked into the last volume base of network, that is, classify layer and recurrence layer.They divide
Yong Yu frame not classified and be returned.Classify layer include 2 elements, for differentiate be parts and non-parts estimation
Probability.It includes 4 coordinate elements (x, y, w, h), the roughing position for determining parts to return layer.RPN is a full convolution
Neural network, main purpose generate svm classifier and predict required image block.By in the image block of each prediction with sample image
The image block of middle mark carries out overlapping ratio calculation, and the image block of prediction of the overlapping ratio more than 0.8 is selected to be trained as SVM
Positive sample selects negative sample of the image block of prediction of the overlapping ratio less than 0.3 as training, and will be overlapped ratio between 0.3
And the image block of the prediction between 0.8 is cast out, and training sample set is obtained；According to the prediction probability of image block, from positive sample
The classification based training that highest 300 image blocks of prediction probability carry out variety classes parts as the input of SVM classifier is chosen,
RPN networks and SVM are dissolved by sharing last convolution layer parameter in a network, reduce training parameter.It is used when training
Multiple scales and multiple scale base frames, come promoted unconventional scale and ratio parts detection, pass through RPN networks before this
The image block that training sample is concentrated is obtained, the roughing region of parts to be detected is obtained, roughing region is then sent to SVM points
Class device is differentiated, judges whether the roughing region of parts to be detected is best region by feature vector, if so, indicating
It is accurately positioned to position, parts detection model is obtained after the completion of various type component training.
The present embodiment loads good trained network model before, is adopted in real time with high speed camera first in detection process
Collect the image for including parts to be detected and image to be detected is loaded into parts detection model when a given pulse signal
In, during Internet communication, the scoring parameters and location parameter of parts to be detected can be with parallel processings, therefore use GPU simultaneously
The method of row processing accelerates the propagation of network, finally obtains the score and position data of each parts to be detected, is accurately positioned
It detects by actual test to the position of each parts and needs 0.18s per pictures, can meet zero in industrial production line
The real-time requirement detected and count of component.
With reference to figure 2, a kind of intelligent visual detection system for industrial production line of the invention, including：
Image acquisition units, for carrying out sample image acquisition to the parts on various types of production line, and to sample graph
As being labeled；
Training unit is modeled, for according to sample image, being handled by RPN networks and SVM classifier, obtains parts inspection
Survey model；
Object detection unit, for according to obtained parts detection model, being detected, obtaining to parts to be detected
The detection data of parts to be detected.
It is further used as preferred embodiment, described image collecting unit specifically includes：
Interception unit carries out image interception for the parts to required detection in every sample image, obtains parts
Image block；
Unit is marked, the image block for being obtained according to interception obtains the coordinate information of each image block and to sample graph
Image block as in is labeled, the image block marked.
It is further used as preferred embodiment, the modeling training unit includes：
RPN processing units obtain the feature of last layer for carrying out RPN convolutional network operation processings to sample image
Figure；
Convolution unit obtains feature vector for preset convolution kernel and obtained characteristic pattern to be carried out convolution；
Image block predicting unit, for according to the size category of sample image and the length-width ratio type of image block, passing through volume
The image block that product core is predicted；
Probability prediction unit, for the last volume base access classification layer and recurrence layer of RPN convolutional networks to be carried out probability
Prediction, obtains the prediction probability of the image block of each prediction；
Sample selecting unit, for overlap ratio will to be carried out with the image block marked in sample image in the image block of each prediction
Rate calculates, the positive sample for selecting the image block of prediction of the overlapping ratio more than 0.8 to be trained as SVM, and selection overlapping ratio is less than
Negative sample of the image block of 0.3 prediction as training, and the image block that prediction of the ratio between 0.3 and 0.8 will be overlapped
Cast out, obtains training sample set；
Model training unit, for the prediction probability according to image block, from the middle highest n of selection prediction probability of positive sample
A image block carries out the classification based training of variety classes parts as the input of SVM classifier, is obtained by RPN networks when training
The image block that training sample is concentrated, obtains the roughing region of parts to be detected, roughing region is then sent to SVM classifier
Differentiated, judge whether the roughing region of parts to be detected is best region by feature vector, if so, indicating accurate
Position is navigated to, parts detection model is obtained after the completion of various type component training, wherein n is preset value.
It is further used as preferred embodiment, the object detection unit includes：
Collecting unit obtains part diagram picture to be detected for carrying out Image Acquisition to parts to be detected；
Detection unit for part diagram picture to be detected to be loaded into parts detection model, and is located parallel by GPU
Reason mode accelerates the propagation of network, obtains score and the position data of parts to be detected to get to the inspection of parts to be detected
Measured data.
A kind of intelligent visual detection device for industrial production line of the invention, including：
Memory, for storing program；
Processor, for executing described program, described program to be used for industrial production described in the processor execution
The intelligent visual detection method of line.
From the foregoing it can be that a kind of intelligent visual detection method, system and device for industrial production line of the invention
By the way that RPN networks and SVM classifier are combined, even for the less industrial production line parts of textural characteristics and surface light
Sliding packaging bag can also train the parts detection model that can indicate object features by convolutional neural networks, and parts are used in combination
Detection model can predict the exact position of target to be detected, and parts detection model for illumination, target occlusion and
Deformation all has good adaptability, effectively improves the accuracy rate of detection, meets the requirement detected in real time in industrial production.
It is to be illustrated to the preferable implementation of the present invention, but the invention is not limited to the implementation above
Example, those skilled in the art can also make various equivalent variations or be replaced under the premise of without prejudice to spirit of that invention
It changes, these equivalent deformations or replacement are all contained in the application claim limited range.
Claims (9)
1. a kind of intelligent visual detection method for industrial production line, which is characterized in that include the following steps：
Sample image acquisition is carried out to the parts on various types of production line, and sample image is labeled；
According to sample image, is handled by RPN networks and SVM classifier, obtain parts detection model；
According to obtained parts detection model, parts to be detected are detected, obtain the testing number of parts to be detected
According to.
2. a kind of intelligent visual detection method for industrial production line according to claim 1, it is characterised in that：It is described
Sample image is labeled, the step for specifically include：
Image interception is carried out to the parts of required detection in every sample image, obtains the image block of parts；
According to the obtained image block of interception, the coordinate information of each image block is obtained and to the image block in sample image into rower
Note, the image block marked.
3. a kind of intelligent visual detection method for industrial production line according to claim 2, it is characterised in that：It is described
According to sample image, handled by RPN networks and SVM classifier, obtain parts detection model, the step for include：
RPN convolutional network operation processings are carried out to sample image, obtain the characteristic pattern of last layer；
Preset convolution kernel and obtained characteristic pattern are subjected to convolution, obtain feature vector；
According to the length-width ratio type of the size category of sample image and image block, the image block predicted by convolution kernel；
By the last volume base access classification layer of RPN convolutional networks and layer progress probabilistic forecasting is returned, obtains the figure of each prediction
As the prediction probability of block；
Overlapping ratio calculation will be carried out in the image block of each prediction with the image block marked in sample image, selection overlapping ratio is big
In the positive sample that the image block of 0.8 prediction is trained as SVM, select the image block of prediction of the overlapping ratio less than 0.3 as
Trained negative sample, and the image block for being overlapped prediction of the ratio between 0.3 and 0.8 is cast out, obtain training sample set；
According to the prediction probability of image block, from the highest n image block of middle selection prediction probability of positive sample as SVM classifier
Input carry out variety classes parts classification based training, training when by RPN networks obtain training sample concentrate image block,
The roughing region of parts to be detected is obtained, roughing region, which is then sent to SVM classifier, differentiates, passes through feature vector
Judge whether the roughing region of parts to be detected is best region, if so, expression, which is accurately positioned, arrives position, various types of zero
Parts detection model is obtained after the completion of part training, wherein n is preset value.
4. a kind of intelligent visual detection method for industrial production line according to claim 1, it is characterised in that：It is described
The obtained parts detection model of basis, parts to be detected are detected, the detection data of parts to be detected is obtained,
The step for include：
Image Acquisition is carried out to parts to be detected, obtains part diagram picture to be detected；
Part diagram picture to be detected is loaded into parts detection model, and network is accelerated by GPU parallel processing manners
It propagates, obtains score and the position data of parts to be detected to get to the detection data of parts to be detected.
5. a kind of intelligent visual detection system for industrial production line, which is characterized in that including：
Image acquisition units, for on various types of production line parts carry out sample image acquisition, and to sample image into
Rower is noted；
Training unit is modeled, for according to sample image, being handled by RPN networks and SVM classifier, obtains parts detection mould
Type；
Object detection unit, for according to obtained parts detection model, being detected, obtaining to be checked to parts to be detected
Survey the detection data of parts.
6. a kind of intelligent visual detection system for industrial production line according to claim 5, it is characterised in that：It is described
Image acquisition units specifically include：
Interception unit carries out image interception for the parts to required detection in every sample image, obtains the figure of parts
As block；
Unit is marked, the image block for being obtained according to interception obtains the coordinate information of each image block and in sample image
Image block be labeled, the image block marked.
7. a kind of intelligent visual detection system for industrial production line according to claim 6, it is characterised in that：It is described
Modeling training unit includes：
RPN processing units obtain the characteristic pattern of last layer for carrying out RPN convolutional network operation processings to sample image；
Convolution unit obtains feature vector for preset convolution kernel and obtained characteristic pattern to be carried out convolution；
Image block predicting unit, for according to the size category of sample image and the length-width ratio type of image block, passing through convolution kernel
The image block predicted；
Probability prediction unit, for the last volume base access classification layer and recurrence layer of RPN convolutional networks to be carried out probabilistic forecasting,
Obtain the prediction probability of the image block of each prediction；
Sample selecting unit, for overlapping ratiometer will to be carried out with the image block marked in sample image in the image block of each prediction
Calculation, the positive sample for selecting the image block of prediction of the overlapping ratio more than 0.8 to be trained as SVM, selection overlapping ratio are less than 0.3
Negative sample of the image block of prediction as training, and the image block for being overlapped prediction of the ratio between 0.3 and 0.8 is cast out,
Obtain training sample set；
Model training unit, for the prediction probability according to image block, from the highest n figure of the middle selection prediction probability of positive sample
Classification based training as block as the input progress variety classes parts of SVM classifier,
The image block that training sample is concentrated is obtained by RPN networks when training, obtains the roughing region of parts to be detected, then
Roughing region is sent to SVM classifier to differentiate, by feature vector judge parts to be detected roughing region whether
For best region, position is arrived if so, indicating to be accurately positioned, parts detection mould is obtained after the completion of various type component training
Type, wherein n are preset value.
8. a kind of intelligent visual detection system for industrial production line according to claim 5, it is characterised in that：It is described
Object detection unit includes：
Collecting unit obtains part diagram picture to be detected for carrying out Image Acquisition to parts to be detected；
Detection unit for part diagram picture to be detected to be loaded into parts detection model, and passes through the parallel processing sides GPU
Formula accelerates the propagation of network, obtains score and the position data of parts to be detected to get to the testing number of parts to be detected
According to.
9. a kind of intelligent visual detection device for industrial production line, which is characterized in that including：
Memory, for storing program；
Processor, for executing described program, described program makes the processor execute such as any one of Claims 1 to 4 institute
The intelligent visual detection method for industrial production line stated.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710908716.5A CN108305242A (en) | 2017-09-29 | 2017-09-29 | A kind of intelligent visual detection method, system and device for industrial production line |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710908716.5A CN108305242A (en) | 2017-09-29 | 2017-09-29 | A kind of intelligent visual detection method, system and device for industrial production line |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108305242A true CN108305242A (en) | 2018-07-20 |
Family
ID=62869392
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710908716.5A Pending CN108305242A (en) | 2017-09-29 | 2017-09-29 | A kind of intelligent visual detection method, system and device for industrial production line |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108305242A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109304306A (en) * | 2018-09-19 | 2019-02-05 | 广东省智能制造研究所 | Production line articles sorting method, system and articles sorting system |
| CN109454195A (en) * | 2018-10-31 | 2019-03-12 | 安徽巨自动化装备有限公司 | A kind of aluminium vehicle body self-piercing riveting riveting mould failure visible detection method |
| CN112258485A (en) * | 2020-10-27 | 2021-01-22 | 南京清湛人工智能研究院有限公司 | Industrial visual inspection method based on transfer learning and knowledge graph |
| CN116452840A (en) * | 2023-06-19 | 2023-07-18 | 济宁联威车轮制造有限公司 | Automobile part assembly position vision checking system based on numerical control machine |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106599939A (en) * | 2016-12-30 | 2017-04-26 | 深圳市唯特视科技有限公司 | Real-time target detection method based on region convolutional neural network |
| CN107016357A (en) * | 2017-03-23 | 2017-08-04 | 北京工业大学 | A kind of video pedestrian detection method based on time-domain convolutional neural networks |
-
2017
- 2017-09-29 CN CN201710908716.5A patent/CN108305242A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106599939A (en) * | 2016-12-30 | 2017-04-26 | 深圳市唯特视科技有限公司 | Real-time target detection method based on region convolutional neural network |
| CN107016357A (en) * | 2017-03-23 | 2017-08-04 | 北京工业大学 | A kind of video pedestrian detection method based on time-domain convolutional neural networks |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109304306A (en) * | 2018-09-19 | 2019-02-05 | 广东省智能制造研究所 | Production line articles sorting method, system and articles sorting system |
| CN109304306B (en) * | 2018-09-19 | 2020-08-11 | 广东省智能制造研究所 | Production line object sorting method and system and object sorting system |
| CN109454195A (en) * | 2018-10-31 | 2019-03-12 | 安徽巨自动化装备有限公司 | A kind of aluminium vehicle body self-piercing riveting riveting mould failure visible detection method |
| WO2020087725A1 (en) * | 2018-10-31 | 2020-05-07 | 安徽巨一自动化装备有限公司 | Method for visually detecting failure of self-piercing riveting die in aluminum bodywork |
| CN112258485A (en) * | 2020-10-27 | 2021-01-22 | 南京清湛人工智能研究院有限公司 | Industrial visual inspection method based on transfer learning and knowledge graph |
| CN116452840A (en) * | 2023-06-19 | 2023-07-18 | 济宁联威车轮制造有限公司 | Automobile part assembly position vision checking system based on numerical control machine |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Hu et al. | Detection of PCB surface defects with improved faster-RCNN and feature pyramid network | |
| CN108898047B (en) | Pedestrian detection method and system based on blocking and shielding perception | |
| CN107229930B (en) | Intelligent identification method for numerical value of pointer instrument | |
| CN103390164B (en) | Method for checking object based on depth image and its realize device | |
| CN108305242A (en) | A kind of intelligent visual detection method, system and device for industrial production line | |
| JP3549569B2 (en) | Target pattern detection method in video | |
| US20140341421A1 (en) | Method for Detecting Persons Using 1D Depths and 2D Texture | |
| CN104346802B (en) | A kind of personnel leave the post monitoring method and equipment | |
| CN108681693A (en) | Licence plate recognition method based on trusted area | |
| CN108805016A (en) | A kind of head and shoulder method for detecting area and device | |
| CN106018422B (en) | Based on matched special-shaped stamping parts profile defects vision detection system and method | |
| CN111242899B (en) | Image-based flaw detection method and computer-readable storage medium | |
| CN110490842A (en) | A kind of steel strip surface defect detection method based on deep learning | |
| CN109632808A (en) | Seamed edge defect inspection method, device, electronic equipment and storage medium | |
| CN108764338B (en) | Pedestrian tracking method applied to video analysis | |
| CN111127417B (en) | Printing defect detection method based on SIFT feature matching and SSD algorithm improvement | |
| CN109800755A (en) | A kind of remote sensing image small target detecting method based on Analysis On Multi-scale Features | |
| CN110334594A (en) | A kind of object detection method based on batch again YOLO algorithm of standardization processing | |
| CN107480585A (en) | Object detection method based on DPM algorithms | |
| CN109711326A (en) | A kind of video object detection method based on shallow-layer residual error network | |
| CN116385430A (en) | Machine vision flaw detection method, device, medium and equipment | |
| Selamet et al. | Automatic detection and classification of defective areas on metal parts by using adaptive fusion of faster R-CNN and shape from shading | |
| CN110517497A (en) | A kind of road traffic classification method, device, equipment, medium | |
| Li et al. | Autofeeding system for assembling the CBCs on automobile engine based on 3-D vision guidance | |
| CN113822836A (en) | Method of marking an image |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180720 |