CN112819824A - Neural network of visual inspection system, and inspection system and method including the same - Google Patents
Neural network of visual inspection system, and inspection system and method including the same Download PDFInfo
- Publication number
- CN112819824A CN112819824A CN202110233030.7A CN202110233030A CN112819824A CN 112819824 A CN112819824 A CN 112819824A CN 202110233030 A CN202110233030 A CN 202110233030A CN 112819824 A CN112819824 A CN 112819824A
- Authority
- CN
- China
- Prior art keywords
- feature matrix
- layer
- ceramic substrate
- inspection
- size
- 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
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 31
- 238000000034 method Methods 0.000 title claims abstract description 17
- 238000007689 inspection Methods 0.000 title claims description 41
- 238000011179 visual inspection Methods 0.000 title claims description 19
- 238000001514 detection method Methods 0.000 claims abstract description 92
- 239000000919 ceramic Substances 0.000 claims abstract description 75
- 239000000758 substrate Substances 0.000 claims abstract description 75
- 238000007599 discharging Methods 0.000 claims abstract description 11
- 239000011159 matrix material Substances 0.000 claims description 105
- 238000005070 sampling Methods 0.000 claims description 25
- 230000005284 excitation Effects 0.000 claims description 15
- 238000004364 calculation method Methods 0.000 claims description 10
- 238000010606 normalization Methods 0.000 claims description 10
- 230000002950 deficient Effects 0.000 claims description 9
- 239000000284 extract Substances 0.000 claims description 9
- 230000007547 defect Effects 0.000 claims description 5
- 239000006185 dispersion Substances 0.000 claims description 5
- 238000011084 recovery Methods 0.000 claims description 4
- 230000000007 visual effect Effects 0.000 abstract description 9
- 238000011027 product recovery Methods 0.000 description 6
- 238000004519 manufacturing process Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000026676 system process Effects 0.000 description 1
Images
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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Quality & Reliability (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Investigating Materials By The Use Of Optical Means Adapted For Particular Applications (AREA)
Abstract
The invention discloses a neural network of a visual detection system, and a detection system and a method comprising the neural network. The automatic detection device comprises a feeding portion and a discharging portion, a detection portion is arranged between the feeding portion and the discharging portion, the detection portion comprises a first detection area, a second detection area and a third detection area, the first detection area comprises a first detection camera and a first portal frame, a first rail is arranged on the first portal frame, the automatic detection device further comprises a first slider connected with the first rail in a sliding mode, a first sucker connected with the first slider, the first sucker is used for adsorbing a ceramic substrate to be detected, the first detection camera is arranged upwards, and the moving track of the first sucker is located above the first detection camera. The visual detection system of the invention has high detection precision and high detection speed.
Description
Technical Field
The invention relates to the technical field of visual detection, in particular to a neural network of a visual detection system, and a detection system and a detection method comprising the neural network.
Background
In the production process of the ceramic substrate, due to uncertainty of the production process, the ceramic substrate may have cracks or have insufficient flatness on the side surface, and thus, the appearance of the ceramic substrate needs to be inspected. Although a detection system for identifying by using a CCD camera is available at present, the existing detection system cannot accurately identify defects on the surface of a ceramic substrate. Moreover, when each surface of the ceramic substrate needs to be detected, most of the existing visual detection systems have low detection efficiency or need more manual intervention, so that the detection efficiency of the ceramic substrate is not high, and the detection precision is low.
Disclosure of Invention
In order to overcome the above disadvantages, an object of the present invention is to provide a visual inspection system for a ceramic substrate, which can rapidly photograph the upper and lower surfaces and the respective sides of the ceramic substrate, greatly improve the inspection efficiency of the ceramic substrate, and accurately calculate and identify the defects of the ceramic substrate.
The neural network for a visual inspection system of a ceramic substrate of the present invention comprises:
(1) the input layer is used for inputting pictures with the size of 224 x 224, reducing the size of a feature matrix by using a convolution layer with a convolution kernel of 3 and the step length of 2 for the input images, and then extracting the feature matrix D1 by using a batch normalization and excitation function RELU 6;
(2) extracting the features of the image by using the down-sampling layer for multiple times, and reducing the size of the feature matrix to obtain feature matrices D2, D3, D4 and D5 respectively;
(3) processing the feature matrix D2 obtained in the step (2) by using an upper sampling layer to obtain S1_1, and then adding the S1_1 and the D1 to realize cross-layer connection;
(4) obtaining S1_2, S1_3, S2_1, S2_2 and C3, respectively, using the same method as step (3);
(5) calculating the feature matrix D4 obtained in the step (2) by using a soft attention calculation layer to obtain a feature matrix A1, multiplying the feature matrix A1 and the feature matrix D5 obtained after the feature matrix D passes through an upper sampling layer to realize weight assignment of features, and obtaining a feature matrix U4;
(6) calculating the feature matrix S2_2 obtained in the step (4) by using a soft attention calculation layer to obtain a feature matrix A2, multiplying the feature matrix obtained after the A2 and the D3 pass through an upper sampling layer to realize weight assignment of features, and obtaining a feature matrix U2;
(7) adding a feature matrix obtained by upsampling U2 and S1_3 to realize cross-layer connection and generating a feature matrix U1;
(8) scaling the feature matrix size to be consistent with the input matrix size using quadratic linear interpolation for U1, and generating the final feature matrix using convolution layer with convolution kernel of 1,
the downsampling layer is constructed by first reducing the size of the feature matrix using a convolutional layer with a convolution kernel of 3 and step size 2, then extracting the feature map using the batch normalization and excitation function RELU6, then further extracting the features using a convolutional layer with a convolution kernel of 3 and step size 1,
the up-sampling layer firstly enlarges the size of a two-time characteristic matrix by using a quadratic linear interpolation method, then extracts a characteristic map by using a convolution layer with a convolution kernel of 1, and finally reduces gradient dispersion by using batch standardization and an excitation function RELU6 to improve the training speed;
the soft attention computation layer extracts feature matrices by first using convolution layers with convolution kernel 3 and then computing feature weights using sigmoid function.
The visual inspection system of the ceramic substrate based on the neural network comprises an inspection part and a server, wherein the server comprises the neural network, and the server is used for processing the image obtained by the inspection part and inputting the preprocessed image into the neural network for defect inspection.
The detection device comprises a feeding portion and a discharging portion, a detection portion is arranged between the feeding portion and the discharging portion, the detection portion comprises a first detection area, a second detection area and a third detection area, the first detection area comprises a first detection camera and a first portal frame, a first rail is arranged on the first portal frame, a first slider connected with the first rail in a sliding mode is further included, a first sucker connected with the first slider is further included, the first sucker is used for adsorbing a ceramic substrate to be detected, the first detection camera is arranged upwards, and the first sucker can pass through the position above the first detection camera.
Furthermore, the first driving portion is used for driving the first sliding block to move, the first driving portion comprises a first driving motor and a first screw rod which is coaxially connected with the first driving motor, the first screw rod is in threaded connection with the first sliding block and penetrates through the middle of the first sliding block, and a clamping joint which is clamped with the first rail is arranged on the first sliding block.
Furthermore, the second detection area comprises a second portal frame, a second track is arranged on the second portal frame, the second detection area further comprises a second sliding block in sliding connection with the second track, a rotatable second sucking disc is arranged on the second sliding block, and a pair of second detection cameras are oppositely arranged on two sides of the second track.
Still further, the rotary air cylinder is connected with the second sliding block, and a rotary head of the rotary air cylinder penetrates through the second sliding block and is connected with the second suction cup.
Furthermore, the third detection area comprises a third detection camera arranged at the upper end of the second portal frame, and the second sliding block can pass below the third detection camera.
Furthermore, the feeding part comprises at least one feeding disc and a feeding manipulator, the feeding manipulator is located between the first portal frame and the second portal frame, and the discharging part comprises a discharging manipulator close to the third detection area, a good product recovery disc and a defective product recovery disc.
The invention discloses a visual detection method of a ceramic substrate based on a neural network, which comprises the following steps:
s1, moving the ceramic substrate to be detected to be opposite to the detection camera, taking at least two pictures of any one surface of the ceramic substrate, and transmitting the pictures to the machine;
s2, inputting the image of the ceramic substrate into a neural network of a machine table for image processing, and identifying defective products;
s3, moving each surface of the ceramic substrate to be opposite to the detection camera in sequence, taking more than two pictures of each surface, transmitting the pictures to the machine, inputting the pictures into the neural network of the machine for image processing and identifying defective products,
the neural network includes:
a neural network for a visual inspection system for a ceramic substrate, comprising:
(1) the input layer is used for inputting pictures with the size of 224 x 224, reducing the size of a feature matrix by using a convolution layer with a convolution kernel of 3 and the step length of 2 for the input images, and then extracting the feature matrix D1 by using a batch normalization and excitation function RELU 6;
(2) extracting the features of the image by using the down-sampling layer for multiple times, and reducing the size of the feature matrix to obtain feature matrices D2, D3, D4 and D5 respectively;
(3) processing the feature matrix D2 obtained in the step (2) by using an upper sampling layer to obtain S1_1, and then adding the S1_1 and the D1 to realize cross-layer connection;
(4) obtaining S1_2, S1_3, S2_1, S2_2 and C3, respectively, using the same method as step (3);
(5) calculating the feature matrix D4 obtained in the step (2) by using a soft attention calculation layer to obtain a feature matrix A1, multiplying the feature matrix A1 and the feature matrix D5 obtained after the feature matrix D passes through an upper sampling layer to realize weight assignment of features, and obtaining a feature matrix U4;
(6) calculating the feature matrix S2_2 obtained in the step (4) by using a soft attention calculation layer to obtain a feature matrix A2, multiplying the feature matrix obtained after the A2 and the D3 pass through an upper sampling layer to realize weight assignment of features, and obtaining a feature matrix U2;
(7) adding a feature matrix obtained by upsampling U2 and S1_3 to realize cross-layer connection and generating a feature matrix U1;
(8) scaling the feature matrix size to be consistent with the input matrix size using quadratic linear interpolation for U1, and generating the final feature matrix using convolution layer with convolution kernel of 1,
the downsampling layer is constructed by first reducing the size of the feature matrix using a convolutional layer with a convolution kernel of 3 and step size 2, then extracting the feature map using the batch normalization and excitation function RELU6, then further extracting the features using a convolutional layer with a convolution kernel of 3 and step size 1,
the up-sampling layer firstly enlarges the size of a two-time characteristic matrix by using a quadratic linear interpolation method, then extracts a characteristic map by using a convolution layer with a convolution kernel of 1, and finally reduces gradient dispersion by using batch standardization and an excitation function RELU6 to improve the training speed;
the soft attention computation layer extracts feature matrices by first using convolution layers with convolution kernel 3 and then computing feature weights using sigmoid function.
Further, the inspection cameras include a first inspection camera, a second inspection camera, and a third inspection camera,
in the step S1, the ceramic substrate to be detected is moved to a position opposite to the first detection camera, and the lower surface of the ceramic substrate is photographed;
in the step S3, the ceramic substrate can be moved to a position opposite to a second detection camera, and the second detection camera takes a picture of the side of the ceramic substrate; the ceramic substrate can be moved to a position opposite to the third detection camera, and the upper surface of the ceramic substrate is photographed.
The invention has the following beneficial effects:
the visual detection system can select all unqualified products, zero missing detection is realized, qualified products can be prevented from being regarded as unqualified products, and the misjudgment rate is effectively reduced;
secondly, the visual detection system can accurately judge and calculate the defects of unqualified products, and the detection precision is high;
secondly, the visual detection system can greatly improve the detection speed and effectively save the labor cost.
Drawings
FIG. 1 is a diagram of a neural network according to a first embodiment;
FIG. 2 is a schematic top view of a second embodiment of the present invention;
fig. 3 is a schematic perspective view of a second gantry and a second driving part according to a second embodiment of the present invention.
In the figure:
1. a feeding part; 21. a first detection zone; 211. a first detection camera; 212. a first gantry; 213. a first track; 214. a first slider; 215. a first suction cup; 216. a first drive motor; 217. a first lead screw; 22. a second detection zone; 221. a second gantry; 222. a second track; 224. a second slider; 225. a second detection camera; 226. a second drive motor; 227. a second lead screw; 228. a second suction cup; 23. a third detection zone; 231. a third detection camera; 24. a rotating cylinder; 25. feeding a material plate; 26. a feeding manipulator; 3. a discharge part; 31. a feeding manipulator; 32. a good product recovery tray; 33. and a defective product recovery tray.
Detailed Description
The following detailed description of the preferred embodiments of the present invention, taken in conjunction with the accompanying drawings, will make the advantages and features of the invention easier to understand by those skilled in the art, and thus will clearly and clearly define the scope of the invention.
The first embodiment is as follows:
the embodiment is a visual inspection method of a ceramic substrate based on a neural network, which comprises the following steps:
s1, moving the ceramic substrate to be detected to be opposite to the detection camera, taking at least two pictures of any one surface of the ceramic substrate, and transmitting the pictures to the machine;
s2, inputting the image of the ceramic substrate into a neural network of a machine table for image processing, and identifying defective products;
and S3, sequentially moving each surface of the ceramic substrate to be opposite to the detection camera, respectively taking more than two pictures of each surface, transmitting the pictures to the machine, inputting the pictures into a neural network of the machine for image processing, and identifying defective products.
Referring to fig. 1, the neural network of the present embodiment includes:
(1) the input layer is used for inputting pictures with the size of 224 x 224, reducing the size of the feature matrix by using a convolution layer with the convolution kernel of 3 and the step length of 2 for the input images, and then extracting the feature matrix D1 by using a batch normalization and excitation function RELU 6;
(2) extracting the features of the image by using the down-sampling layer for multiple times, and reducing the size of the feature matrix to obtain feature matrices D2, D3, D4 and D5 respectively;
(3) processing the feature matrix D2 obtained in the step (2) by using an upper sampling layer to obtain S1_1, and then adding the S1_1 and the D1 to realize cross-layer connection;
(4) obtaining S1_2, S1_3, S2_1, S2_2 and C3, respectively, using the same method as step (3);
(5) calculating the feature matrix D4 obtained in the step (2) by using a soft attention calculation layer to obtain a feature matrix A1, multiplying the feature matrix A1 and the feature matrix D5 obtained after the feature matrix D passes through an upper sampling layer to realize weight assignment of features, and obtaining a feature matrix U4;
(6) calculating the feature matrix S2_2 obtained in the step (4) by using a soft attention calculation layer to obtain a feature matrix A2, multiplying the feature matrix obtained after the A2 and the D3 pass through an upper sampling layer to realize weight assignment of features, and obtaining a feature matrix U2;
(7) adding a feature matrix obtained by upsampling U2 and S1_3 to realize cross-layer connection and generating a feature matrix U1;
(8) scaling the feature matrix size to be consistent with the input matrix size using quadratic linear interpolation for U1, and generating the final feature matrix using convolution layer with convolution kernel of 1,
the downsampling layer of the present invention is made by first reducing the size of the feature matrix using a convolutional layer with a convolution kernel of 3 and step size of 2, then extracting the feature map using the batch normalization and excitation function RELU6, and then further extracting the features using a convolutional layer with a convolution kernel of 3 and step size of 1.
The up-sampling layer of the invention firstly uses a quadratic linear interpolation method to enlarge the size of a two-times characteristic matrix, then uses a convolution layer with a convolution kernel of 1 to extract a characteristic map, and finally uses batch standardization and an excitation function RELU6 to reduce gradient dispersion and improve training speed.
The soft attention computation layer of the present invention extracts feature matrices by first using a convolution layer with convolution kernel 3 and then calculates feature weights using sigmoid function.
Example two:
referring to fig. 2 and 3, as a preferred embodiment of the visual inspection system for ceramic substrates according to the present invention, the visual inspection system comprises a loading portion 1, a detecting portion and a discharging portion 3, which are sequentially arranged, wherein the detecting portion comprises a first detecting area 21, a second detecting area 22 and a third detecting area 23, which are sequentially arranged, the first detecting area 21 is used for detecting the lower surface of the ceramic substrate, the second detecting area 22 is used for detecting the side surface of the ceramic substrate, and the third detecting area 23 is used for detecting the upper surface of the ceramic substrate.
The first detection area 21 of the invention comprises a first detection camera 211 and a first portal frame 212, wherein a first track 213 is arranged in the middle of the first portal frame 212 in the height direction along the width direction, the first detection area further comprises a first slide block 214 in sliding connection with the first track 213, a first suction disc 215 is arranged on the lower surface of the first slide block 214, the first suction disc 215 is connected with an external air source and is used for adsorbing a ceramic substrate to be detected, the first detection camera 211 is arranged upwards, and the moving track of the first suction disc 215 is positioned above the first detection camera 211. Thus, the first inspection camera 211 performs visual inspection of the ceramic substrate while the ceramic substrate is moved above the first inspection camera 211.
The first detection area 21 of the present invention further includes a first driving portion for driving the first sliding block 214 to move, the first driving portion includes a first driving motor 216 and a first lead screw 217 coaxially connected to the first driving motor 216, the first lead screw 217 is screwed to the first sliding block 214 and passes through the middle of the first sliding block 214, and a clamping joint engaged with the first track 213 is disposed on the first sliding block 214. The first lead screw 217 is disposed in a direction parallel to the first rail 213. Thus, the first slider 214 can be repeatedly moved along the first track 213 by the first driving motor 216.
The second detection area 22 of the present invention includes a second gantry 221, and the second gantry 221 of the present embodiment is disposed in a direction perpendicular to the first gantry 212. A second rail 222 is provided at the middle portion of the second portal frame 221 along the width direction thereof, and further includes a second slider 224 slidably connected to the second rail 222, a rotatable second suction cup 228 is provided on the second slider 224, and a pair of second detection cameras 225 is provided at both sides of the second rail 222 in an opposing manner. The second chuck 228 has a height higher than that of the second rail 222, thereby facilitating photographing of the ceramic substrate by the second inspection cameras 225, and the two second inspection cameras 225 are located at the same distance from the ceramic substrate on the second chuck 228. The second slider 224 of the second detection area 22 of the present embodiment is driven by a second driving portion to move, the structure of the second driving portion of the present embodiment is similar to that of the first driving portion, and the second driving portion includes a second driving motor 226 and a second lead screw 227 coaxially connected to the second driving motor.
The present invention further includes a rotation cylinder 24 disposed on the lower surface of the second slider 224, a through hole for passing a rotation head of the rotation cylinder 24 is disposed in the middle of the second slider 224, and the rotation cylinder 24 is connected to the second suction cup 228. Therefore, the ceramic substrate can be rotated by the rotary cylinder 24. Therefore, after the second inspection camera 225 photographs a pair of oppositely disposed side surfaces of the ceramic substrate, the second chuck 228 is rotated by 90 degrees by the rotation cylinder 24, and thus the second inspection camera 225 photographs the other two side surfaces of the ceramic substrate, thereby realizing image recognition of the four side surfaces of the ceramic substrate.
The third detection area 23 of the present invention includes a third detection camera 231 disposed at the upper end of the second gantry 221, and the third detection camera 231 is located above the moving track of the second slider 224. Thus, when the second slider 224 moves below the third inspection camera 231, the third inspection camera 231 photographs the upper surface of the ceramic substrate.
The loading part 1 of the present invention comprises at least one loading tray 25 and a loading robot 26, the loading robot 26 being located in the middle of the first gantry 212 and the second gantry 221. The loading robot 26 is used to move the ceramic substrate from the loading tray 25 to the first chuck 215, and can move from the first chuck 215 to be located on the second chuck 228. The discharge unit 3 of the present invention includes a feeding robot 31, a good product recovery tray 32, and a defective product recovery tray 33, which are close to the third detection area 23.
The working process of the visual inspection system of the ceramic substrate comprises the following steps: the loading manipulator 26 moves the ceramic substrate from the loading tray 25 to the lower part of the first sucker 215, after the first sucker 215 sucks the ceramic substrate, the first sucker 215 moves to the upper part of the first detection camera 211 along the first track 213 under the action of the first driving motor 216, and the first detection camera 211 performs photographing processing; subsequently, the first chuck 215 continues to move to the end of the first rail 213, the loading robot 26 moves the ceramic substrate above the second chuck 228, and the second chuck 228 sucks the ceramic substrate; subsequently, the second suction cup 228 is driven by the second driving motor 226 to move to the middle of the two second detection cameras 225, and the second detection cameras 225 perform photographing processing on a pair of oppositely arranged side surfaces of the ceramic substrate; subsequently, under the driving of the rotary cylinder 24, the second suction cup 228 rotates 90 degrees, and the second detection camera 225 photographs the other pair of side surfaces of the ceramic substrate; subsequently, the second suction cup 228 continues to move to a position below the third detection camera 231, and the third detection camera 231 performs photographing processing on the ceramic substrate; the detection system processes the pictures obtained by photographing each surface of the ceramic substrate and judges the quality of the ceramic substrate; subsequently, the second suction cup 228 is moved to the end of the second rail 222 by the second driving motor 226, and is placed on the corresponding good recovery tray 32 and the bad recovery tray 33 by the feeding robot 31.
The above embodiments are merely illustrative of the technical concept and features of the present invention, and the present invention is not limited thereto, and any equivalent changes or modifications made according to the spirit of the present invention should be included in the scope of the present invention.
Claims (10)
1. A neural network for a visual inspection system for a ceramic substrate, comprising:
(1) the input layer is used for inputting pictures with the size of 224 x 224, reducing the size of a feature matrix by using a convolution layer with a convolution kernel of 3 and the step length of 2 for the input images, and then extracting the feature matrix D1 by using a batch normalization and excitation function RELU 6;
(2) extracting the features of the image by using the down-sampling layer for multiple times, and reducing the size of the feature matrix to obtain feature matrices D2, D3, D4 and D5 respectively;
(3) processing the feature matrix D2 obtained in the step (2) by using an upper sampling layer to obtain S1_1, and then adding the S1_1 and the D1 to realize cross-layer connection;
(4) obtaining S1_2, S1_3, S2_1, S2_2 and C3, respectively, using the same method as step (3);
(5) calculating the feature matrix D4 obtained in the step (2) by using a soft attention calculation layer to obtain a feature matrix A1, multiplying the feature matrix A1 and the feature matrix D5 obtained after the feature matrix D passes through an upper sampling layer to realize weight assignment of features, and obtaining a feature matrix U4;
(6) calculating the feature matrix S2_2 obtained in the step (4) by using a soft attention calculation layer to obtain a feature matrix A2, multiplying the feature matrix obtained after the A2 and the D3 pass through an upper sampling layer to realize weight assignment of features, and obtaining a feature matrix U2;
(7) adding a feature matrix obtained by upsampling U2 and S1_3 to realize cross-layer connection and generating a feature matrix U1;
(8) scaling the feature matrix size to be consistent with the input matrix size using quadratic linear interpolation for U1, and generating the final feature matrix using convolution layer with convolution kernel of 1,
the downsampling layer is constructed by first reducing the size of the feature matrix using a convolutional layer with a convolution kernel of 3 and step size 2, then extracting the feature map using the batch normalization and excitation function RELU6, then further extracting the features using a convolutional layer with a convolution kernel of 3 and step size 1,
the up-sampling layer firstly enlarges the size of a two-time characteristic matrix by using a quadratic linear interpolation method, then extracts a characteristic map by using a convolution layer with a convolution kernel of 1, and finally reduces gradient dispersion by using batch standardization and an excitation function RELU6 to improve the training speed;
the soft attention computation layer extracts feature matrices by first using convolution layers with convolution kernel 3 and then computing feature weights using sigmoid function.
2. A visual inspection system for ceramic substrates based on a neural network, comprising an inspection part and a server, wherein the server comprises the neural network of claim 1, and the server is used for processing the image obtained by the inspection part and inputting the preprocessed image into the neural network for defect inspection.
3. The visual inspection system for ceramic substrates based on neural networks as claimed in claim 2, further comprising a loading portion and a discharging portion, wherein the inspection portion is disposed between the loading portion and the discharging portion, the inspection portion comprises a first inspection area, a second inspection area and a third inspection area, the first inspection area comprises a first inspection camera and a first portal frame, a first rail is disposed on the first portal frame, the first inspection area further comprises a first slider connected with the first rail in a sliding manner, the first inspection area further comprises a first sucker connected with the first slider, the first sucker is used for adsorbing the ceramic substrate to be inspected, the first inspection camera is disposed upward, and the first sucker can pass over the first inspection camera.
4. The visual inspection system of a ceramic substrate as claimed in claim 3, further comprising a first driving portion for driving the first slide block to move, wherein the first driving portion comprises a first driving motor and a first lead screw coaxially connected to the first driving motor, the first lead screw is in threaded connection with the first slide block and passes through the middle of the first slide block, and a clamping head engaged with the first rail is disposed on the first slide block.
5. The system of claim 4, wherein the second inspection area comprises a second gantry, a second rail is disposed on the second gantry, the system further comprises a second slider slidably connected to the second rail, a second rotatable suction cup is disposed on the second slider, and a pair of second inspection cameras is disposed on two opposite sides of the second rail.
6. The system of claim 5, further comprising a rotary cylinder coupled to the second slide, wherein a spin head of the rotary cylinder passes through the second slide and is coupled to a second chuck.
7. The system of claim 6, wherein the third inspection area comprises a third inspection camera disposed at an upper end of the second gantry, and the second slide block can pass under the third inspection camera.
8. The visual inspection system of ceramic substrates of claim 7, wherein the feeding portion comprises at least one feeding tray and a feeding manipulator, the feeding manipulator is located between the first portal frame and the second portal frame, and the discharging portion comprises a discharging manipulator, a good recovery tray and a bad recovery tray near the third detection area.
9. A visual inspection method of a ceramic substrate based on a neural network is characterized by comprising the following steps:
s1, moving the ceramic substrate to be detected to be opposite to the detection camera, taking at least two pictures of any one surface of the ceramic substrate, and transmitting the pictures to the machine;
s2, inputting the image of the ceramic substrate into a neural network of a machine table for image processing, and identifying defective products;
s3, moving each surface of the ceramic substrate to be opposite to the detection camera in sequence, taking more than two pictures of each surface, transmitting the pictures to the machine, inputting the pictures into the neural network of the machine for image processing and identifying defective products,
the neural network includes:
a neural network for a visual inspection system for a ceramic substrate, comprising:
(1) the input layer is used for inputting pictures with the size of 224 x 224, reducing the size of a feature matrix by using a convolution layer with a convolution kernel of 3 and the step length of 2 for the input images, and then extracting the feature matrix D1 by using a batch normalization and excitation function RELU 6;
(2) extracting the features of the image by using the down-sampling layer for multiple times, and reducing the size of the feature matrix to obtain feature matrices D2, D3, D4 and D5 respectively;
(3) processing the feature matrix D2 obtained in the step (2) by using an upper sampling layer to obtain S1_1, and then adding the S1_1 and the D1 to realize cross-layer connection;
(4) obtaining S1_2, S1_3, S2_1, S2_2 and C3, respectively, using the same method as step (3);
(5) calculating the feature matrix D4 obtained in the step (2) by using a soft attention calculation layer to obtain a feature matrix A1, multiplying the feature matrix A1 and the feature matrix D5 obtained after the feature matrix D passes through an upper sampling layer to realize weight assignment of features, and obtaining a feature matrix U4;
(6) calculating the feature matrix S2_2 obtained in the step (4) by using a soft attention calculation layer to obtain a feature matrix A2, multiplying the feature matrix obtained after the A2 and the D3 pass through an upper sampling layer to realize weight assignment of features, and obtaining a feature matrix U2;
(7) adding a feature matrix obtained by upsampling U2 and S1_3 to realize cross-layer connection and generating a feature matrix U1;
(8) scaling the feature matrix size to be consistent with the input matrix size using quadratic linear interpolation for U1, and generating the final feature matrix using convolution layer with convolution kernel of 1,
the downsampling layer is constructed by first reducing the size of the feature matrix using a convolutional layer with a convolution kernel of 3 and step size 2, then extracting the feature map using the batch normalization and excitation function RELU6, then further extracting the features using a convolutional layer with a convolution kernel of 3 and step size 1,
the up-sampling layer firstly enlarges the size of a two-time characteristic matrix by using a quadratic linear interpolation method, then extracts a characteristic map by using a convolution layer with a convolution kernel of 1, and finally reduces gradient dispersion by using batch standardization and an excitation function RELU6 to improve the training speed;
the soft attention computation layer extracts feature matrices by first using convolution layers with convolution kernel 3 and then computing feature weights using sigmoid function.
10. The neural network-based visual inspection method of ceramic substrate according to claim 9, wherein the inspection camera includes a first inspection camera, a second inspection camera, and a third inspection camera,
in the step S1, the ceramic substrate to be detected is moved to a position opposite to the first detection camera, and the lower surface of the ceramic substrate is photographed;
in the step S3, the ceramic substrate can be moved to a position opposite to a second detection camera, and the second detection camera takes a picture of the side of the ceramic substrate; the ceramic substrate can be moved to a position opposite to the third detection camera, and the upper surface of the ceramic substrate is photographed.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110233030.7A CN112819824B (en) | 2021-03-03 | 2021-03-03 | Neural network for vision inspection system, inspection system and method including the same |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110233030.7A CN112819824B (en) | 2021-03-03 | 2021-03-03 | Neural network for vision inspection system, inspection system and method including the same |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112819824A true CN112819824A (en) | 2021-05-18 |
| CN112819824B CN112819824B (en) | 2024-04-26 |
Family
ID=75862623
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110233030.7A Active CN112819824B (en) | 2021-03-03 | 2021-03-03 | Neural network for vision inspection system, inspection system and method including the same |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112819824B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113239945A (en) * | 2021-07-12 | 2021-08-10 | 中国人民解放军国防科技大学 | Infrared weak and small target detection method utilizing trajectory constraint |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH0644378A (en) * | 1992-07-23 | 1994-02-18 | Matsushita Electric Ind Co Ltd | Holonic visual recognition device |
| CN105891215A (en) * | 2016-03-31 | 2016-08-24 | 浙江工业大学 | Welding visual detection method and device based on convolutional neural network |
| CN108074231A (en) * | 2017-12-18 | 2018-05-25 | 浙江工业大学 | Magnetic sheet surface defect detection method based on convolutional neural network |
| CN110059728A (en) * | 2019-03-26 | 2019-07-26 | 浙江科技学院 | RGB-D image vision conspicuousness detection method based on attention model |
| CN215416732U (en) * | 2021-03-03 | 2022-01-04 | 科为升视觉技术（苏州）有限公司 | Visual inspection system of ceramic substrate |
-
2021
- 2021-03-03 CN CN202110233030.7A patent/CN112819824B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH0644378A (en) * | 1992-07-23 | 1994-02-18 | Matsushita Electric Ind Co Ltd | Holonic visual recognition device |
| CN105891215A (en) * | 2016-03-31 | 2016-08-24 | 浙江工业大学 | Welding visual detection method and device based on convolutional neural network |
| CN108074231A (en) * | 2017-12-18 | 2018-05-25 | 浙江工业大学 | Magnetic sheet surface defect detection method based on convolutional neural network |
| CN110059728A (en) * | 2019-03-26 | 2019-07-26 | 浙江科技学院 | RGB-D image vision conspicuousness detection method based on attention model |
| CN215416732U (en) * | 2021-03-03 | 2022-01-04 | 科为升视觉技术（苏州）有限公司 | Visual inspection system of ceramic substrate |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113239945A (en) * | 2021-07-12 | 2021-08-10 | 中国人民解放军国防科技大学 | Infrared weak and small target detection method utilizing trajectory constraint |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112819824B (en) | 2024-04-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| EP1112550B1 (en) | An automated wafer defect inspection system and a process of performing such inspection | |
| US10733723B2 (en) | Methods and system for improved quality inspection | |
| US10706525B2 (en) | Methods and systems for improved quality inspection | |
| CN215416732U (en) | Visual inspection system of ceramic substrate | |
| CN108288288B (en) | Method, device and system for measuring precision shaft dimension based on visual identification | |
| CN105784716B (en) | Friction plate Quality Inspection System | |
| JP3522280B2 (en) | Method and apparatus for a ball bond inspection system | |
| CN109142371A (en) | High density flexible exterior substrate defect detecting system and method based on deep learning | |
| WO2023168972A1 (en) | Linear array camera-based copper surface defect detection method and apparatus | |
| CN108344743A (en) | One kind being based on machine vision drug blister package defect inspection method and system | |
| CN112816418B (en) | Mobile phone metal middle frame defect imaging system and detection method | |
| CN111707614B (en) | Optical chip surface contamination defect detection method and line laser vision detection system | |
| CN113670923B (en) | Omnibearing high-precision detection device and method for defects of notebook computer shell | |
| CN109738454A (en) | A kind of soft-package battery tab detection device and method | |
| CN114252452A (en) | Online detection device and method for appearance defects and contour dimension of small-sized revolving body | |
| CN113762427B (en) | Feeding abnormity detection method in industrial automation detection scene | |
| CN108015770A (en) | Position of manipulator scaling method and system | |
| CN112819824A (en) | Neural network of visual inspection system, and inspection system and method including the same | |
| WO2024148645A1 (en) | Apparatus for estimating under monocular infrared thermal imaging vision pose of object grasped by manipulator, and method thereof | |
| CN112014407A (en) | Method for detecting surface defects of integrated circuit wafer | |
| CN116071315A (en) | Product visual defect detection method and system based on machine vision | |
| CN113953208B (en) | Full-automatic sorting device and method for electronic components | |
| CN118351074A (en) | Depth information processing method based on visual reconstruction | |
| CN212275624U (en) | Floor type detection equipment | |
| CN114187269A (en) | Method for rapidly detecting surface defect edge of small-sized device |
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 |