CN102854194A - Object surface defect detection method and apparatus based on linear array CCD - Google Patents
Object surface defect detection method and apparatus based on linear array CCD Download PDFInfo
- Publication number
- CN102854194A CN102854194A CN2012103346430A CN201210334643A CN102854194A CN 102854194 A CN102854194 A CN 102854194A CN 2012103346430 A CN2012103346430 A CN 2012103346430A CN 201210334643 A CN201210334643 A CN 201210334643A CN 102854194 A CN102854194 A CN 102854194A
- Authority
- CN
- China
- Prior art keywords
- output
- array ccd
- neural network
- neuron
- defective
- 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
Images
Landscapes
- Image Processing (AREA)
- Investigating Materials By The Use Of Optical Means Adapted For Particular Applications (AREA)
Abstract
The invention discloses an object surface defect detection method and an object surface defect detection apparatus based on a linear array CCD. According to the apparatus, the image information of surfaces of the objects on a production line is collected by using a fixed linear array CCD; defect identification and defect location determination are carried out by using a method of non-interlaced defect detection with a BP neural network algorithm according to the collected image information. The apparatus comprises an optical image collection module, a signal conditioning module, an A/D conversion module, a driving module, a power supply module, and a programmable logic device and peripheral circuits and a software system thereof. The programmable logic device is the core of the apparatus. The method and the apparatus provided by the invention have the advantages of high detection speed, high detection accuracy, simple structure, and low cost. The method and the apparatus are suitable to be used in single defect detections upon surface images of objects on the production line.
Description
Technical field
The present invention relates to a kind of object surface defect inspection method and device based on line array CCD.
Background technology
In the material and material manufactured goods of all conglomeraties, surface imperfection is one of key factor that affects product quality, not only affect the outward appearance of product such as defectives such as the scabbing of steel plate, aluminium sheet, Copper Foil, pits, more seriously reduced corrosion resistivity, wearing quality and the fatigue strength etc. of product; In addition, in high quality polymer film application, the defective of membraneous material can affect the performance of product equally, and for example for large screen flat plate display, the defective of membraneous material can affect the electrical insulation properties of protection screen; For the high-quality wrappage, the defective of membraneous material can reduce thermotolerance, the corrosion resistivity of packing; In the printing industry journey, owing to reasons such as techniques, printed matter tends to occur the open defects such as aberration, defect point, the line of ink marker, thus the appearance that causes printing substandard products or even waste product.Therefore, if can carry out 100% automatic detection and be used for quality control and manufactured goods screening detecting the information that obtains starting material and finished product in process of production, will be the important means of improving the quality of products.
Traditional detection method has the detection method of surface flaw of non-automaticization and the detection method of surface flaw of robotization; Wherein, the detection method of surface flaw of non-automaticization has artificial visual sampling observation method and surperficial strobe light to detect two kinds; The detection method of surface flaw of robotization has the Computer Vision Detection method of laser scanning method and CCD imaging etc.
Artificial visual: detect this and rely on the visual inspection defective, but the naked eyes detectability is limited after all, in the situation that object is from A car sped at the moment, naked eyes can't in time focus on, thereby produce " motion blur sense ".The visual not only detection efficiency of old friend worker is low, and cost is high, labour intensity is large.
Strobe light detects: it detects principle is that the pulse flash with 10 ~ 30us can cause the retina arrest reaction, plays the effect of camera shutter.This detection method is further developed, be about to stroboscopic light source and be combined with special-purpose video camera, observe defective by monitor.This method cost is low, but automaticity is also low.
The laser scanning inspection method: laser scanning inspection is owned by France in the category of vision-based detection.When the laser beam after focusing on dropped on object to be detected surperficial, according to character and the structure of material surface, radiant rays was scattered more or less.If there is defective in body surface, inevitable so that scattering strength changes with respect to normal surface, detect it and change the defective that just can detect body surface.Data volume is very large in this defect inspection process, utilize general purpose microprocessor to be difficult to accomplish real-time processing, so detection speed is slow; If improve detection speed, then need to adopt special machine to carry out data and process, cause the high cost of defects detection.
CCD imaging detecting method: CCD (Charge Coupled Devices) is charge-coupled image sensor, is a kind of novel solid imaging device.It is the Analogous Integrated Electronic Circuits chip that develops on mountain, extensive silicon integrated circuit technique basis, and the conversion of light harvesting electricity, light integration, three kinds of functions of scanning are integrated.Its essential part is by the photosensitive element array of MOS and read shift register and form.The CCD device has that volume is little, lightweight, vibration and shock resistant, be subjected to ambient electromagnetic field affect little, operating distance is large, measuring accuracy is high, low cost and other advantages, be widely used in the measurement and control of various industry spot.Traditional CCD imaging detecting method is mainly Array CCD splicing method.
Array CCD splicing method: detection system adopts a plurality of area array CCD cameras, and the visual field overlaps mutually between the adjacent camera, realizes the collection to plate surface information.The resolution that gathers image depends on the number of CCD camera sensing device, the camera collection of same type to the resolution of image fix, can change by the size of adjusting the camera collection image accuracy of detection of system.The advantage of this detection method is: the scope of application is wider, and accommodative ability of environment is strong, can realize simultaneously the judgement of effects on surface defective many kinds of parameters.Shortcoming is: the uniformity requirement to light source is very high, the heterogeneity of illumination is proofreaied and correct by the enlargement factor of adjusting vision signal, thereby obtain unified grayscale image, this finally causes the reduction of signal to noise ratio (S/N ratio), because the signal noise in dark territory also is exaggerated simultaneously, from figure image intensifying light and shade contrast and light source heterogeneity concerning image effect correction angle, the one dimension image that the relative line sweep of two dimensional image obtains seems more complicated, so the unstable of light source is the main cause that causes the detection system erroneous judgement; 2, owing to imaging viewing field crossover between the Array CCD Camera is not undetected to guarantee, so the image redundancy data volume is larger; Image is processed, classification of defects is finished by software usually, and the software amount of calculation is large.
Therefore, existing surface defects detection system has the shortcomings such as the data volume of detection is large, detection speed is slow, testing cost is high, pick-up unit is complicated.
Summary of the invention
Technical matters to be solved by this invention provides a kind of object surface defect inspection method and device based on line array CCD, should be fast based on object surface defect inspection method and the device detection speed of line array CCD, Detection accuracy is high, and simple in structure, with low cost, be suitable for the single defective of object surface image on the production line is detected online.
The technical solution of invention is as follows:
A kind of body surface defect inspection method based on line array CCD, the irradiation that light source sends is on body surface, the image imaging of body surface is to the pixel array surface of line array CCD device, so that the output of line array CCD device is used for characterizing the analog voltage signal S (t) of light intensity; In the line array CCD device each is as the corresponding analog voltage signal S (t) of number;
Adopt the BP neural network to analog voltage signal S (t) processing of classifying, thereby identify the body surface defective;
The input layer of described BP neural network is a neuron, and input signal is analog voltage signal S (t);
The hidden layer of described BP neural network is one deck, comprises 3 neurons;
The output layer of described BP neural network is a neuron, and output 0 or 1 signal represents that respectively body surface zone corresponding to current pixel point is defective or normal; Because position and the body surface zone of pixel in line array CCD have one-to-one relationship, therefore, be the defective locations place that object can be determined in 0 the position of pixel in the line array CCD device according to output signal.
Before using the BP neural network, first the BP neural network is carried out data pre-service and training;
The pretreated process of described data is: use first the value of the analog voltage signal S (t) in the sample divided by the value of the maximum analog voltage signal S (t) of these row, so that its input value is limited on the interval [0,1]; And, represent " defective " with 0, represent " normally " with 1, as the expection output of BP neural network;
Described training process is:
(3) weights of this network of initialization and deviation: the initialization weights of network and deviation are taken the decimal between [1,1] that random function generates;
(4) train and the iteration of scanning process:
The actual value of output neuron〉0.9 o'clock just think that the Boolean quantity of output of this moment is 1, export and thought that the Boolean quantity of output at this moment was 0 at＜0.1 o'clock;
Learning rate η is made as 0.9, just refer to the hereinafter neuronic deviation of output layer according to the output valve of the end condition neural network of appointment and the error of real output value less than 0.1[error herein, refer in the present invention deviation θ 5], adopt a plurality of samples, adopt the method for example renewal to train and iteration to weights and deviation;
1. with respect to front one deck i, the clean input Ij that calculates neuron j is I
j=∑ (W
IjO
i)+θ
j
2. calculate neuron output: the unipolarity Sigmoid function that uses logarithmic
The output of each neuron j is mapped to interval [0,1];
3. calculate reverse propagated error ERRj: to each neuron j of output layer, ERRj=Oj* (1-Oj) * (Tj-Oj), wherein, Tj is the actual output of the known class label of training sample; [Tj is the actual quantization output of training sample, can only get 0 or 1, and Oj does not disperse, be continuous, its variation range is [0.1], for native system, because output only has two classes: 0 or 1, therefore condition can suitably be relaxed, when herein Oj output〉just think that 1, Oj output thought 0 at＜0.1 o'clock 0.9 the time.】
4. calculate the error E RRj of hidden layer neuron j: each the neuron j from last to first hidden layer, ERRj=Oj* (1-Oj) * ∑ (ERRk*Wjk), K are the node serial number of output layer;
5. refreshing weight: to each weights Wij in the network, calculate Δ Wij=η * ERRj*Oj with following two formulas;
Wij=Wij+ΔWij；
6. upgrade deviation: to each the deviation θ in the network
j, calculate with following two formulas
Δθ
j=η*ERRj；
θ
j=θ
j+Δθ
j；
After training is finished, from the correlation rule of BP neural network extraction about the input and output class.
A kind of body surface defect detecting device based on line array CCD adopts aforesaid body surface defect inspection method based on line array CCD, comprises optical imagery acquisition module, programmable logic device (PLD), driver module, signal condition module and A/D modular converter;
The logical driver module of programmable logic device (PLD) drives the work of optical imagery acquisition module; The output signal of optical imagery acquisition module is input to programmable logic device (PLD) through signal condition module and A/D modular converter successively and processes;
The optical imagery acquisition module comprises light source, optical lens and imageing sensor.
Described body surface defect detecting device based on line array CCD also comprises display screen and the warning audio amplifier that is connected with programmable logic device (PLD).
Beneficial effect:
The invention discloses a kind of object surface defect inspection method and device based on line array CCD, this device gathers the image information of object surface on the production line by fixing line array CCD, then according to the image information that collects, utilize the BP neural network algorithm, adopt line by line detection whether to exist the method for defective to realize the identification of defective and determining of defective locations; Device comprises optical imagery acquisition module, signal condition module, A/D modular converter, driver module, power module, programmable logic device (PLD) and peripheral circuit thereof and its software systems.The above device is as core take programmable logic device (PLD).Detection speed of the present invention is high, and Detection accuracy is high, and simple in structure, with low cost, is suitable for the detection to the defective that the body surface image is single on the production line.
Every frame that line array CCD is collected (being interpreted as of frame: the corresponding output data of each pixel of line array CCD, the data that the whole pixels of linear array CCD scanning are once exported are a frame, system whenever collects frame data, just these frame data are preserved, and separately these frame data are processed) view data processes separately, and judge separately whether this frame image data the inside exists defective data, and unlike traditional scheme, first view data is all gathered and finish, carry out again graphical analysis, processing.
The BP neural network algorithm is to process iteratively one group of training sample, and neural network forecast and the actual known class label of each sample are relatively learnt.Oppositely revise its weights for each sample, so that the square-error between neural network forecast and the actual class is minimum.The BP neural network algorithm iterates according to optimum training criterion, determines and the continuous neural network structure of adjusting, and by iterative modifications, learning process stops when weights are restrained.Therefore, it has the advantages such as error is little, convergence good, dynamic is good, the result is objective.
Adopt the linear array CCD scanning method, its advantage is:
1, owing to self high scan rate and high resolving power, the surface is detected can reach higher resolution and precision at plate width direction;
2, contrast area array CCD scan method, correcting distorted error and optical heterogeneity are more easy;
3, irredundant data or redundant data amount are very little [because the imaging viewing field crossover is not undetected to guarantee between the Array CCD Camera, so the image redundancy data volume is larger, then there is not this problem in line array CCD, therefore irredundant data or redundant data amount are very little], linear array CCD scanning is easy to realize in the row, algorithm in the ranks, and image processing algorithm relevant and line-to-line correlation is very effective to defect detection in the row.
Gather the image information of body surface on the production line by fixing line array CCD, then according to the image information that collects, adopt to detect line by line whether to exist the method for defective to come the information such as defect recognition and definite defective locations.
Whether exist the principle of defective be: utilize programmable logic device (PLD) that every frame image data that line array CCD collects is processed separately if detecting line by line, and judge separately whether have defective data in this frame image data the inside [defining of defective data and normal data is not unalterable, because the surface of different materials, defective data and normal data are all different, therefore the need of defective data and normal data define according to concrete material surface.When system just brings into operation, have a self study stage.In this stage, can offer systematic learning to normal sample and defect sample respectively, learning phase finishes, and feature (size of the level value of line array CCD output) normal and that defect sample has can be learnt by system.Therefore in this system, be not that to have a kind of value of fixed data be the value of corresponding defective data, but according to different surfacings, defective data has different features.The adaptive ability that is system is strong, has certain artificial intelligence.], and unlike traditional scheme, view data is all gathered finish first, carry out again graphical analysis, processing.To there being the row of defective, need further to process: if this row is the beginning of a new defective, then according to running situation, record information and the audio alerts such as position of this defective.Wherein, the defect recognition principle is:
1, learning phase.Before data are classified, there is a self study stage.This stage adopts the BP neural network algorithm to extract the material surface feature.
2, the Data classification stage.To the data that every row collects, use the BP neural network algorithm to carry out the classification of data, be divided into two classes: normal and defective.If the data of one's own profession all belong to " normally " class, this capable defective that do not exist then; If there are the data that belong to " defective " class in one's own profession, then there is defective in this row.
In real time display defect position and warning, detection speed is fast, and Detection accuracy is high, and simple in structure, with low cost.
Description of drawings
Fig. 1 is system architecture diagram;
Fig. 2 is one dimension electronic camera system principle of work schematic diagram;
Fig. 3 is the process flow diagram of defect identification method;
Fig. 4 is an object surface schematic diagram that has defective.
Fig. 5 is the topology diagram of the BP neural network among the present invention.
Embodiment
Below with reference to the drawings and specific embodiments the present invention is described in further details:
Embodiment 1:
Referring to Fig. 1, present invention includes optical imagery acquisition module, signal condition module, A/D modular converter, driver module, power module, programmable logic device (PLD) and peripheral circuit thereof.Described programmable logic device (PLD) is the core of this system, links to each other with A/D modular converter, driver module, power module etc. respectively.
The optical imagery acquisition module.This module is comprised of light source, optical lens and imageing sensor, and this module is responsible for gathering the body surface image.What imageing sensor adopted is line array CCD.This module is powered by power module, and accepts the driving of driver module, and the body surface image information that it collects is passed to the signal condition module.
The signal condition module.This module comprises amplifies and filtering two parts, is prior art.Because the scope of the outputting analog signal scope of CCD device and the analog input signal of A/D conversion chip is not mated, therefore before the A/D conversion, must amplify (concept that amplification herein is a broad sense), make it be suitable for the input of analog to digital converter (ADC); Filtering circuit is used for the interference noise of filtered signal simultaneously.This module is powered by power module, and its input is the output of optical imagery acquisition module, and its output is the input of A/D modular converter.
Driver module.Because the signal level of the drive signal level during CCD work and programmable logic device (PLD) output is not mated, so need to adopt voltage changer to form driving circuit programmable logic device (PLD) output timing signal is converted to the required level of CCD, be prior art.
The A/D modular converter.The A/D modular converter becomes analog signal conversion corresponding digital signal output under sampling clock control, and the digital data transmission after will changing by data line is in programmable logic device (PLD), so that programmable logic device is processed data.This module is powered by power module, and its input is the output of signal condition module and the control command of programmable logic device (PLD), and its output is digitized simulating signal, outputs to programmable logic device (PLD).
Programmable logic device (PLD) and peripheral circuit thereof.Programmable logic device (PLD) is the control core of whole system, cooperates with the interface of data transmission module for CCD provides suitable driving pulse sequential, A/D conversion and control, data to process to reach, and has realized needed all Digital Logic in the system.
Power module is used for powering to the each several part of system, because the each several part of system possibility supply voltage not identical (if any 5V, 3.3V, 1.2V etc.), therefore this module output multi-channel DC voltage.
Principle of work of the present invention is as follows:
Referring to Fig. 2, the irradiation that light source sends is being with on the defective body surface, and this moment, CCD device and peripheral circuit and optical system formed the one dimension electronic camera system.Under irradiation light, the defective in the material is imaged onto CCD device pixel array surface, and through CCD output is the analog voltage signal S (t) that is loaded with defect information (S (t) is directly proportional with incident intensity) of continuous distribution.Then the signal condition module is just nursed one's health S (t) signal, sends into the A/D modular converter after conditioning finishes and carries out digitizing, sends into programmable logic device (PLD) after the digitizing and processes.What native system adopted is to detect line by line the method that whether has defective to come the information such as defect recognition and definite defective locations, namely utilize programmable logic device (PLD) that every frame image data that line array CCD collects is processed separately, and judge separately whether this frame image data the inside exists defective data, and unlike traditional scheme, first view data is all gathered and finish, carry out again graphical analysis, processing.To there being the row of defective, need further to process: if this row is the beginning of a new defective, then according to running situation, record information and the audio alerts such as position of this defective.
A certain example detection sample is as shown in Figure 4:
An instantiation is as follows:
1, learning phase.Before data are classified, there is a self study stage.This stage adopts the BP neural network algorithm to extract the material surface feature.Step is as follows:
1), data pre-service.Before beginning training, need to standardize and to the data category attribute recompile of discrete type to the sample data that collects.Use first in the sample corresponding magnitude of voltage divided by the maximum attribute value of these row, so that its input value is limited on the interval [0,1], for the data category attribute of discrete type, recompile represents " defective " with 0, represents " normally " with 1, exports as expecting.Data after the standardization are as shown in table 2.[output voltage of CCD is the output voltage of each pixel, only have a pixel large although the size of defective is general unlikely, but our scheme is to adopt the method that judges whether line by line to exist defective, judge separately namely whether the every delegation that collects exists defective, and concrete reason has above been set forth.What we adopted is the line array CCD that 1024 pixels are arranged, the every run-down of line array CCD is just exported the magnitude of voltage that represents 1024 pixel light intensity serially so, these 1024 magnitudes of voltage are exactly a frame, and then separately these 1024 magnitudes of voltage are processed, judge in these 1024 voltages whether have the voltage that belongs to " defective class ".The direction of motion of material to be detected is vertical with the direction of scanning of line array CCD, therefore in the situation that line array CCD does not move, material movement to be detected just can realize the scanning of material all surfaces to be detected.】
Table 1 sample data (only providing representational several data in this table, because other data are all similar with these)
| Sample number | CCD output voltage/V | | Remarks | |
| 1 | 1.282 | Defective | The dark colour defective | |
| 2 | 1.294 | Defective | The dark colour defective |
| 3 | 1.368 | Normally | |
| 4 | 1.383 | Normally | |
| 5 | 1.392 | Normally | |
| 6 | 2.003 | Defective | The light colour defective |
| 7 | 2.007 | Defective | The light colour defective |
| ...... |
Sample data after table 2 standardization
| Sample number | CCD output voltage/ | Data type | |
| 1 | 0.427 | 0 | |
| 2 | 0.431 | 0 | |
| 3 | 0.456 | 1 | |
| 4 | 0.461 | 1 | |
| 5 | 0.464 | 1 | |
| 6 | 0.668 | 0 | |
| 7 | 0.669 | 0 | |
| ...... |
2), planned network topological structure.The key of Network Topology Design is to determine neuron number and each neuron initial value and the deviation of hidden layer; Network through training if its accuracy can not be accepted, then must re-start topology design or use different initial weights and deviation instead.
According to analysis, establish 1 input layer, 1 comprises 3 neuronic hidden layers and 1 output layer (neuron value 0 expression data " defective " class is got 1 expression data " normally " class.Network topology structure as shown in Figure 5.
3), the weights of this network of initialization and deviation.The initialization weights of network and deviation are taken the decimal between [1,1] that random function generates, and initialization weights and the deviation of network are as shown in table 3.[θ 2-5 represent respectively hidden layer neuron 2,3,4 and the deviation of output layer neuron 5]
Table 3
| W12 | W13 | W14 | W25 | W35 | W45 | θ2 | θ3 | θ4 | θ5 |
| -0.1 | 0.3 | 0.1 | -0.1 | 0.4 | 0.2 | -0.3 | 0.4 | 0.2 | 0.2 |
4), train and the iteration of scanning process.According to the end condition of learning rate and appointment (output valve of neural network and the error of real output value are less than a certain threshold value. for native system, because output only has two classes: 0 or 1, therefore condition can suitably be relaxed, output〉0.9 o'clock just think 1, export and thought 0 at＜0.1 o'clock.), adopt the method for example renewal to train and the scanning process iteration to weights and deviation.
Learning rate η is made as 0.9, according to the end condition (output valve of neural network and the error of real output value are less than 0.1) of appointment the method for weights and the renewal of deviation employing example is trained and the scanning process iteration.At first, get No. 1 sample in the table 2, its input 0.427 is offered network, calculate each neuronic clean input, output and error, and backpropagation errors, again refreshing weight and deviation;
1. calculate the clean input of neuron j: with respect to front one deck i, the clean input Ij that calculates neuron j is I
j=∑ (W
IjO
i)+θ
j
2. calculate neuron output: the unipolarity Sigmoid function that uses logarithmic
The output of each neuron j is mapped to interval [0,1];
3. calculate reverse propagated error: to each neuron j of output layer, ERRj=Oj* (1-Oj) * (Tj-Oj), wherein, Tj is the actual output of the known class label of training sample;
4. calculate the error E RRj of hidden layer neuron j: each the neuron j from last to first hidden layer, ERRj=Oj* (1-Oj) * ∑ (ERRk*Wjk), K are the node serial number of output layer;
5. refreshing weight: to each weights Wij in the network, calculate with following two formulas,
ΔWij=η*ERRj*Oj，
Wij=Wij+ΔWij；
6. upgrade deviation: to each the deviation θ in the network
i, calculate with following two formulas
Δθ
j=η*ERRj；
θ
j=θ
j+Δθ
j；
In the scan iterations, each neuronic clean input, output, error, weights, deviation are upgraded as shown in table 4 for the first time.
Then, get sample No. 2, its input 0.431 is offered network, according to weights and the deviation behind No. 1 Sample Refreshment, calculate each neuronic clean input, output and error, and weights and the deviation again upgraded.All samples are repeated said process, finish scan iterations the 1st time.If scan iterations does not satisfy end condition for the first time, then begin the next round scan iterations, until satisfy end condition, iteration finishes.
Table 4
2) derive correlation rule.After above-mentioned training, the BP neural network can be extracted the correlation rule about the input and output class, i.e. corresponding which kind of output of which kind of data input.
2, the Data classification stage.Use the BP neural network algorithm to carry out Data classification, be adopt the BP algorithm with each property value of training sample as input, actual class as output.To the BP neural network after the training, by the clustering processing of beta pruning, neuron or active value, derive the correlation rule of input layer and output layer, can realize the classification of concrete data according to these rules.The native system data are divided into two classes: normal and defective.If the data of one's own profession all belong to " normally " class, this capable defective that do not exist then; If there are the data that belong to " defective " class in one's own profession, then there is defective in this row.
The defect recognition software flow pattern as shown in Figure 3.
Experiment shows, is that the single defects detection Detection accuracy of body surface image can reach more than 95% on the production line of 200m/min to operating rate.
Claims (4)
1. body surface defect inspection method based on line array CCD, it is characterized in that, the irradiation that light source sends is on body surface, the image imaging of body surface is to the pixel array surface of line array CCD device, so that the output of line array CCD device is used for characterizing the analog voltage signal S (t) of light intensity; In the line array CCD device each is as the corresponding analog voltage signal S (t) of number;
Adopt the BP neural network to analog voltage signal S (t) processing of classifying, thereby identify the body surface defective;
The input layer of described BP neural network is a neuron, and input signal is analog voltage signal S (t);
The hidden layer of described BP neural network is one deck, comprises 3 neurons;
The output layer of described BP neural network is a neuron, and output 0 or 1 signal represents that respectively body surface zone corresponding to current pixel point is defective or normal; Because position and the body surface zone of pixel in line array CCD have one-to-one relationship, therefore, be the defective locations place that object can be determined in 0 the position of pixel in the line array CCD device according to output signal.
2. the body surface defect inspection method based on line array CCD according to claim 1 is characterized in that, before using the BP neural network, first the BP neural network is carried out data pre-service and training;
The pretreated process of described data is: use first the value of the analog voltage signal S (t) in the sample divided by the value of the maximum analog voltage signal S (t) of these row, so that its input value is limited on the interval [0,1]; And, represent " defective " with 0, represent " normally " with 1, as the expection output of BP neural network;
Described training process is:
(1) weights of this network of initialization and deviation: the initialization weights of network and deviation are taken the decimal between [1,1] that random function generates;
(2) train and the iteration of scanning process:
The actual value of output neuron〉0.9 o'clock just think that the Boolean quantity of output of this moment is 1, export and thought that the Boolean quantity of output at this moment was 0 at＜0.1 o'clock;
Learning rate η is made as 0.9, little according to the error of the output valve of the end condition neural network of appointment and real output value
In 0.1, adopt a plurality of samples, adopt the method for example renewal to train and iteration to weights and deviation;
1. with respect to front one deck i, the clean input Ij that calculates neuron j is I
j=∑ (W
IjO
i)+θ
j
2. calculate neuron output: the unipolarity Sigmoid function that uses logarithmic
The output of each neuron j is mapped to interval [0,1];
3. calculate reverse propagated error ERRj: to each neuron j of output layer, ERRj=Oj* (1-Oj) * (Tj-Oj), wherein, Tj is the actual output of the known class label of training sample;
4. calculate the error E RRj of hidden layer neuron j: each the neuron j from last to first hidden layer, ERRj=Oj* (1-Oj) * ∑ (ERRk*Wjk), K are the node serial number of output layer;
5. refreshing weight: to each weights Wij in the network, calculate Δ Wij=η * ERRj*Oj with following two formulas;
Wij=Wij+ΔWij；
6. upgrade deviation: to each the deviation θ in the network
j, calculate with following two formulas
Δθ
j=η*ERRj；
θ
j=θ
j+Δθ
j；
After training is finished, from the correlation rule of BP neural network extraction about the input and output class.
3. body surface defect detecting device based on line array CCD, it is characterized in that, adopt claim 1 or 2 described body surface defect inspection methods based on line array CCD, comprise optical imagery acquisition module, programmable logic device (PLD), driver module, signal condition module and A/D modular converter;
The logical driver module of programmable logic device (PLD) drives the work of optical imagery acquisition module; The output signal of optical imagery acquisition module is input to programmable logic device (PLD) through signal condition module and A/D modular converter successively and processes;
The optical imagery acquisition module comprises light source, optical lens and imageing sensor.
4. the body surface defect detecting device based on line array CCD according to claim 1 is characterized in that, also comprises the display screen and the warning audio amplifier that are connected with programmable logic device (PLD).
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201210334643.0A CN102854194B (en) | 2012-09-11 | 2012-09-11 | Object surface defect detection method and apparatus based on linear array CCD |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201210334643.0A CN102854194B (en) | 2012-09-11 | 2012-09-11 | Object surface defect detection method and apparatus based on linear array CCD |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN102854194A true CN102854194A (en) | 2013-01-02 |
| CN102854194B CN102854194B (en) | 2014-11-05 |
Family
ID=47400970
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201210334643.0A Expired - Fee Related CN102854194B (en) | 2012-09-11 | 2012-09-11 | Object surface defect detection method and apparatus based on linear array CCD |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102854194B (en) |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103841726A (en) * | 2014-03-06 | 2014-06-04 | 河南科技大学 | Luminance real-time adjusting device of light source in pole piece surface defect detection and adjusting method thereof |
| CN104713584A (en) * | 2013-12-17 | 2015-06-17 | 致茂电子股份有限公司 | Correcting method for optical detecting device |
| CN104748856A (en) * | 2013-12-31 | 2015-07-01 | 致茂电子股份有限公司 | Optical detection device |
| CN105571500A (en) * | 2016-01-26 | 2016-05-11 | 中塑联新材料科技湖北有限公司 | Random tester for plastic sheet production line |
| CN107123107A (en) * | 2017-03-24 | 2017-09-01 | 广东工业大学 | Cloth defect inspection method based on neutral net deep learning |
| CN107300559A (en) * | 2017-08-25 | 2017-10-27 | 山东众鑫电子材料有限公司 | A kind of Kapton defect detection system and method |
| CN107705293A (en) * | 2017-09-14 | 2018-02-16 | 广东工业大学 | A kind of hardware dimension measurement method based on CCD area array cameras vision-based detections |
| CN108509947A (en) * | 2018-01-29 | 2018-09-07 | 佛山市南海区广工大数控装备协同创新研究院 | A kind of automatic identification polishing process based on artificial neural network |
| CN108672316A (en) * | 2018-03-27 | 2018-10-19 | 哈尔滨理工大学 | A kind of micro parts quality detecting system based on convolutional neural networks |
| CN108846841A (en) * | 2018-07-02 | 2018-11-20 | 北京百度网讯科技有限公司 | Display screen quality determining method, device, electronic equipment and storage medium |
| CN110254468A (en) * | 2019-06-20 | 2019-09-20 | 吉林大学 | A kind of raceway surface defect intelligent online detection device and detection method |
| CN110286138A (en) * | 2018-12-27 | 2019-09-27 | 合刃科技（深圳）有限公司 | Information detecting method, apparatus and system |
| CN110473179A (en) * | 2019-07-30 | 2019-11-19 | 上海深视信息科技有限公司 | A kind of film surface defects detection method, system and equipment based on deep learning |
| CN111272763A (en) * | 2018-12-04 | 2020-06-12 | 通用电气公司 | System and method for workpiece inspection |
| CN111398293A (en) * | 2020-04-08 | 2020-07-10 | 重庆引尖机电有限公司 | Spare part production detecting system |
| CN111693531A (en) * | 2019-03-14 | 2020-09-22 | 顶级手套(国际)有限公司 | Artificial intelligence for defect detection of polymer products in production line |
| RU2795303C1 (en) * | 2022-08-02 | 2023-05-02 | Общество с ограниченной ответственностью "Объединенная Компания РУСАЛ Инженерно-технологический центр" | Method for automatic continuous surface quality control |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH04122847A (en) * | 1990-09-14 | 1992-04-23 | Omron Corp | Optical apparatus for inspecting defect |
| JP2004013300A (en) * | 2002-06-04 | 2004-01-15 | Denso Wave Inc | Information code reader |
| CN1696675A (en) * | 2005-05-31 | 2005-11-16 | 哈尔滨工业大学 | Image type online flaws detection equipment in line array for surface of strip steel, and detection method |
| CN102183525A (en) * | 2011-01-20 | 2011-09-14 | 单宝华 | Bridge apparent state automatic detection device and method based on charge coupled device (CCD) array photographic technique |
| CN202351182U (en) * | 2011-11-16 | 2012-07-25 | 福建标新易开盖集团有限公司 | Online high-speed detection system for defects on surface of tinplate |
-
2012
- 2012-09-11 CN CN201210334643.0A patent/CN102854194B/en not_active Expired - Fee Related
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH04122847A (en) * | 1990-09-14 | 1992-04-23 | Omron Corp | Optical apparatus for inspecting defect |
| JP2004013300A (en) * | 2002-06-04 | 2004-01-15 | Denso Wave Inc | Information code reader |
| CN1696675A (en) * | 2005-05-31 | 2005-11-16 | 哈尔滨工业大学 | Image type online flaws detection equipment in line array for surface of strip steel, and detection method |
| CN102183525A (en) * | 2011-01-20 | 2011-09-14 | 单宝华 | Bridge apparent state automatic detection device and method based on charge coupled device (CCD) array photographic technique |
| CN202351182U (en) * | 2011-11-16 | 2012-07-25 | 福建标新易开盖集团有限公司 | Online high-speed detection system for defects on surface of tinplate |
Non-Patent Citations (4)
| Title |
|---|
| 李晶等: "基于BP神经网络的外姿态测量系统线阵CCD标定", 《仪器仪表学报》, vol. 31, no. 5, 31 May 2010 (2010-05-31), pages 1138 - 1141 * |
| 王庆等: "基于BP 神经网络的冷轧带钢表面缺陷分类研究", 《福建电脑》, no. 2, 31 December 2007 (2007-12-31), pages 97 - 108 * |
| 石争浩，冯亚宁等: "基于机器视觉的铸件表面缺陷人工神经网络检测方法", 《重型机械》, no. 2, 31 December 2004 (2004-12-31), pages 52 - 5 * |
| 胡亮等: "基于线阵CCD 钢板表面缺陷在线检测系统的研究", 《计量学报》, vol. 26, no. 3, 31 July 2005 (2005-07-31), pages 201 - 3 * |
Cited By (22)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104713584A (en) * | 2013-12-17 | 2015-06-17 | 致茂电子股份有限公司 | Correcting method for optical detecting device |
| CN104713584B (en) * | 2013-12-17 | 2017-04-12 | 致茂电子股份有限公司 | Correcting method for optical detecting device |
| CN104748856A (en) * | 2013-12-31 | 2015-07-01 | 致茂电子股份有限公司 | Optical detection device |
| CN103841726A (en) * | 2014-03-06 | 2014-06-04 | 河南科技大学 | Luminance real-time adjusting device of light source in pole piece surface defect detection and adjusting method thereof |
| CN103841726B (en) * | 2014-03-06 | 2015-09-09 | 河南科技大学 | The real-time adjusting device of light-source brightness and method of adjustment thereof in pole piece defects detection |
| CN105571500A (en) * | 2016-01-26 | 2016-05-11 | 中塑联新材料科技湖北有限公司 | Random tester for plastic sheet production line |
| CN107123107A (en) * | 2017-03-24 | 2017-09-01 | 广东工业大学 | Cloth defect inspection method based on neutral net deep learning |
| CN107300559A (en) * | 2017-08-25 | 2017-10-27 | 山东众鑫电子材料有限公司 | A kind of Kapton defect detection system and method |
| CN107705293A (en) * | 2017-09-14 | 2018-02-16 | 广东工业大学 | A kind of hardware dimension measurement method based on CCD area array cameras vision-based detections |
| CN108509947A (en) * | 2018-01-29 | 2018-09-07 | 佛山市南海区广工大数控装备协同创新研究院 | A kind of automatic identification polishing process based on artificial neural network |
| CN108672316A (en) * | 2018-03-27 | 2018-10-19 | 哈尔滨理工大学 | A kind of micro parts quality detecting system based on convolutional neural networks |
| CN108846841A (en) * | 2018-07-02 | 2018-11-20 | 北京百度网讯科技有限公司 | Display screen quality determining method, device, electronic equipment and storage medium |
| US11488294B2 (en) | 2018-07-02 | 2022-11-01 | Beijing Baidu Netcom Science Technology Co., Ltd. | Method for detecting display screen quality, apparatus, electronic device and storage medium |
| CN111272763A (en) * | 2018-12-04 | 2020-06-12 | 通用电气公司 | System and method for workpiece inspection |
| CN111272763B (en) * | 2018-12-04 | 2023-04-07 | 通用电气公司 | System and method for workpiece inspection |
| CN110286138A (en) * | 2018-12-27 | 2019-09-27 | 合刃科技（深圳）有限公司 | Information detecting method, apparatus and system |
| CN111693531A (en) * | 2019-03-14 | 2020-09-22 | 顶级手套(国际)有限公司 | Artificial intelligence for defect detection of polymer products in production line |
| CN110254468A (en) * | 2019-06-20 | 2019-09-20 | 吉林大学 | A kind of raceway surface defect intelligent online detection device and detection method |
| CN110473179A (en) * | 2019-07-30 | 2019-11-19 | 上海深视信息科技有限公司 | A kind of film surface defects detection method, system and equipment based on deep learning |
| CN110473179B (en) * | 2019-07-30 | 2022-03-25 | 上海深视信息科技有限公司 | Method, system and equipment for detecting surface defects of thin film based on deep learning |
| CN111398293A (en) * | 2020-04-08 | 2020-07-10 | 重庆引尖机电有限公司 | Spare part production detecting system |
| RU2795303C1 (en) * | 2022-08-02 | 2023-05-02 | Общество с ограниченной ответственностью "Объединенная Компания РУСАЛ Инженерно-технологический центр" | Method for automatic continuous surface quality control |
Also Published As
| Publication number | Publication date |
|---|---|
| CN102854194B (en) | 2014-11-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN102854194B (en) | Object surface defect detection method and apparatus based on linear array CCD | |
| Deitsch et al. | Automatic classification of defective photovoltaic module cells in electroluminescence images | |
| Ren et al. | State of the art in defect detection based on machine vision | |
| Zhou et al. | An automatic surface defect inspection system for automobiles using machine vision methods | |
| CN109550712B (en) | Chemical fiber filament tail fiber appearance defect detection system and method | |
| CN109239102B (en) | CNN-based flexible circuit board appearance defect detection method | |
| Bai et al. | An optimized railway fastener detection method based on modified Faster R-CNN | |
| CN104112269B (en) | A kind of solar battery laser groove parameter detection method and system based on machine vision | |
| CN106875373B (en) | Mobile phone screen MURA defect detection method based on convolutional neural network pruning algorithm | |
| Horstmeyer et al. | Convolutional neural networks that teach microscopes how to image | |
| He et al. | Deep regression neural network for industrial surface defect detection | |
| CN107944504B (en) | Board recognition and machine learning method and device for board recognition and electronic equipment | |
| CN102854191A (en) | Real-time visual detection and identification method for high speed rail surface defect | |
| EP3669227A1 (en) | Identifying the quality of the cell images acquired with digital holographic microscopy using convolutional neural networks | |
| CN107966454A (en) | A kind of end plug defect detecting device and detection method based on FPGA | |
| TWI834005B (en) | Deep learning networks for nuisance filtering | |
| CN110490842B (en) | Strip steel surface defect detection method based on deep learning | |
| CN111965636A (en) | Night target detection method based on millimeter wave radar and vision fusion | |
| Jiang et al. | A deep neural networks approach for pixel-level runway pavement crack segmentation using drone-captured images | |
| JP2016181098A (en) | Area detection device and area detection method | |
| CN114119526A (en) | Steel plate surface defect detection and identification system and method based on machine vision | |
| CN111487192A (en) | Machine vision surface defect detection device and method based on artificial intelligence | |
| Yauney et al. | Convolutional neural network for combined classification of fluorescent biomarkers and expert annotations using white light images | |
| CN111145145A (en) | Image surface defect detection method based on MobileNet | |
| Medina et al. | Automated visual inspection of road surface cracks |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20141105 Termination date: 20160911 |