CN114549391A - Circuit board surface defect detection method based on polarization prior - Google Patents

Circuit board surface defect detection method based on polarization prior Download PDF

Info

Publication number
CN114549391A
CN114549391A CN202011353398.9A CN202011353398A CN114549391A CN 114549391 A CN114549391 A CN 114549391A CN 202011353398 A CN202011353398 A CN 202011353398A CN 114549391 A CN114549391 A CN 114549391A
Authority
CN
China
Prior art keywords
image
visible light
module
light polarization
defect
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202011353398.9A
Other languages
Chinese (zh)
Inventor
赵永强
姚乃夫
李梦珂
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Northwestern Polytechnical University
Shenzhen Institute of Northwestern Polytechnical University
Original Assignee
Northwestern Polytechnical University
Shenzhen Institute of Northwestern Polytechnical University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Northwestern Polytechnical University, Shenzhen Institute of Northwestern Polytechnical University filed Critical Northwestern Polytechnical University
Priority to CN202011353398.9A priority Critical patent/CN114549391A/en
Publication of CN114549391A publication Critical patent/CN114549391A/en
Pending legal-status Critical Current

Links

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
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
    • G01N21/84—Systems specially adapted for particular applications
    • G01N21/88—Investigating the presence of flaws or contamination
    • G01N21/95—Investigating the presence of flaws or contamination characterised by the material or shape of the object to be examined
    • G01N21/956—Inspecting patterns on the surface of objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
    • G01N21/84—Systems specially adapted for particular applications
    • G01N21/88—Investigating the presence of flaws or contamination
    • G01N21/95—Investigating the presence of flaws or contamination characterised by the material or shape of the object to be examined
    • G01N21/956—Inspecting patterns on the surface of objects
    • G01N2021/95638—Inspecting patterns on the surface of objects for PCB's
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10048—Infrared image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20212—Image combination
    • G06T2207/20221—Image fusion; Image merging
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30108—Industrial image inspection
    • G06T2207/30141—Printed circuit board [PCB]

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Quality & Reliability (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Chemical & Material Sciences (AREA)
  • Analytical Chemistry (AREA)
  • Biochemistry (AREA)
  • Immunology (AREA)
  • Pathology (AREA)
  • Investigating Materials By The Use Of Optical Means Adapted For Particular Applications (AREA)

Abstract

The invention discloses a method for detecting surface defects of a circuit board based on polarization prior, which comprises the following steps: s1, acquiring an infrared intensity image and a visible light polarization image of the circuit board to be tested; s2, demosaicing the visible light polarization image to obtain a Stokes vector image, and calculating according to the Stokes vector image to obtain: a visible light polarization degree image, a visible light polarization angle image; and S3, inputting the infrared intensity image and the visible light polarization image obtained in the S1, and the visible light polarization degree image and the visible light polarization angle image obtained in the S2 into a pre-trained deep neural network to obtain a defect detection result. The method solves the problems that the efficiency is low when the traditional image processing method is used for detection in the prior art, and defects are difficult to detect by using a single-input deep learning-based method.

Description

Circuit board surface defect detection method based on polarization prior
Technical Field
The invention belongs to the technical field of defect detection, and particularly relates to a circuit board surface defect detection method based on polarization priori.
Background
The circuit board is used as a basic component of electronic equipment, and the importance of the circuit board is self-evident, and the integration level of components is higher and higher as the design of the circuit board is more and more complex. In the actual production process, due to various reasons such as fluctuation of production environment, non-standard manufacturing process and the like, defects are inevitably generated on the surfaces of electronic products and various parts in the production flow. These defects affect the product performance and cause significant economic losses to the user. The existing manual visual detection method has the defects of strong subjectivity, limited human eye spatial resolution, large uncertainty and the like, and can not meet the high-speed and high-accuracy detection requirements of the modern industry. The traditional image processing method is complex in processing, needs a standard circuit board image as a reference and is poor in real-time performance, and a deep learning-based method usually adopts a single information source as input, so that the efficiency is not high when certain defects are detected or the situation that the defects cannot be identified exists.
Disclosure of Invention
The invention aims to provide a method for detecting surface defects of a circuit board based on polarization prior, which aims to solve the problems that the efficiency is low when the traditional image processing method is used for detection in the prior art, and the defects are difficult to detect by using a single-input deep learning-based method.
The invention adopts the following technical scheme: a method for detecting surface defects of a circuit board based on polarization prior comprises the following steps:
s1, acquiring an infrared intensity image and a visible light polarization image of the circuit board to be tested;
s2, demosaicing the visible light polarization image to obtain a Stokes vector image, and calculating according to the Stokes vector image to obtain: a visible light polarization degree image, a visible light polarization angle image;
and S3, inputting the infrared intensity image and the visible light polarization image obtained in the S1, and the visible light polarization degree image and the visible light polarization angle image obtained in the S2 into a pre-trained deep neural network to obtain a defect detection result.
Further, the training process of the deep neural network comprises the following steps: interpolating the infrared intensity image to obtain the resolution ratio same as that of the visible light image, and then performing visual angle conversion to accurately align the infrared intensity image with the visible light image; marking the visible light image according to the defect type, wherein the infrared intensity image shares the marking result of the visible light image; training the depth neural network by using the infrared intensity image, the visible light image and the labeling result to obtain an optimal weight coefficient;
wherein the visible light image comprises a visible light polarization image, a visible light polarization degree image, and a visible light polarization angle image.
Further, the deep neural network includes: the system comprises a feature extraction module, a multi-scale feature self-adaptive fusion module and a defect classification positioning module;
the system comprises a characteristic extraction module, a feature extraction module and a feature extraction module, wherein the characteristic extraction module is used for extracting low-level structural characteristics and high-level semantic characteristics of infrared and visible light images; the multi-scale feature self-adaptive fusion module is used for self-adaptively learning and fusing the proportion according to the contribution of different levels to the detection of the defects with different sizes; and the defect classification positioning module is used for classifying the defect types and regressing the defect positions according to the fused features.
Furthermore, the feature extraction module comprises not less than four convolution modules, each convolution module comprises a plurality of convolution kernels, a normalization layer and an activation layer, the descending scale of the first convolution layer of each convolution module is 1/2, and the size of the subsequent feature graph in the same convolution module is kept unchanged.
Further, the multi-scale feature adaptive fusion module comprises a plurality of convolution kernels and an activation layer, the input of the multi-scale feature adaptive fusion module is the output of each convolution module of the feature extraction module, and the output of the multi-scale feature adaptive fusion module is the result of multiplying the weight learned by the module and the output of each hierarchy of each convolution module of the feature extraction module.
Furthermore, the defect classification positioning module comprises two branches which are a classification branch and a regression branch respectively, the classification branch is composed of at least four convolution kernels, the probability of whether each corresponding position of each point is a defect is finally output, the regression branch is composed of at least four convolution kernels, and the regression branch is finally output as the regression quantity of four coordinates of the corresponding candidate frame of the defect position.
The invention has the beneficial effects that:
(1) the method can fully utilize the polarization characteristics of multiple information sources caused by factors such as material, light source direction and the like on the surface of the circuit board to effectively improve the accuracy and efficiency of defect detection.
(2) By adopting the multi-scale feature self-adaptive fusion module, low-level structural features and high-level semantic features can be effectively fused, the accuracy of defect detection of different sizes is improved, and the interference of non-related levels on detection is reduced.
(3) The network adopts a full convolution module, so that the weight parameters are greatly reduced, and the overfitting phenomenon can be effectively prevented.
Drawings
FIG. 1 is a schematic flow chart of a method for detecting defects of a circuit board according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of a defect detection network according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a deep neural network training process according to an embodiment of the present invention;
FIGS. 4(a) and 4(b) are respectively a visible light polarization image and an infrared intensity image of a circuit board to be tested according to an embodiment of the present invention;
FIGS. 5(a) - (d) are respectively a visible light polarization image, a visible light polarization degree image, a visible light polarization angle image and an infrared intensity image after the conversion of the viewing angle and the resolution in the embodiment of the present invention;
FIG. 6 shows a defect detection result according to an embodiment of the present invention.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
The specific embodiment of the present invention provides a method for detecting defects on a surface of a circuit board based on a polarization image of a polarization image, fig. 1 is a schematic flow chart of the method for detecting defects on a circuit board according to the embodiment of the present invention, and as shown in fig. 1, the method includes:
and S1, acquiring an infrared intensity image and a visible light polarization image of the circuit board to be tested.
In the embodiment of the invention, a rotatable polaroid is arranged in front of the infrared camera and the focal plane visible light polarization camera or the common visible light camera to acquire the infrared intensity image and the visible light polarization image.
Optionally, when the circuit board image is collected, parameters such as the angle of the infrared camera and the visible light polarization camera, the focal length of the lens and the like can be adjusted to enable the circuit board to be within the field of view of the two cameras and to be imaged clearly.
S2, demosaicing the visible light polarization image to obtain a Stokes vector image, and calculating according to the Stokes vector image to obtain: a visible light polarization degree image, a visible light polarization angle image.
In the embodiment of the invention, methods such as linear interpolation, bilinear interpolation, bicubic interpolation and the like can be adopted to demosaic the polarization image to obtain four polarization angle images, and then the four polarization angle images are used for calculating to obtain a polarization degree image and a polarization angle image, wherein the formula is as follows:
I0，I45，I90，I135＝Demosaic(Ipol)，
Figure BDA0002801961520000041
S1＝I0-I90，
S2＝I45-I135，
Figure BDA0002801961520000042
Figure BDA0002801961520000043
wherein Demosaic () is demosaicing operation, I0，I45，I90，I135For four polarization angle images, S0As intensity images, S1And S2For the difference image corresponding to the vertical direction, DoP is the polarization degree image and AoP is the polarization angle image.
And S3, inputting the infrared intensity image and the visible light polarization image obtained in the S1, and the visible light polarization degree image and the visible light polarization angle image obtained in the S2 into a pre-trained deep neural network to obtain a defect detection result.
Hereinafter, the visible light image includes a visible light polarization image, a visible light polarization degree image, and a visible light polarization angle image.
Examples
In an embodiment of the present invention, the visible light image includes a visible light polarization image obtained by demosaicing and a difference image corresponding to a vertical direction, and the polarization degree and polarization angle image includes a visible light polarization degree image and a visible light polarization angle image calculated from the visible light polarization image. When the deep neural network is input, all the images are superposed according to channels to form a three-dimensional tensor which is expressed as
Figure BDA0002801961520000051
C, H, W are the number of channels, height and width of tensor, respectively, and the probability of a certain defect in the range of a certain candidate frame and the position and width and height size information of the candidate frame are output after the pre-trained network operation.
Based on the above embodiments, further, the structure of the defect detection deep neural network is explained below. Fig. 2 is a schematic structural diagram of a defect detection network according to an embodiment of the present invention.
As shown in fig. 2, the deep neural network is composed of three parts, namely, a feature extraction module, a multi-scale feature adaptive fusion module and a defect classification and localization module.
The feature extraction module is used for extracting low-level structural features and high-level semantic features of the infrared and visible light images, and in the embodiment, the feature extraction module is composed of 5 convolution modules, namely conv2_ x, conv3_ x, conv4_ x and conv5_ x, each convolution module comprises a plurality of convolution kernels, a normalization layer and an activation layer, the last convolution layer of each residual block is output as the output of each level, namely C2, C3, C4 and C5, and each level is 1/4, 1/8, 1/16 and 1/32 of the input image scale.
And the multi-scale feature self-adaptive fusion module is used for self-adaptively learning the fusion proportion according to the contributions of different levels to the detection of the defects with different sizes. The characteristic pyramid structure from top to bottom is realized by up-sampling the characteristic graph of the upper layer in a proportion of 2, and the sampling method generally comprises nearest neighbor, bilinear, double-cubic and the like. In the corresponding connection part, the characteristic diagram from bottom to top is added with the characteristic diagram from top to bottom pixel by pixel through reducing the channel number by 1 multiplied by 1 convolution layer. To reduce the noise of the upsampling, the fused feature map is subjected to a 3 × 3 convolution operation to generate a final feature map. The final signature is represented as P2, P3, P4, P5, corresponding to the four bottom-up levels C2, C3, C4, C5. Let the outputs of the P2, P3, P4 and P5 characteristic diagrams be x2, x3, x4 and x5 respectively,
inputting the corresponding convolution network learning self-adaptive weights of the input signals respectively to be alphai、βi、γi、θiWherein i ∈ [2, 3, 4, 5 ]]Four weights per level satisfy αi+βi+γi+θi1 and αi，βi，γi，θi∈[0，1]The fused output is represented as yi＝αix2→i+βix3→i+γix4→i+θix5→i，i∈[2，3，4，5]Wherein x is2→i、x3→i、x4→i、x5→iRespectively, from 2, 3, 4, 5 levels down-sampled or up-sampled to the i level. The weights and number of fusions can be adjusted accordingly according to different pyramid levels.
And the defect classification positioning module is used for classifying the defect types and regressing the defect positions according to the fused features. The classification branch predicts the probability of an object at each pixel position, the frame regression branch is used for regressing the offset between each anchor frame and the nearest real labeling frame, and the two branches are composed of full convolution layers with the same structure. The output of all pyramids is classified and regression information is extracted through not less than four 3 × 3 convolutional layers, the final classification branch output is W × H × KA, the regression branch output is W × H × 4A, where K is the number of categories of all targets, a is the number of all anchor frames, the number of channels of the feature map may be 64, 128, 256, 512, and in this embodiment, 256 is selected as the number of channels.
After the deep neural network is constructed, a visible light polarization image and an infrared intensity image of the circuit board to be tested containing four open circuit defects are respectively obtained by using a visible light polarization camera and an infrared camera, and the obtained image resolutions are 2448 × 2048 and 640 × 512 respectively, as shown in fig. 4(a) and 4 (b). Because the focal planes, the lens parameters and the placing positions of different cameras are different, the visual fields and the resolutions are inconsistent, four images I with different polarization angles are obtained after demosaicing the visible light polarization images0，I45，I90，I135Then, the visible light polarization degree and polarization angle image is calculated. Selecting four corresponding feature points from the infrared intensity image and the visible light image to calculate homography matrixes of two image planes, converting the infrared intensity image to the same resolution and the same visual angle with the corresponding visible light image through the homography matrixes, and removing edge areas to obtain four images with the same resolution and strict alignment, as shown in fig. 5. Compared with the detection of a single visible light polarization image only through color information, the defects can be detected through characteristic information such as extra materials, geometrical shapes, heat radiation and the like on the surface of the circuit board by utilizing the remaining three images. Therefore, the four images are superposed according to the channels to form a four-channel image, the four-channel image is input into the deep neural network, calculation is sequentially carried out through the convolution layer of the deep neural network, the defect type with the confidence coefficient larger than 0.1 and the offset between the defect type and the corresponding anchor frame are output, the offset is converted into four parameters of horizontal and vertical coordinates, width and height in the image through regression calculation, and the four parameters are marked in the visible light polarization image through a rectangular frame. The method can detect all four open circuit defects, and the confidence coefficients are 0.99 and fourThe individual open defect locations are shown as points 1-4 marked in fig. 6.
Based on the above embodiments, further, the deep neural network model training process is explained below. Fig. 3 is a deep neural network training flow chart.
The method comprises the steps of firstly, carrying out interpolation on an infrared intensity image to obtain the resolution ratio which is the same as that of a visible light image, and then carrying out visual angle conversion to enable the infrared intensity image to be accurately aligned with the visible light image.
In this embodiment, the visible light image includes a visible light polarization image and the calculated visible light polarization degree and visible light polarization angle images. The resolution of the infrared intensity image is lower than that of the visible light image, the lens parameters, the relative angle of the camera and the view field are inconsistent, and the infrared intensity image needs to be interpolated and then subjected to view angle conversion so that the infrared intensity image and the visible light image are accurately aligned. As an example, a linear interpolation method can be adopted to up-sample the infrared intensity image to the same resolution as the visible light image, then at least four characteristic points of the infrared intensity image and the visible light image are selected to calculate a homography matrix of the circuit board plane, and finally the infrared intensity image is converted into the same visual angle as the visible light image through homography conversion, so that the infrared intensity image and the visible light image are accurately aligned.
And secondly, labeling the visible light image according to the defect type, wherein the infrared intensity image shares the labeling result of the visible light image.
In this embodiment, a set of input images includes an infrared intensity image and a corresponding visible light image. Since the multiple images are aligned accurately in the previous step, only one visible light image needs to be labeled, and the rest images in the group share the labeling result. As an example, a rectangular frame may be used to label a region where a defect is located in the image, and the region may be labeled according to different defect types. The defect types are common defects on the surface of the circuit board, and include but are not limited to short circuit, open circuit, mouse bite, burr defect and the like.
And step three, training the deep neural network by using the infrared intensity image, the visible light image and the labeling result to obtain an optimal weight coefficient.
In this embodiment, a supervised learning method may be adopted to train the deep neural network, the training image group and the corresponding labeling result are input into the deep neural network, and the predicted category and region are output, the classification loss of the network adopts a cross entropy loss function, and the regression loss of the predicted region frame adopts an L1 loss function, which is expressed as
Figure BDA0002801961520000081
Wherein y isiThe value is labeled for the category and,
Figure BDA0002801961520000082
is a class prediction value, riThe value is marked on the frame and the frame,
Figure BDA0002801961520000083
and updating the weight for the frame prediction value through a gradient descent optimization algorithm through multiple iterations, so that the difference between the predicted defect region and type of the network and the actual marking value is minimized, and an optimal weight coefficient can be obtained to finish the training of the network.
In the prior art, a single source image of a visible light or infrared intensity image is generally used as input to detect the defects of the circuit board, only the color information of the surface of the circuit board is utilized and is easily influenced by a light source, and the method additionally utilizes a polarization degree image and a polarization angle image as input, so that the influence of various interferences can be reduced, and the detection stability is improved. The visible light polarization degree image is related to the refractive index of the material on the surface of the circuit board and the light incidence angle, and can be helpful for detecting the defects of the surface caused by the change of the material; the visible light polarization angle image is related to the azimuth angle of the reflected light on the surface of the circuit board, and can be helpful for detecting the defects of the surface caused by uneven shape. The infrared intensity image can reflect the radiation quantity, and whether defects exist in the infrared intensity image can be judged according to the uniformity of the radiation quantity of the same device. The infrared intensity image can be used as the supplement of the visible light polarization degree and polarization angle image, and the defect missing detection rate and the misjudgment rate are reduced.
The method adopts a multi-scale feature self-adaptive fusion module which can self-adaptively fuse feature levels of different sizes according to different weights according to the contribution degree of feature levels of different sizes to the detection of defects of different sizes, is suitable for the detection of the defects of different sizes and improves the detection accuracy.

Claims (6)

1. A method for detecting surface defects of a circuit board based on polarization prior is characterized by comprising the following steps:
s1, acquiring an infrared intensity image and a visible light polarization image of the circuit board to be tested;
s2, demosaicing the visible light polarization image to obtain a Stokes vector image, and calculating according to the Stokes vector image to obtain: a visible light polarization degree image, a visible light polarization angle image;
and S3, inputting the infrared intensity image and the visible light polarization image obtained in the S1, and the visible light polarization degree image and the visible light polarization angle image obtained in the S2 into a pre-trained deep neural network to obtain a defect detection result.
2. The method of claim 1, wherein the training process of the deep neural network comprises: interpolating the infrared intensity image to obtain the resolution ratio same as that of the visible light image, and then performing visual angle conversion to accurately align the infrared intensity image with the visible light image; marking the visible light image according to the defect type, wherein the infrared intensity image shares the marking result of the visible light image; training the deep neural network by using the infrared intensity image, the visible light image and the labeling result to obtain an optimal weight coefficient;
wherein the visible light image comprises a visible light polarization image, a visible light polarization degree image, and a visible light polarization angle image.
3. The method of claim 1, wherein the deep neural network comprises: the system comprises a feature extraction module, a multi-scale feature self-adaptive fusion module and a defect classification positioning module;
the system comprises a feature extraction module, a feature extraction module and a feature extraction module, wherein the feature extraction module is used for extracting low-level structural features and high-level semantic features of infrared and visible light images; the multi-scale feature self-adaptive fusion module is used for self-adaptively learning and fusing the proportion according to the contributions of different levels to the detection of the defects with different sizes; and the defect classification positioning module is used for classifying the defect types and regressing the defect positions according to the fused features.
4. The method of claim 3, wherein the feature extraction modules include no less than four convolution modules, the convolution modules are composed of a plurality of convolution kernels, normalization layers, and activation layers, the first convolution layer descent scale of each convolution module is 1/2, and the size of the subsequent feature map in the same convolution module remains unchanged.
5. The method of claim 3, wherein the multi-scale feature adaptive fusion module comprises a plurality of convolution kernels and an activation layer, wherein the input of the multi-scale feature adaptive fusion module is the output of each convolution module of the feature extraction module, and the output of the multi-scale feature adaptive fusion module is the result of multiplying the weight learned by the module by the output of each convolution module of the feature extraction module at each level.
6. The method of claim 3, wherein the defect classification and localization module comprises two branches, namely a classification branch and a regression branch, the classification branch is composed of not less than four convolution kernels, the final output is the probability of whether each corresponding position of each point is a defect, the regression branch is composed of not less than four convolution kernels, and the final output is the regression quantity of four coordinates of the corresponding candidate frame of the defect position.
CN202011353398.9A 2020-11-26 2020-11-26 Circuit board surface defect detection method based on polarization prior Pending CN114549391A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011353398.9A CN114549391A (en) 2020-11-26 2020-11-26 Circuit board surface defect detection method based on polarization prior

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011353398.9A CN114549391A (en) 2020-11-26 2020-11-26 Circuit board surface defect detection method based on polarization prior

Publications (1)

Publication Number Publication Date
CN114549391A true CN114549391A (en) 2022-05-27

Family

ID=81668033

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011353398.9A Pending CN114549391A (en) 2020-11-26 2020-11-26 Circuit board surface defect detection method based on polarization prior

Country Status (1)

Country Link
CN (1) CN114549391A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116433532A (en) * 2023-05-06 2023-07-14 合肥工业大学 Infrared polarized image fusion denoising method based on attention-guided filtering
CN116433661A (en) * 2023-06-12 2023-07-14 锋睿领创(珠海)科技有限公司 Method, device, equipment and medium for detecting semiconductor wafer by multitasking

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116433532A (en) * 2023-05-06 2023-07-14 合肥工业大学 Infrared polarized image fusion denoising method based on attention-guided filtering
CN116433532B (en) * 2023-05-06 2023-09-26 合肥工业大学 Infrared polarized image fusion denoising method based on attention-guided filtering
CN116433661A (en) * 2023-06-12 2023-07-14 锋睿领创(珠海)科技有限公司 Method, device, equipment and medium for detecting semiconductor wafer by multitasking
CN116433661B (en) * 2023-06-12 2023-08-18 锋睿领创(珠海)科技有限公司 Method, device, equipment and medium for detecting semiconductor wafer by multitasking

Similar Documents

Publication Publication Date Title
CN109584248B (en) Infrared target instance segmentation method based on feature fusion and dense connection network
CN111223088B (en) Casting surface defect identification method based on deep convolutional neural network
CN106875373B (en) Mobile phone screen MURA defect detection method based on convolutional neural network pruning algorithm
CN112288008B (en) Mosaic multispectral image disguised target detection method based on deep learning
CN112967243A (en) Deep learning chip packaging crack defect detection method based on YOLO
CN111524135A (en) Image enhancement-based method and system for detecting defects of small hardware fittings of power transmission line
CN112836713A (en) Image anchor-frame-free detection-based mesoscale convection system identification and tracking method
CN113591795B (en) Lightweight face detection method and system based on mixed attention characteristic pyramid structure
CN113920107A (en) Insulator damage detection method based on improved yolov5 algorithm
CN111553949A (en) Positioning and grabbing method for irregular workpiece based on single-frame RGB-D image deep learning
CN113516656B (en) Defect image data processing simulation method based on ACGAN and Cameralink cameras
CN114549391A (en) Circuit board surface defect detection method based on polarization prior
CN113191204B (en) Multi-scale blocking pedestrian detection method and system
CN111209858A (en) Real-time license plate detection method based on deep convolutional neural network
CN111582074A (en) Monitoring video leaf occlusion detection method based on scene depth information perception
CN111967288A (en) Intelligent three-dimensional object identification and positioning system and method
CN112633220A (en) Human body posture estimation method based on bidirectional serialization modeling
CN114332942A (en) Night infrared pedestrian detection method and system based on improved YOLOv3
CN111368637B (en) Transfer robot target identification method based on multi-mask convolutional neural network
CN115775236A (en) Surface tiny defect visual detection method and system based on multi-scale feature fusion
CN115661611A (en) Infrared small target detection method based on improved Yolov5 network
CN117474863A (en) Chip surface defect detection method for compressed multi-head self-attention neural network
CN114627044A (en) Solar photovoltaic module hot spot detection method based on deep learning
CN117079125A (en) Kiwi fruit pollination flower identification method based on improved YOLOv5
CN114926456A (en) Rail foreign matter detection method based on semi-automatic labeling and improved deep learning

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