A high-speed and high-precision method for measuring appearance defects and circuits of ceramic copper-clad substrates

Abstract

The present invention relates to a high-speed and high-precision ceramic copper-clad substrate appearance defect and circuit measurement method, the steps of which include: image acquisition, cutout, partition processing, defect detection, secondary screening, circuit measurement and result determination. The method uses defect detection to ensure the comprehensiveness of appearance detection, and secondary screening to ensure the accuracy of appearance detection, thereby ensuring that the missed detection rate and false detection rate are reduced when the number of sub-boards is plural and the types of appearance defects are numerous, thereby reducing the scrap cost. The same sub-board image is used to simultaneously complete appearance detection and circuit detection, the detection types are comprehensive, and the steps are saved.

Description

Technical Field

[0001] The present invention relates to the field of optical detection technology, and in particular to a high-speed and high-precision ceramic copper-clad substrate appearance defect and circuit measurement method. Background Art

[0002] In recent years, with the rapid development of new energy vehicles, aerospace, energy storage, smart grids and smart home appliances, high-voltage and high-power semiconductor modules have been widely used. With the development of electronic technology, the integration of chips is getting larger and larger, the wiring width of circuits is getting finer and finer, and the power dissipation per unit area is getting larger and larger. In a high-power and high-voltage environment, it is inevitable that the heat will increase, which will easily lead to the loss of the original function of semiconductor devices.

[0003] Semiconductor power modules have high output power and high heat generation. Poor heat dissipation of the chip can easily cause the bonding wire to peel or fuse, leading to module failure. The chip's current carrying and heat dissipation are mainly completed by the ceramic copper-clad substrate in the module. The ceramic copper-clad substrate absorbs the heat generated by the chip and conducts it to the heat sink, realizing heat exchange between the chip and the outside world.

[0004] In the field of semiconductor packaging, copper-clad ceramic substrates are widely used due to their excellent thermal conductivity, insulation and mechanical properties. Among them, direct bonded copper (DBC) and active metal brazing (AMB) are the two main ceramic copper-clad technologies. In terms of material, ceramic substrates are superior to ordinary glass fiber PCB boards due to their heat dissipation performance, current carrying capacity, insulation, thermal expansion coefficient, etc.; in terms of copper cladding, ceramic substrates are made by combining copper foil and ceramic substrates in a high temperature environment through high / low temperature co-firing, copper plating, copper cladding, etc., with strong bonding, copper foil that is not easy to fall off, high reliability, and stable performance in high temperature and high humidity environments.

[0005] In semiconductor process technology, copper-clad ceramic substrates are mainly used as carriers for various power electronic device chips. In the packaging of IGBT modules (insulated gate bipolar transistors), copper-clad ceramic substrates with IGBT chips are usually welded to the module bottom plate to form a heat dissipation channel for the IGBT chip. The copper-clad ceramic substrates absorb the heat generated by the chip and conduct it to the heat sink, thereby achieving heat exchange between the chip and the outside world to achieve heat dissipation.

[0006] Chinese patent CN114354491A discloses a DCB ceramic substrate defect detection method based on machine vision. The method first obtains a standard image of the substrate, establishes an image template, then collects an image of the product to be tested, performs image preprocessing, matches the image template with the test image, locates the module area to be tested, and then performs defect detection and marks the defect type, thereby realizing the classification of defect types.

[0007] Chinese patent CN 108346148A discloses a method for detecting oxidation areas of high-density flexible IC substrates, which uses a high-precision camera to take pictures, moves the stage to obtain images of various parts of the flexible IC substrate, corrects image distortion, and then fuses the images to obtain a complete image of the flexible IC substrate. This can increase the clarity of the image, thereby accurately determining the oxidation area.

[0008] In the actual manufacturing process of ceramic copper-clad substrates, when etching the circuits on the copper surface, defects such as open circuits and short circuits may occur, and the substrate will lose its original controllable conductive ability. In addition, when copper and ceramic are attached, there may be defects such as sintering bubbles, lumps, copper defects, dirt, oxidation and scratches, which will damage the bonding strength between the ceramic and the copper surface, and further cause thermal failure and other problems in subsequent products. Ceramic copper-clad substrates are generally manufactured as large motherboards first, and then cut into multiple daughter boards. The quality of independent daughter boards is unrelated, so the defects of the daughter boards need to be detected separately to determine whether they are defective products. These defects are numerous and complex, and there is also a probability of misjudgment, which leads to a high probability that the daughter boards will be judged as defective products due to misdetection, which will bring unnecessary scrap costs.

[0009] Therefore, it is necessary to improve the detection method to solve the above problems. Summary of the invention

[0010] A main purpose of the present invention is to provide a high-speed and high-precision ceramic copper-clad substrate appearance defect and circuit measurement method, which can fully detect the ceramic copper-clad motherboard, distinguish the range of sub-boards, and then detect the defect types according to the ceramic surface and the copper surface respectively, screen out suspected defects, and then perform secondary judgment. Not only is the defect detection type comprehensive, but also the misjudgment is rare.

[0011] The present invention achieves the above-mentioned purpose through the following technical scheme: a high-speed and high-precision ceramic copper-clad substrate appearance defect and circuit measurement method, the steps comprising:

[0012] S1. Image acquisition: Use an industrial camera to collect appearance photos of the ceramic copper-clad motherboard to obtain a complete image of the ceramic copper-clad motherboard;

[0013] S2, cutting out: segmenting and cutting out the ceramic copper-clad motherboard image according to the number of sub-boards in the ceramic copper-clad motherboard and the arrangement of the sub-boards, to obtain a sub-board image of each sub-board;

[0014] S3, partition processing: for each sub-board image, the copper area and ceramic area in the sub-board are extracted respectively by using the different colors of ceramic and copper under the customized light source, and the image is enhanced;

[0015] S4, defect detection: determine whether the copper area has a first suspected defect, and whether the ceramic area has a second suspected defect. If the first suspected defect or the second suspected defect exists, extract the corresponding defect thumbnail, and then execute step S5. The first suspected defect includes sintering bubbles, lumps, copper defects, oxidation, dirt and scratches, and the second suspected defects include open circuits, short circuits, bosses, line defects and ceramic pollution. Otherwise, execute step S6;

[0016] S5, secondary screening: the defect thumbnail is transmitted to the classification convolutional neural network for classification, and the first suspected defect and the second suspected defect detected are secondary judged as real defects or over-inspected defects. If they are real defects, the defects are recorded; if they are over-inspected defects, the defects are not recorded;

[0017] S6, circuit measurement: perform circuit measurement for each sub-board image, and measure the position and size of key objects in some copper areas and ceramic areas according to the drawing specifications. The key objects are local strip areas of ceramics, copper surfaces or green oil that need to be judged for size accuracy. If the size of the key objects exceeds the accuracy range, it is determined that the circuit is defective;

[0018] S7. Result determination: If there are real defects and / or bad circuits in the sub-board image, the sub-board is determined to be a bad board, otherwise it is determined to be a qualified board.

[0019] Specifically, the complete ceramic copper-clad motherboard image in step S1 is composed of multiple local images; the left-right size of the ceramic copper-clad motherboard is M, and the top-bottom size is N. The range of the local image is the field of view of the industrial camera, the left-right size is m, the top-bottom size is n, the number of puzzles is z=x×y, x is M / m rounded up, y is N / n rounded up, the acquired local images are sorted in the order of left first, right second, top third, and bottom, and then named Image1-Imagez in sequence, and the grayscale value of the pixel point of each local image is calculated.

[0020] Furthermore, before the image acquisition step is performed, the industrial camera is calibrated first, and the calibration method is as follows: prepare a calibration plate that meets the camera field of view and accuracy requirements, and the calibration plate has a known geometric shape and size; place the calibration plate under the camera in different postures and angles to take a group of pictures; extract feature points from the acquired calibration plate image, calculate the internal and external parameters of the camera and the distortion coefficient of the lens according to the posture transformation of the feature points, and finally apply the calculated distortion coefficient to the local image to convert the local image into a rectangular plane image without distortion.

[0021] Furthermore, the local image stitching method is as follows: generate an empty image Image0 that meets the size of the entire ceramic copper-clad motherboard, set the coordinates of the upper left corner vertex of Image0 to (0, 0), and map the grayscale value of each pixel in Image1 to Image0 one by one according to the actual position of the local image in the motherboard, and the mapping starting point starts from (0, 0);

[0022] Calculate the overlapping pixels of Image2 and Image1, and map the grayscale value of each pixel remaining after removing the overlapping pixels in Image2 to Image0. The mapping starting point of Image2 starts from the upper right corner vertex of the pixel mapped from Image1.

[0023] According to the order of the partial images, the mapping of Image3-Imagez to Image0 is completed in sequence from left to right and from top to bottom, so as to obtain the complete ceramic copper-clad motherboard image.

[0024] Specifically, the specific method of the cutout step is:

[0025] According to the design drawings of the ceramic copper-clad motherboard, the design width of the ceramic copper-clad motherboard, the design height of the ceramic copper-clad motherboard, the design size of the sub-board, the design height of the horizontal cutting line and the design width of the vertical cutting line are extracted. The sub-boards are arranged in an array. The range of the sub-boards consists of the minimum rectangular copper area and the ceramic boundary around it. The design size of the sub-board includes the outward expansion width and outward expansion height of the ceramic boundary. The horizontal cutting line is between the upper and lower adjacent sub-boards and runs through the width direction of the ceramic copper-clad motherboard from left to right. The vertical cutting line is between the left and right adjacent sub-boards and runs through the height direction of the ceramic copper-clad motherboard from top to bottom. The horizontal cutting line intersects with the vertical cutting line to form a grid-shaped cutting area;

[0026] Taking the designed height of the horizontal cutting line as the height and the designed width of the ceramic copper-clad motherboard as the width, a rectangular structural element W is generated, and the ceramic copper-clad motherboard image is traversed using W, and the ceramic area that can be covered by W is retained to obtain images of several horizontal dividing strips; taking the designed width of the vertical cutting line as the width and the designed height of the ceramic copper-clad motherboard as the height, a rectangular structural element H is generated, and the ceramic copper-clad motherboard image is traversed using H, and the ceramic area that can be covered by H is retained to obtain images of several vertical dividing strips; combining the images of all the horizontal dividing strips and the images of the vertical dividing strips to obtain an image of the complete dividing strip;

[0027] The image of the separation zone is removed from the image of the ceramic copper-clad motherboard to obtain images of several minimum rectangular copper-containing areas, the four corner points of each minimum rectangular copper-containing area are located, and the outer expansion corresponding to the four corner points is determined according to the width and height of the ceramic boundary in the motherboard drawing, thereby determining the positions of the four outer expansion vertices corresponding to the four corner points, and the four outer expansion vertices are connected to obtain the actual daughterboard outline;

[0028] All daughter board images are cut out from the ceramic copper-clad motherboard image according to the actual daughter board outline, and the individual daughter board images are sorted in the order of top first then bottom, and left first then right.

[0029] Specifically, the specific method of image enhancement is:

[0030] Use Fourier transform on the sub-plate image obtained by cutting out, convert the image from the spatial domain to the frequency domain, and obtain the original frequency domain image ImageP;

[0031] Generate a Gaussian frequency domain filter with specific specifications and specific resolution according to the size of the sub-board image, multiply the value of each pixel on ImageP by the value of the corresponding pixel on the frequency domain filter, and obtain a new frequency domain image ImagePF;

[0032] Use inverse Fourier transform on ImagePF to convert the image from frequency domain to spatial domain to obtain a new spatial domain image PFImage;

[0033] Subtract the grayscale value of the corresponding pixel on PFImage from the grayscale value of each pixel on the sub-plate image, and multiply the difference by an enhancement coefficient to obtain the enhanced sub-plate image ImageZ.

[0034] Furthermore, the specific method of the defect detection is:

[0035] For images with obvious background / foreground distinction after enhancement, use binarization to extract suspected pixels of defects, and group suspected pixels with the same gray value range together as suspected areas;

[0036] Use Blob analysis to extract and mark the connected domains of the suspected area. Each marked Blob represents a suspected target.

[0037] The features of the suspected targets are calculated, and the suspected targets are screened according to the features of the sub-board appearance defects. The ones that are finally retained are regarded as suspected defects.

[0038] Furthermore, the specific method of the secondary screening is:

[0039] Based on the convolutional neural network YOLO-V8 model, the real defect images and inspected defect images collected in the early stage are used as training samples, and the classification model is obtained after training;

[0040] Input the defect thumbnail corresponding to the suspected defect area into the classification model to obtain the classification result of the defect image output by the classification convolutional neural network;

[0041] Judging the classification results, if the defect category belongs to a category of real defects, the suspected defect is a real defect, otherwise it is an over-inspected defect and does not need to be detected;

[0042] If no real inspection defect is detected, the process of the current sub-board is qualified, otherwise it is unqualified.

[0043] Specifically, the line measurement method is:

[0044] Import the CAD drawing of the daughter board, extract the contour information, the design position information of the key object and the standard size of the key object on the CAD drawing, wherein the key object is a local strip area of ​​ceramic, copper surface or green oil that needs to be judged for dimensional accuracy;

[0045] Mapping the CAD contour information and the design position information of the key object onto the sub-board image to find the key object;

[0046] Establishing a straight line perpendicular to the extension direction of the key object, and taking the length of the line segment intercepted by the straight line and the actual outline of the key object as the actual size of the key object;

[0047] Calculate the average gray value of the pixels on each line segment, form a contour curve with the arrangement of gray values, and use a filter to smooth the contour;

[0048] Calculate the first-order derivative of the smooth contour. The sub-pixel positions of all local extreme values ​​of the first-order derivative are edge candidate points. The edge candidate points are represented by the vector of the derivative pointing to the corresponding position of the smooth grayscale contour. The edge candidate points whose absolute values ​​are greater than a given threshold are regarded as image edges. The discrete edge points on both sides are fitted into straight line segments respectively. The distance between the two edge line segments is the actual size of the focus object.

[0049] The actual size of the key object is compared with the standard size on the CAD drawing. If the difference between the two is within the allowable range of process error, the process of the current sub-board is determined to be qualified, otherwise it is unqualified.

[0050] The beneficial effects of the technical solution of the present invention are:

[0051] 1. This method uses defect detection to ensure the comprehensiveness of appearance inspection, and secondary screening to ensure the accuracy of appearance inspection, thereby ensuring that the missed detection rate and false detection rate are reduced when there are multiple sub-boards and many types of appearance defects, thereby reducing scrap costs.

[0052] 2. Use the same daughter board image to complete appearance inspection and circuit inspection at the same time, with comprehensive inspection types and saved steps.

[0053] 3. The field of view of the industrial camera is narrowed by taking partial images, and several relatively clear partial images covering the entire surface of the ceramic copper-clad motherboard are obtained. The partial images are distorted and adjusted, and then spliced ​​into a complete ceramic copper-clad motherboard image, which is clear and accurate.

[0054] 4. The image is converted from the spatial domain to the frequency domain for processing through image enhancement methods, making the defect edge clear. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is an arrangement relationship diagram of a local image without distortion in an embodiment;

[0056] Figure 2 This is a complete image of a ceramic copper-clad motherboard in the embodiment;

[0057] Figure 3 It is the sub-board image in the embodiment.

[0058] The numbers in the figure represent:

[0059] 1-Copper area, 2-Ceramic area, 3-Ceramic boundary. DETAILED DESCRIPTION

[0060] The present invention is further described in detail below with reference to specific embodiments.

[0061] Example:

[0062] The present invention relates to a high-speed and high-precision ceramic copper-clad substrate appearance defect and circuit measurement method, the steps comprising:

[0063] S1. Image acquisition: Use an industrial camera to capture appearance photos of the ceramic copper clad motherboard to obtain a complete image of the ceramic copper clad motherboard.

[0064] Ceramic copper clad motherboards are manufactured according to motherboard drawings. However, during the manufacturing process, there may be slight differences in the sub-board structures on the ceramic copper clad motherboard due to some uncontrollable reasons. The purpose of the test is to confirm whether these slight differences affect the qualification of the sub-boards. The prerequisite for accurate detection is to obtain accurate images of the ceramic copper clad motherboard. For ordinary industrial cameras, the farther the shooting distance, the larger the field of view, but the lower the clarity. Therefore, in order to obtain a more accurate image of the ceramic copper clad motherboard, the field of view of the industrial camera is generally narrowed to obtain several relatively clear local images covering the entire surface of the ceramic copper clad motherboard, and then spliced ​​into a ceramic copper clad motherboard image. For example Figure 1 and Figure 2 As shown, in the embodiment, the complete ceramic copper-clad motherboard image is composed of 4×3 partial images, and the size of each partial image is consistent.

[0065] However, the local images captured by industrial cameras are distorted, which makes it impossible to stitch adjacent local images together. The distortion of industrial cameras is mainly divided into barrel distortion and pincushion distortion. The magnification of the center area of ​​the optical axis in the field of view is much greater than that of the edge area, which produces barrel distortion, making the center area of ​​the image clearer, while the surrounding edge areas will expand and deform; the magnification of the edge area in the field of view is much greater than that of the center area of ​​the optical axis, which produces pincushion distortion, making the surrounding edge areas of the image clearer, while the center area will be compressed and deformed. Therefore, it is necessary to obtain the intrinsic and extrinsic parameters of the camera through camera calibration to perform distortion correction, so as to ensure that the image is clear and accurate. Therefore, the industrial camera must be calibrated before the image acquisition step.

[0066] The specific calibration method for industrial cameras is as follows: prepare a calibration plate that meets the camera's field of view and accuracy requirements, and the calibration plate has a known geometric shape and size; place the calibration plate under the camera in different postures and angles to take a set of pictures; extract feature points from the acquired calibration plate images, and calculate the camera's internal and external parameters and the lens's distortion coefficient based on the posture transformation of the feature points. Finally, apply the calculated distortion coefficient to the local image to convert the local image into a rectangular plane image without distortion.

[0067] After calibration, the edge of the local image can more realistically reflect its planar state, and adjacent local images can confirm which part belongs to the overlapping area.

[0068] The left and right dimensions of the ceramic copper-clad motherboard are M, and the upper and lower dimensions are N. The range of the local image is the field of view of the industrial camera, with the left and right dimensions being m, and the upper and lower dimensions being n. The number of puzzle pieces is z=x×y, where x is M / m rounded up, and y is N / n rounded up. The acquired local images are sorted in the order of left to right, top to bottom, and then named Image1-Imagez (the maximum in the embodiment is 12), and the grayscale value of the pixel point of each local image is calculated.

[0069] The specific stitching method of the partial image is as follows: generate an empty image Image0 that meets the size of the entire ceramic copper-clad motherboard, set the coordinates of the upper left corner vertex of Image0 to (0, 0), and map the grayscale value of each pixel in Image1 to Image0 one by one according to the actual position of the partial image in the motherboard, and the mapping starting point starts from (0, 0);

[0070] Calculate the overlapping pixels of Image2 and Image1, and map the grayscale value of each pixel remaining after removing the overlapping pixels in Image2 to Image0. The mapping starting point of Image2 starts from the upper right corner vertex of the pixel mapped from Image1.

[0071] According to the order of the partial images, the mapping of Image3-Imagez to Image0 is completed from left to right and from top to bottom, so as to obtain a complete ceramic copper-clad motherboard image.

[0072] The range of the local image is not restricted by the size of the daughter board, so when shooting with an industrial camera, it is sufficient to ensure that the range of all local images can cover the entire mother board. This is applicable to various daughter board models.

[0073] S2, cutout: according to the number of daughter boards in the ceramic copper-clad motherboard and the arrangement of the daughter boards, the ceramic copper-clad motherboard image is segmented and cutout to obtain the daughter board image of each daughter board. Figure 3 shown.

[0074] The specific method of the cutout step is:

[0075] According to the design drawings of the ceramic copper-clad motherboard, the design width of the ceramic copper-clad motherboard, the design height of the ceramic copper-clad motherboard, the design size of the sub-board, the design height of the horizontal cutting line and the design width of the vertical cutting line are extracted. The sub-boards are arranged in an array. The range of the sub-boards is composed of the minimum rectangular copper-containing area and the ceramic boundary around it. The design size of the sub-board includes the outward expansion width and outward expansion height of the ceramic boundary. The horizontal cutting line is between the upper and lower adjacent sub-boards and runs through the width direction of the ceramic copper-clad motherboard from left to right. The vertical cutting line is between the left and right adjacent sub-boards and runs through the height direction of the ceramic copper-clad motherboard from top to bottom. The horizontal cutting line intersects with the vertical cutting line to form a grid-like cutting area;

[0076] A rectangular structure element W is generated with the design height of the horizontal cutting line as the height and the design width of the ceramic copper-clad motherboard as the width, and W is used to traverse the ceramic copper-clad motherboard image, and the ceramic area 2 that can be covered by W is retained to obtain images of several horizontal dividing strips; a rectangular structure element H is generated with the design width of the vertical cutting line as the width and the design height of the ceramic copper-clad motherboard as the height, and H is used to traverse the ceramic copper-clad motherboard image, and the ceramic area 2 that can be covered by H is retained to obtain images of several vertical dividing strips; the images of all the horizontal dividing strips and the images of the vertical dividing strips are combined to obtain an image of the complete dividing strip;

[0077] The image of the separation zone is removed from the image of the ceramic copper-clad motherboard to obtain images of several minimum rectangular copper-containing areas, the four corner points of each minimum rectangular copper-containing area are located, and the outer expansion corresponding to the four corner points is determined according to the width and height of the ceramic boundary in the motherboard drawing, thereby determining the positions of the four outer expansion vertices corresponding to the four corner points, and the four outer expansion vertices are connected to obtain the actual daughterboard outline;

[0078] All daughter board images are cut out from the ceramic copper-clad motherboard image according to the actual daughter board outline, and the individual daughter board images are sorted in the order of top first, then bottom, and left first, then right.

[0079] The ceramic cutting line is a daughterboard partition area with a width of zero to several millimeters. Because the daughterboard design outline on the CAD drawing generally contains the ceramic boundary, and there is no obvious boundary between the ceramic boundary and the ceramic cutting line on the daughterboard image, it is necessary to find it by expansion. Because the daughterboard is based on copper area 1, and copper area 1 and ceramic area 2 have a clear boundary on the motherboard image, the daughterboard outline needs to rely on copper area 1 to find. In fact, there is a ceramic surface inside copper area 1 instead of a complete rectangular area, but to determine the outline of the daughterboard, you only need to first determine the minimum rectangular range covering copper area 1, that is, the minimum rectangular copper-containing area. Ceramic boundary 3 is expanded from the surrounding of the minimum rectangular copper-containing area. Before determining the minimum rectangular copper-containing area, the ceramic area 2 in the complete ceramic copper-clad motherboard image must be removed according to the size of the horizontal cutting line and the vertical cutting line, so two rectangular structural elements W and H must be constructed. The area swept by the rectangular structure element W will extend from the lower boundary of the copper area 1 of the previous row of sub-boards to the upper boundary of the copper area 1 of the next row of sub-boards. The area swept by the rectangular structure element H will extend from the right boundary of the copper area 1 of the left row of sub-boards to the left boundary of the copper area 1 of the right row of sub-boards. The minimum rectangular copper-containing area is obtained by removing these two swept areas.

[0080] S3, partition processing: for each sub-board image, using the different colors of ceramic and copper under the customized light source, the copper area 1 and the ceramic area 2 in the substrate are extracted respectively, and the image is enhanced.

[0081] Ceramic copper-clad motherboard is made by sintering copper foil directly onto the ceramic surface. The copper surface has a bumpy and granular feel, not a smooth plane. When defects appear on the copper surface, especially when the degree of the defect is mild or the area is small, the particles on the copper surface will continuously scatter the light reflected by the defect, making the imaging of the defect blurred and causing unclear edges. The image can be regarded as a superposition of waves of various frequencies. The frequency feature is the grayscale change feature of the image. The grayscale change of low-frequency features is not obvious, such as the overall outline of the image, and the grayscale change of high-frequency features is drastic, such as image edges and noise. Minor defects are difficult to distinguish from the background due to blurred edges, so it is necessary to convert the image from the spatial domain to the frequency domain for processing to make the defect edges clear.

[0082] The specific methods of image enhancement are:

[0083] Use Fourier transform on the sub-plate image obtained by cutting out, convert the image from the spatial domain to the frequency domain, and obtain the original frequency domain image ImageP;

[0084] Generate a Gaussian frequency domain filter with specific specifications and specific resolution according to the size of the sub-board image, multiply the value of each pixel on ImageP by the value of the corresponding pixel on the frequency domain filter, and obtain a new frequency domain image ImagePF;

[0085] Use inverse Fourier transform on ImagePF to convert the image from frequency domain to spatial domain to obtain a new spatial domain image PFImage;

[0086] Subtract the grayscale value of the corresponding pixel on PFImage from the grayscale value of each pixel on the sub-plate image, and multiply the difference by an enhancement coefficient to obtain the enhanced sub-plate image ImageZ.

[0087] S4, defect detection: determine whether there is a first suspected defect in copper area 1 and whether there is a second suspected defect in ceramic area 2. If there is a first suspected defect or a second suspected defect, extract the corresponding defect thumbnail, and then execute step S5. The first suspected defect includes sintering bubbles, lumps, copper defects, oxidation, dirt and scratches, and the second suspected defects include open circuits, short circuits, bosses, line defects and ceramic pollution; otherwise, execute step S6.

[0088] The specific methods of defect detection are:

[0089] For images with obvious background / foreground distinction after enhancement, use binarization to extract suspected pixels of defects, and group suspected pixels with the same gray value range together as suspected areas;

[0090] Use Blob analysis to extract and mark the connected domains of the suspected area. Each marked Blob represents a suspected target.

[0091] The features of the suspected targets are calculated, and the suspected targets are screened according to the features of the sub-board appearance defects. The ones that are finally retained are regarded as suspected defects.

[0092] S5. Secondary screening: The defect image is sent to the classification convolutional neural network for classification, and a secondary judgment is made as to whether the first suspected defect and the second suspected defect detected are real defects or over-inspected defects. If they are real defects, the defects are recorded; if they are over-inspected defects, the defects are not recorded.

[0093] The specific method of secondary screening is:

[0094] Based on the convolutional neural network YOLO-V8 model, the real defect images and inspected defect images collected in the early stage are used as training samples, and the classification model is obtained after training;

[0095] Input the defect thumbnail corresponding to the suspected defect area into the classification model to obtain the classification result of the defect image output by the classification convolutional neural network;

[0096] Judging the classification results, if the defect category belongs to a category of real defects, the suspected defect is a real defect, otherwise it is an over-inspected defect and does not need to be detected;

[0097] If no real inspection defect is detected, the process of the current sub-board is qualified, otherwise it is unqualified.

[0098] This method uses defect detection to ensure the comprehensiveness of appearance detection, and secondary screening to ensure the accuracy of appearance detection, thereby ensuring that the missed detection rate and false detection rate are reduced when there are multiple sub-boards and many types of appearance defects, thereby reducing scrap costs.

[0099] S6. Circuit measurement: Perform circuit measurement for each sub-board image. According to the drawing specifications, perform position and size measurement on the key objects of the copper area 1 and the ceramic area 2. The key objects are the local strip areas of the ceramic, copper surface or green oil that need to be judged for dimensional accuracy. If the size of the key object exceeds the accuracy range, it is judged as a defective circuit.

[0100] The line measurement method is:

[0101] Import the CAD drawing of the daughter board, extract the contour information on the CAD drawing, the design position information of the key objects, and the standard size of the key objects. The key objects are the local strip areas of ceramics, copper surfaces or green oils that need to be judged for dimensional accuracy;

[0102] Map the CAD contour information and the design position information of the key object onto the sub-board image to find the key object;

[0103] A straight line perpendicular to the extension direction of the key object is established, and the length of the line segment intercepted by the straight line and the actual contour of the key object is taken as the actual size of the key object;

[0104] Calculate the average gray value of the pixels on each line segment, form a contour curve with the arrangement of gray values, and use a filter to smooth the contour;

[0105] Calculate the first-order derivative of the smooth contour. The sub-pixel positions of all local extreme values ​​of the first-order derivative are edge candidate points. The edge candidate points are represented by the vector of the derivative pointing to the corresponding position of the smooth grayscale contour. The edge candidate points whose absolute values ​​are greater than a given threshold are regarded as image edges. The discrete edge points on both sides are fitted into straight line segments respectively. The distance between the two edge line segments is the actual size of the focus object.

[0106] Compare the actual size of the key object with the standard size on the CAD drawing. If the difference between the two is within the allowable range of process error, the process of the current sub-board is judged to be qualified, otherwise it is unqualified.

[0107] The same sub-board image is used to complete both appearance inspection and circuit inspection at the same time, with comprehensive inspection types and fewer steps. For circuit inspection, the image boundary on the sub-board image is related, so it is only necessary to accurately extract the image boundary and obtain the actual size, which can be compared with the standard size to distinguish whether it is qualified or not.

[0108] S7. Result determination: If there are real defects and / or bad circuits in the sub-board image, the sub-board is determined to be a bad board, otherwise it is determined to be a qualified board.

[0109] The above are only some embodiments of the present invention. For those skilled in the art, several modifications and improvements can be made without departing from the creative concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A high-speed and high-precision ceramic copper-clad substrate appearance defect and circuit measurement method, characterized in that The steps include: S1. Image acquisition: Use an industrial camera to collect appearance photos of the ceramic copper-clad motherboard to obtain a complete image of the ceramic copper-clad motherboard; S2, cutting out: segmenting and cutting out the image of the ceramic copper-clad motherboard according to the number of daughter boards in the ceramic copper-clad motherboard and the arrangement of the daughter boards, to obtain a daughter board image of each daughter board; S3, partition processing: for each sub-board image, the copper area and ceramic area in the sub-board are extracted respectively by using the different colors of ceramic and copper under the customized light source, and the image is enhanced; S4, defect detection: determine whether the copper area has a first suspected defect, and whether the ceramic area has a second suspected defect. If the first suspected defect or the second suspected defect exists, extract the corresponding defect thumbnail, and then execute step S5. The first suspected defect includes sintering bubbles, lumps, copper defects, oxidation, dirt and scratches, and the second suspected defects include open circuits, short circuits, bosses, line defects and ceramic pollution. Otherwise, execute step S6; S5, secondary screening: the defect thumbnail is transmitted to the classification convolutional neural network for classification, and the first suspected defect and the second suspected defect detected are secondary judged as real defects or over-inspected defects. If they are real defects, the defects are recorded; if they are over-inspected defects, the defects are not recorded; S6, circuit measurement: perform circuit measurement for each sub-board image, and measure the position and size of key objects in some copper areas and ceramic areas according to the drawing specifications. The key objects are local strip areas of ceramics, copper surfaces or green oil that need to be judged for size accuracy. If the size of the key objects exceeds the accuracy range, it is determined that the circuit is defective; S7. Result determination: If there are real defects and / or bad circuits in the sub-board image, the sub-board is determined to be a bad board, otherwise it is determined to be a qualified board.

2. The high-speed and high-precision ceramic copper-clad substrate appearance defect and line measurement method according to claim 1 is characterized in that: In step S1, the complete ceramic copper-clad motherboard image is stitched together by multiple local images; the left-right size of the ceramic copper-clad motherboard is M, and the top-bottom size is N. The range of the local image is the field of view of the industrial camera, the left-right size is m, the top-bottom size is n, the number of puzzles is z=x×y, x is M / m rounded up, y is N / n rounded up, the acquired local images are sorted in the order of left first, right second, top third, and bottom, and then named Image1-Imagez in sequence, and the grayscale value of the pixel point of each local image is calculated.

3. The high-speed and high-precision ceramic copper-clad substrate appearance defect and line measurement method according to claim 2 is characterized in that: Before the image acquisition step is performed, the industrial camera is calibrated first, and the calibration method is as follows: prepare a calibration plate that meets the camera field of view and accuracy requirements, and the calibration plate has a known geometric shape and size; place the calibration plate under the camera in different postures and angles to take a group of pictures; extract feature points from the acquired calibration plate image, calculate the internal and external parameters of the camera and the distortion coefficient of the lens according to the posture transformation of the feature points, and finally apply the obtained distortion coefficient to the local image to convert the local image into a rectangular plane image without distortion.

4. The high-speed and high-precision ceramic copper-clad substrate appearance defect and line measurement method according to claim 2, characterized in that: The local image stitching method is as follows: generate an empty image Image0 that meets the size of the entire ceramic copper-clad motherboard, set the coordinates of the upper left corner vertex of Image0 to (0, 0), and map the grayscale value of each pixel in Image1 to Image0 one by one according to the actual position of the local image in the motherboard, and the mapping starting point starts from (0, 0); Calculate the overlapping pixels of Image2 and Image1, and map the grayscale value of each pixel remaining after removing the overlapping pixels in Image2 to Image0. The mapping starting point of Image2 starts from the upper right corner vertex of the pixel mapped from Image1. According to the order of the partial images, the mapping of Image3-Imagez to Image0 is completed in sequence from left to right and from top to bottom, so as to obtain the complete ceramic copper-clad motherboard image.

5. The high-speed and high-precision ceramic copper-clad substrate appearance defect and line measurement method according to claim 1 is characterized in that: The specific method of the cutout step is: According to the design drawings of the ceramic copper-clad motherboard, the design width of the ceramic copper-clad motherboard, the design height of the ceramic copper-clad motherboard, the design size of the sub-board, the design height of the horizontal cutting line and the design width of the vertical cutting line are extracted. The sub-boards are arranged in an array. The range of the sub-boards is composed of the minimum rectangular copper-containing area and the ceramic boundary around it. The design size of the sub-board includes the outward expansion width and outward expansion height of the ceramic boundary. The horizontal cutting line is between the upper and lower adjacent sub-boards and runs through the width direction of the ceramic copper-clad motherboard from left to right. The vertical cutting line is between the left and right adjacent sub-boards and runs through the height direction of the ceramic copper-clad motherboard from top to bottom. The horizontal cutting line intersects with the vertical cutting line to form a grid-shaped cutting area; Taking the designed height of the horizontal cutting line as the height and the designed width of the ceramic copper-clad motherboard as the width, a rectangular structural element W is generated, and the ceramic copper-clad motherboard image is traversed using W, and the ceramic area that can be covered by W is retained to obtain images of several horizontal dividing strips; taking the designed width of the vertical cutting line as the width and the designed height of the ceramic copper-clad motherboard as the height, a rectangular structural element H is generated, and the ceramic copper-clad motherboard image is traversed using H, and the ceramic area that can be covered by H is retained to obtain images of several vertical dividing strips; combining the images of all the horizontal dividing strips and the images of the vertical dividing strips to obtain an image of the complete dividing strip; The image of the separation zone is removed from the image of the ceramic copper-clad motherboard to obtain images of several minimum rectangular copper-containing areas, the four corner points of each minimum rectangular copper-containing area are located, and the outer expansion corresponding to the four corner points is determined according to the width and height of the ceramic boundary in the motherboard drawing, thereby determining the positions of the four outer expansion vertices corresponding to the four corner points, and the four outer expansion vertices are connected to obtain the actual daughterboard outline; All daughter board images are cut out from the ceramic copper-clad motherboard image according to the actual daughter board outline, and the individual daughter board images are sorted in the order of top first then bottom, and left first then right.

6. The high-speed and high-precision ceramic copper-clad substrate appearance defect and line measurement method according to claim 1 is characterized in that: The specific method of image enhancement is: Use Fourier transform on the sub-plate image obtained by cutting out, convert the image from the spatial domain to the frequency domain, and obtain the original frequency domain image ImageP; Generate a Gaussian frequency domain filter with specific specifications and specific resolution according to the size of the sub-board image, multiply the value of each pixel on ImageP by the value of the corresponding pixel on the frequency domain filter, and obtain a new frequency domain image ImagePF; Use inverse Fourier transform on ImagePF to convert the image from frequency domain to spatial domain to obtain a new spatial domain image PFImage; Subtract the grayscale value of the corresponding pixel on PFImage from the grayscale value of each pixel on the sub-plate image, and multiply the difference by an enhancement coefficient to obtain the enhanced sub-plate image ImageZ.

7. The high-speed and high-precision ceramic copper-clad substrate appearance defect and line measurement method according to claim 6 is characterized in that: The specific method of the defect detection is: For images with obvious background / foreground distinction after enhancement, use binarization to extract suspected pixels of defects, and group suspected pixels with the same gray value range together as suspected areas; Use Blob analysis to extract and mark the connected domains of the suspected area. Each marked Blob represents a suspected target. The features of the suspected targets are calculated, and the suspected targets are screened according to the features of the sub-board appearance defects. The ones that are finally retained are regarded as suspected defects.

8. The high-speed and high-precision ceramic copper-clad substrate appearance defect and line measurement method according to claim 7 is characterized in that: The specific method of the secondary screening is: Based on the convolutional neural network YOLO-V8 model, the real defect images and inspected defect images collected in the early stage are used as training samples, and the classification model is obtained after training; Input the defect thumbnail corresponding to the suspected defect area into the classification model to obtain the classification result of the defect image output by the classification convolutional neural network; Judging the classification results, if the defect category belongs to a category of real defects, the suspected defect is a real defect, otherwise it is an over-inspected defect and does not need to be detected; If no real inspection defect is detected, the process of the current sub-board is qualified, otherwise it is unqualified.

9. The high-speed and high-precision ceramic copper-clad substrate appearance defect and line measurement method according to claim 1, characterized in that: The line measurement method is: Import the CAD drawing of the daughter board, extract the contour information, the design position information of the key object and the standard size of the key object on the CAD drawing, wherein the key object is a local strip area of ​​ceramic, copper surface or green oil that needs to be judged for dimensional accuracy; Mapping the CAD contour information and the design position information of the key object onto the sub-board image to find the key object; Establishing a straight line perpendicular to the extension direction of the key object, and taking the length of the line segment intercepted by the straight line and the actual outline of the key object as the actual size of the key object; Calculate the average gray value of the pixels on each line segment, form a contour curve with the arrangement of gray values, and use a filter to smooth the contour; Calculate the first-order derivative of the smooth contour. The sub-pixel positions of all local extreme values ​​of the first-order derivative are edge candidate points. The edge candidate points are represented by the vector of the derivative pointing to the corresponding position of the smooth grayscale contour. The edge candidate points whose absolute values ​​are greater than a given threshold are regarded as image edges. The discrete edge points on both sides are fitted into straight line segments respectively. The distance between the two edge line segments is the actual size of the focus object. The actual size of the key object is compared with the standard size on the CAD drawing. If the difference between the two is within the allowable range of process error, the process of the current sub-board is determined to be qualified, otherwise it is unqualified.

Images