A wafer six-side detection system and method

Abstract

The present invention discloses a wafer six-side inspection system and method, comprising: a grain loading and correction module for determining the position of the grain to be inspected and obtaining a position compensation value for the grain; a surface feature data acquisition module for acquiring grain surface feature data based on the grain surface profile image; a grain surface quality module for constructing a grain surface quality assessment model and identifying the grain surface quality; a side feature data acquisition module for acquiring grain side feature data based on the grain side profile image; a grain side quality module for constructing a grain side quality assessment model and identifying the grain side quality; and a grain blanking processing module for performing blanking processing on the grain. The technical method provided by the present invention significantly improves the accuracy, efficiency, and intelligence of grain inspection, has broad applicability and scalability, and can be widely used in grain quality inspection for various process types and materials.

Description

Technical Field

[0001] The present invention relates to the field of new generation information technology, and in particular to a wafer six-side detection system and method. Background Art

[0002] With the continuous advancement of semiconductor manufacturing processes, the size of wafer dies is shrinking, their functions are becoming increasingly complex, and the requirements for their appearance quality are becoming increasingly stringent. Before and after wafer dicing and packaging, the dies undergo multi-faceted inspection, especially on the surface and sides. The quality of these inspections directly impacts the final chip performance, reliability, and yield. However, the industry still faces several technical bottlenecks and practical application challenges in six-sided inspection of dies. First, in terms of image feature extraction, traditional inspection methods, which rely primarily on basic algorithms such as edge detection and threshold segmentation, have limited processing power and are unable to cope with complex texture variations, microcracks, defects, or blurred boundaries on the die surface. Die side profiles are further affected by factors such as image angle and edge reflections. This leads to unstable extraction of side image features and a lack of quantitative indicators, making it difficult to provide a reliable basis for subsequent quality assessment. Furthermore, quality assessment methods rely on human experience or fixed thresholds, which are difficult to adapt to image feature variations caused by differences in material type, process type, and specifications. In actual production, the same defect type may exhibit different image characteristics under different process conditions. Fixed threshold judgment results in a high rate of false positives, especially during batch processing, and prevents refined and dynamic quality assessment. Furthermore, variations in lighting conditions are a major interfering factor in grain image inspection. In actual production lines, fluctuations in light intensity, slight changes in illumination angle, or differences in grain surface reflectivity can cause uneven image brightness and grayscale, which in turn affects the stability and accuracy of feature extraction. Existing systems are generally insufficiently robust to illumination variations and lack pixel-level compensation mechanisms. Finally, in terms of multi-process and multi-material compatibility, due to the significant differences in grain surface conditions across materials such as metals, ceramics, and polycrystalline, and processes such as casting, welding, and cutting, existing systems often require frequent manual parameter or model adjustments, are unable to automatically adapt to process variations, and lack versatility and scalability, resulting in complex system debugging and limited applicability. Therefore, existing wafer inspection methods suffer from significant deficiencies in image feature extraction stability, intelligent quality assessment, illumination adaptability, and multi-process compatibility, urgently requiring improvement and optimization. Summary of the Invention

[0003] The present invention addresses the problems existing in the above-mentioned prior art and provides a wafer six-side detection system and method.

[0004] A first embodiment of the present invention provides a wafer six-side inspection system, mainly comprising:

[0005] The grain loading and deflection correction module is used to obtain the grain image to be inspected through the loading and deflection correction camera, convert the edge points in the edge contour of the grain into geometric parameters using the Hough transform algorithm, determine the position of the grain to be inspected, and obtain the position compensation value of the grain;

[0006] The surface feature data acquisition module is used to extract the surface contour image of the grain using the Canny edge detection algorithm, and based on the surface contour image of the grain, obtain the shape factor, reflectivity gradient characteristics and texture characteristics of the grain surface contour image to construct the grain surface feature data;

[0007] The grain surface quality module is used to build a grain surface quality assessment model and identify the grain surface quality based on the grain surface profile image feature data, material, process type and specifications;

[0008] The side feature data acquisition module is used to extract the side profile of the grain using the Canny edge detection algorithm, and based on the grain side profile image, obtain the shape factor, texture characteristics, grain boundary width and grain boundary area of ​​the grain side profile image to construct the grain side feature data;

[0009] The die side quality module is used to build a die side quality assessment model and identify the die side quality based on the die side profile image feature data, material, process type and specifications;

[0010] The grain blanking processing module is used to perform grain blanking processing on the grains according to the grain surface quality assessment results and the grain side quality assessment results.

[0011] A second embodiment of the present invention provides a six-side wafer inspection method, which mainly includes:

[0012] The image of the grain to be inspected is acquired through the loading correction camera. The edge points in the edge contour of the grain are converted into geometric parameters using the Hough transform algorithm to determine the position of the grain to be inspected and obtain the position compensation value of the grain.

[0013] The surface contour image of the grain is extracted by the Canny edge detection algorithm. Based on the surface contour image of the grain, the shape factor, reflectivity gradient characteristics and texture characteristics of the grain surface contour image are obtained to construct the grain surface feature data;

[0014] Based on the grain surface profile image feature data, material, process type and specifications, a grain surface quality assessment model is constructed to identify the grain surface quality;

[0015] The side profile of the grain is extracted using the Canny edge detection algorithm. Based on the grain side profile image, the shape factor, texture characteristics, grain boundary width and grain boundary area of ​​the grain side profile image are obtained to construct the grain side feature data;

[0016] Based on the grain side profile image feature data, material, process type and specifications, a grain side quality assessment model is constructed to identify the grain side quality;

[0017] The grains are cut according to the grain surface quality assessment results and the grain side quality assessment results.

[0018] Furthermore, the method of obtaining an image of the grain to be inspected by a loading correction camera, converting edge points in the edge contour of the grain into geometric parameters using a Hough transform algorithm, determining the position of the grain to be inspected, and obtaining a position compensation value of the grain includes:

[0019] The grains are placed on the grain detection position by manual loading, and the optical camera at the grain detection position is used to obtain the grain position information; the position of the grains is corrected by the loading XYR platform, and the grains are lifted up by the ejector module; the grains are placed on the loading transfer module by the first loading head and moved to the picking position of the second loading head; the loading correction camera at the picking position of the second loading head is used to obtain the grain image to be detected, the edge contour of the grain is obtained by the Canny edge detection algorithm, and the edge contour of the grain is centered by the Hough transform algorithm. The edge points are converted into geometric parameters to determine the position of the grain to be inspected; the initial front / back inspection station image position of the grain of the same specification at the correct imaging position is obtained through the historical inspection records of the grain; according to the correct imaging position of the front / back inspection station of the grain, the initial value of the correct imaging position of the side / end inspection station is set, and the initial values ​​of the correct imaging position of the front / back and side / end are recorded; by comparing the initial values ​​of the correct imaging position of the front / back and side / end with the position of the currently inspected grain, the position compensation value of the grain is obtained.

[0020] Furthermore, the surface contour image of the grain is extracted by the Canny edge detection algorithm, and based on the surface contour image of the grain, the shape factor, reflectivity gradient characteristics and texture characteristics of the grain surface contour image are obtained to construct the grain surface feature data, including:

[0021] The second loading head picks up the grain and places it on the inspection table. The optical camera at the front / back inspection station of the grain is used to obtain the front / back image of the grain. The Canny edge detection algorithm is used to extract the surface contour image of the grain. Based on the surface contour image of the grain, the perimeter and area of ​​the grain surface contour image are determined, and the shape factor calculation formula is used. , calculate the shape factor of the grain surface profile image, where S is the shape factor of the grain, is the perimeter of the grain surface contour, The method comprises the following steps: obtaining the brightness or color distribution data of the grain surface profile image, calculating the reflectivity of each area by the change of the image brightness, obtaining the spatial distribution data of the surface reflectivity, and drawing a reflectivity distribution diagram; performing statistical analysis on the spatial distribution data of the surface reflectivity of the grain surface profile image, obtaining the reflectivity gradient characteristics of the grain surface, including the average reflectivity, the maximum reflectivity and the standard deviation of the gradient direction; converting the RGB channel values ​​of the image into a single grayscale value according to the grain surface profile image using the weighted average method, and obtaining the grayscale image of the grain surface profile; extracting the grayscale co-occurrence matrix according to the grayscale image of the grain surface profile, and calculating the homogeneity based on the formula , calculate the homogeneity H, and based on the contrast calculation formula , calculate the contrast C, and use homogeneity and contrast as texture features of the grain surface profile image, where i is the gray value of a pixel in the image, which is used as the reference pixel, and j is the gray value of the adjacent pixel paired with pixel i in a specific direction and distance, is the value of the i, j position in the gray-level co-occurrence matrix, indicating the frequency of occurrence of the gray-level value i and j pair; the shape factor, reflectivity gradient characteristics and texture characteristics of the grain surface profile image are standardized and combined into a feature vector to constitute the grain surface feature data, which is stored in the grain image monitoring database.

[0022] Furthermore, the step of constructing a grain surface quality assessment model based on grain surface feature data, material, process type, and specifications to identify grain surface quality includes:

[0023] Through the grain image monitoring database, historical grain surface feature data, materials, process types and specifications are obtained, the grain surface quality is marked, and the random forest algorithm is used for model training to build a grain surface quality assessment model. The process types include but are not limited to casting, forging, cutting, welding and heat treatment processes, and the materials include but are not limited to metal materials, ceramic materials and polycrystalline materials. The grain surface quality includes high quality, medium and low quality; based on the real-time acquired grain surface feature data, process type and specifications, the grain surface quality assessment model is used to identify the grain surface quality.

[0024] Furthermore, the side profile of the grain is extracted by the Canny edge detection algorithm, and based on the grain side profile image, the shape factor, texture characteristics, grain boundary width and grain boundary area of ​​the grain side profile image are obtained to construct the grain side feature data, including:

[0025] If the grain on the inspection table moves to the side / end inspection station, the electric slide equipped with the optical camera is controlled to move to the position and focus the image according to the position compensation value of the grain, and the side / end image of the grain is obtained. The side profile of the grain is extracted using the Canny edge detection algorithm, the grain thickness is identified, and the shape factor calculation formula is used to calculate the shape factor of the grain side profile image; based on the grain side profile image, the RGB channel value of the image is converted into a single grayscale value using the weighted average method; if the electric slide equipped with the optical camera moves and the illumination angle changes, the grayscale value compensation unit is used to compensate the grayscale value of each pixel in the grayscale image of the grain side profile according to the illumination angle; based on the compensated grayscale value, the grayscale image of the grain side profile is obtained, and the grayscale co-occurrence matrix is ​​extracted; based on the grayscale co-occurrence matrix of the grain side profile image, the homogeneity is calculated based on the homogeneity calculation formula, and based on the contrast The contrast is calculated using the degree calculation formula, and the homogeneity and contrast are used as the texture features of the grain side profile image; if the grain material is a polycrystalline material, the grayscale image of the grain side profile is binarized, and the area with grayscale below the preset grayscale threshold is marked as the grain boundary area; the grain boundary is refined into a single pixel width, and the distance from the skeleton to the nearest non-grain boundary pixel in the image is calculated to obtain the local grain boundary width of each pixel, and the width values ​​of all grain boundary pixels are averaged to obtain the average grain boundary width; based on the average grain boundary width and the actual pixel size, the physical width is converted to obtain the grain boundary width; based on the number of pixels in the grain boundary area and the actual area of ​​each pixel, the physical area is converted to obtain the grain boundary area; the shape factor, texture features, grain boundary width and grain boundary area of ​​the grain side profile image are standardized and combined into a feature vector to constitute the grain side feature data, which is stored in the grain image monitoring database.

[0026] The method further includes compensating the grayscale value of each pixel in the grain side profile grayscale image according to the illumination angle.

[0027] The method of compensating the grayscale value of each pixel in the grayscale image of the grain side profile according to the illumination angle specifically includes:

[0028] For each pixel in the grain side profile image, calculate the average grayscale value of the neighborhood around the pixel, and use the illumination compensation formula based on the illumination angle , calculate the lighting compensation factor ,in, is the original grayscale value of the pixel (x,y) in the image, is the average grayscale value of the neighborhood around the pixel, which is used to measure the local light intensity. is a constant to prevent division by zero errors and is obtained by fitting historical data. Is a coefficient that controls the intensity of the impact of light changes and is obtained by fitting historical data. It is the angle between the normal of the pixel and the illumination direction, indicating the change in illumination angle and reflecting the change in surface illumination conditions. According to the illumination compensation factor, the formula is used. , perform illumination compensation on each pixel in the grain side profile image to obtain the compensated grayscale value .

[0029] Furthermore, the step of constructing a die side quality assessment model based on die side feature data, material, process type, and specifications to identify die side quality includes:

[0030] Through the grain image monitoring database, historical grain side feature data, materials, process types and specifications are obtained, the grain side quality is marked, and the random forest algorithm is used for model training to build a grain side quality assessment model. The process types include but are not limited to casting, forging, cutting, welding and heat treatment processes, and the materials include but are not limited to metal materials, ceramic materials and polycrystalline materials. The grain side quality includes high quality, medium and low quality. According to the real-time acquired grain side feature data, process type and specifications, the grain side quality assessment model is used to identify the grain side quality.

[0031] Furthermore, the step of performing blanking on the grains according to the grain surface quality assessment results and the grain side quality assessment results includes:

[0032] The grain blanking processing method is determined based on the grain surface quality assessment results and the grain side quality assessment results; if the grain surface quality assessment results or the grain side quality are poor quality, the grain is removed from the inspection table through the NG blanking head and placed in order on the blanking wide crystal ring; if the grain surface quality assessment results and the grain side quality are not poor quality, the grain is removed from the inspection table through the OK blanking head and placed on the OK blanking carrier; if the position compensation value of the grain is greater than the preset compensation value threshold, the grain is removed from the inspection table through the waste discharge head and placed on the re-inspection carrier for re-grain inspection.

[0033] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0034] The present invention provides a six-sided wafer inspection system and method. This system precisely locates the position of the wafer die using a loading correction camera and a Hough transform algorithm, avoiding detection errors caused by positional offsets in traditional methods and significantly improving positioning accuracy. Based on multidimensional features extracted from surface and side images of the wafer, including shape factors, texture features, and reflectivity gradients, the system provides rich and reliable data support for wafer quality assessment, making wafer quality assessment more comprehensive and accurate. By constructing surface and side quality assessment models and integrating information about the wafer's material, process type, and specifications, the system intelligently identifies the quality of the wafer, significantly improving identification stability and accuracy. The system also incorporates a compensation mechanism for illumination variations, effectively eliminating interference caused by differences in illumination angles, enhancing the system's adaptability to environmental changes, and ensuring the stability of image feature extraction. In terms of material unloading control, the system automatically rejects inferior wafers and accurately sorts qualified wafers through intelligent sorting, improving the automation and intelligence of unloading decisions. The present invention significantly improves the accuracy, efficiency and intelligence of grain detection, has wide applicability and scalability, and can be widely used in grain quality detection of various process types and materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flow chart of a wafer six-side inspection system of the present invention;

[0036] Figure 2 This is a flow chart of a wafer six-side detection method of the present invention;

[0037] Figure 3 Schematic diagram of a wafer six-side detection method of the present invention. DETAILED DESCRIPTION

[0038] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] like Figure 1 In this embodiment, a wafer six-side inspection system may specifically include:

[0040] The grain loading and deflection correction module is used to obtain the grain image to be inspected through the loading and deflection correction camera, convert the edge points in the edge contour of the grain into geometric parameters using the Hough transform algorithm, determine the position of the grain to be inspected, and obtain the position compensation value of the grain;

[0041] The surface feature data acquisition module is used to extract the surface contour image of the grain using the Canny edge detection algorithm, and based on the surface contour image of the grain, obtain the shape factor, reflectivity gradient characteristics and texture characteristics of the grain surface contour image to construct the grain surface feature data;

[0042] The grain surface quality module is used to build a grain surface quality assessment model and identify the grain surface quality based on the grain surface profile image feature data, material, process type and specifications;

[0043] The side feature data acquisition module is used to extract the side profile of the grain using the Canny edge detection algorithm, and based on the grain side profile image, obtain the shape factor, texture characteristics, grain boundary width and grain boundary area of ​​the grain side profile image to construct the grain side feature data;

[0044] The die side quality module is used to build a die side quality assessment model and identify the die side quality based on the die side profile image feature data, material, process type and specifications;

[0045] The grain blanking processing module is used to perform grain blanking processing on the grains according to the grain surface quality assessment results and the grain side quality assessment results.

[0046] like Figure 2-3 The present embodiment provides a wafer six-side detection method, which may specifically include:

[0047] Step S101 , obtaining an image of the grain to be inspected through a loading correction camera, converting edge points in the edge contour of the grain into geometric parameters using a Hough transform algorithm, determining the position of the grain to be inspected, and obtaining a position compensation value of the grain.

[0048] The grains are placed on the grain detection position by manual loading, and the optical camera at the grain detection position is used to obtain the grain position information; the position of the grains is corrected by the loading XYR platform, and the grains are lifted up by the ejector module; the grains are placed on the loading transfer module by the first loading head and moved to the picking position of the second loading head; the loading correction camera at the picking position of the second loading head is used to obtain the grain image to be detected, the edge contour of the grain is obtained by the Canny edge detection algorithm, and the edge contour of the grain is centered by the Hough transform algorithm. The edge points are converted into geometric parameters to determine the position of the grain to be inspected; the initial front / back inspection station image position of the grain of the same specification at the correct imaging position is obtained through the historical inspection records of the grain; according to the correct imaging position of the front / back inspection station of the grain, the initial value of the correct imaging position of the side / end inspection station is set, and the initial values ​​of the correct imaging position of the front / back and side / end are recorded; by comparing the initial values ​​of the correct imaging position of the front / back and side / end with the position of the currently inspected grain, the position compensation value of the grain is obtained.

[0049] For example, during the die inspection process, the die is manually loaded and placed at the die inspection location. After measurement by the optical camera at this location, the positions of the four boundaries of the die are: side a' = 305mm, side b' = 105mm, side c' = 315mm, and side d' = 115mm. These are the measured values ​​of the die at the actual inspection location. Position correction is performed using the loading XYR platform. The XYR platform adjusts the lateral and longitudinal position of the die to ensure that the die is aligned with the correct position of the subsequent inspection station. The die is lifted up using a pin module to prevent any slight error in the die position from affecting the subsequent inspection process. The first loading head then removes the die from the position correction area and places it on the loading transfer module. The transfer module then moves the die to the material removal position of the second loading head in preparation for the final image inspection. A loading and correction camera is installed at the retrieving position of the second loading bar. This camera is responsible for capturing images of the die. Using the grain images captured by the loading and correction camera, the Canny edge detection algorithm is used to extract the edge contour of the die. The Hough transform algorithm then converts the edge points in the outline of the die into geometric parameters to determine the precise position of the currently inspected die. The calculated positions of the four boundaries of the die are: a'=305mm, b'=105mm, c'=315mm, and d'=115mm. These data indicate that the boundaries of the die have slightly deviated from the standard position. Based on historical inspection records of the die, the initial image positions of the same specification die at the standard imaging position are obtained. Historical data shows that when the die is in the correct imaging position, the image positions at the front / back inspection stations are: a=300mm, b=100mm, c=310mm, and d=110mm. These are the measured values ​​of the four boundaries of the die at the initial standard position. Based on this historical data, the initial position values ​​for the side / end inspection stations are set based on the correct imaging positions at the front / back inspection stations. These initial values ​​are determined based on the standard positions of the die at other stations and recorded for reference. By comparing the current actual position of the die with the standard positions stored in the historical data, the die position compensation value is calculated. If, during actual measurement, the die position offsets as follows: the difference between a' and a at the front inspection station is 305mm - 300mm = 5mm, the difference between b' and b is 105mm - 100mm = 5mm, the difference between c' and c at the back inspection station is 315mm - 310mm = 5mm, and the difference between d' and d is 115mm - 110mm = 5mm, the calculated position compensation value is a +5mm offset for all sides, meaning that the actual position of the die at the front and back inspection stations is offset by 5mm relative to the standard position.

[0050] Step S102 , extracting the surface contour image of the grain using the Canny edge detection algorithm, and based on the surface contour image of the grain, obtaining the shape factor, reflectivity gradient characteristics and texture characteristics of the surface contour image of the grain, and constructing the surface feature data of the grain.

[0051] The second loading head picks up the grain and places it on the inspection table. The optical camera at the front / back inspection station of the grain is used to obtain the front / back image of the grain. The Canny edge detection algorithm is used to extract the surface contour image of the grain. Based on the surface contour image of the grain, the perimeter and area of ​​the grain surface contour image are determined, and the shape factor calculation formula is used. , calculate the shape factor of the grain surface profile image, where S is the shape factor of the grain, is the perimeter of the grain surface contour, The method comprises the following steps: obtaining the brightness or color distribution data of the grain surface profile image, calculating the reflectivity of each area by the change of the image brightness, obtaining the spatial distribution data of the surface reflectivity, and drawing a reflectivity distribution diagram; performing statistical analysis on the spatial distribution data of the surface reflectivity of the grain surface profile image, obtaining the reflectivity gradient characteristics of the grain surface, including the average reflectivity, the maximum reflectivity and the standard deviation of the gradient direction; converting the RGB channel values ​​of the image into a single grayscale value according to the grain surface profile image using the weighted average method, and obtaining the grayscale image of the grain surface profile; extracting the grayscale co-occurrence matrix according to the grayscale image of the grain surface profile, and calculating the homogeneity based on the formula , calculate the homogeneity H, and based on the contrast calculation formula , calculate the contrast C, and use homogeneity and contrast as texture features of the grain surface profile image, where i is the gray value of a pixel in the image, which is used as the reference pixel, and j is the gray value of the adjacent pixel paired with pixel i in a specific direction and distance, is the value of the i, j position in the gray-level co-occurrence matrix, indicating the frequency of occurrence of the gray-level value i and j pair; the shape factor, reflectivity gradient characteristics and texture characteristics of the grain surface profile image are standardized and combined into a feature vector to constitute the grain surface feature data, which is stored in the grain image monitoring database.

[0052] For example, during a die inspection process, the second loading head successfully picks up the die and places it on the inspection table. The front-side image of the die is then captured using an optical camera at the front / back inspection station. Through image processing, the Canny edge detection algorithm is used to extract the surface contour of the die, and the perimeter of the die surface contour is identified as 200 mm and the area is 50 mm². Based on the extracted contour data, the shape factor calculation formula is used. , calculate the shape factor S of the grain, if the perimeter of the grain 200mm, area If the area of ​​the grain is 50 mm², the shape factor S≈25.46. Using the image brightness or color distribution data, the reflectivity of each area on the grain surface is calculated to obtain the spatial distribution map of the reflectivity. After further analysis, the average reflectivity of the grain surface area is 0.8, the maximum reflectivity is 0.9, and the standard deviation of the gradient direction of the reflectivity is 0.15. These data reveal the illumination reflection characteristics of the grain surface in different areas and the uniformity of the surface reflection. The RGB channel values ​​of the image are converted into a single grayscale value using the weighted average method, and the grayscale image of the grain surface contour is obtained. Based on the grayscale image, the grayscale co-occurrence matrix is ​​extracted, and according to the matrix, combined with the homogeneity calculation formula And contrast calculation formula Homogeneity H and contrast C were calculated as characteristics of the grain surface texture. The calculated homogeneity H was 0.75 and contrast C was 0.65, which represent the smoothness and diversity of the grain surface texture. The shape factor, reflectivity gradient characteristics, and texture characteristics were standardized to generate a comprehensive feature vector, which constitutes the grain surface feature data and is stored in the grain image monitoring database.

[0053] Step S103 : constructing a grain surface quality assessment model based on the grain surface feature data, material, process type and specification to identify the grain surface quality.

[0054] Through the grain image monitoring database, historical grain surface feature data, materials, process types and specifications are obtained, the grain surface quality is marked, and the random forest algorithm is used for model training to build a grain surface quality assessment model. The process types include but are not limited to casting, forging, cutting, welding and heat treatment processes, and the materials include but are not limited to metal materials, ceramic materials and polycrystalline materials. The grain surface quality includes high quality, medium and low quality; based on the real-time acquired grain surface feature data, process type and specifications, the grain surface quality assessment model is used to identify the grain surface quality.

[0055] For example, during a grain surface quality inspection process, the characteristic data of the historical grain surface profile image, the corresponding material, process type and specification are obtained through the grain image monitoring database. From the historical data, a certain specification of grains was extracted, which used a casting process and a metal material, and was marked as high-quality surface quality during the inspection. Specifically, the surface profile image characteristic data of the grain includes a shape factor of 20.5, an average reflectivity of 0.85, a maximum reflectivity of 0.9, a gradient direction standard deviation of 0.1, a grayscale image homogeneity value of 0.8, and a contrast value of 0.6. These historical data are used to train the random forest algorithm to construct a grain surface quality assessment model, which predicts the surface quality grade of the grain based on the characteristic data of the grain surface profile as well as the process type and material. During the training process, the model learns and identifies the characteristics of the surfaces of grains of different quality grades by inputting a large amount of historical data. The quality grades include high quality, medium quality and low quality. During real-time detection, the surface profile image feature data of the current grain to be detected is obtained. If the image features of the grain are a shape factor of 18.7, an average reflectivity of 0.75, a maximum reflectivity of 0.8, a gradient direction standard deviation of 0.15, a homogeneity value of 0.65, and a contrast value of 0.55, and it is obtained that the grain belongs to a forging process, the material is a metal material, and the specifications match the known specifications in the historical data, the trained grain surface quality assessment model is used, and the real-time obtained grain surface profile feature data, process type, and specifications are input into the model for prediction, and the surface quality of the current grain is judged to be medium.

[0056] Step S104 , extracting the side profile of the grain using the Canny edge detection algorithm, and based on the grain side profile image, obtaining the shape factor, texture feature, grain boundary width and grain boundary area of ​​the grain side profile image, and constructing grain side feature data.

[0057] If the grain on the inspection table moves to the side / end inspection station, the electric slide equipped with the optical camera is controlled to move to the position and focus the image according to the position compensation value of the grain, and the side / end image of the grain is obtained. The side profile of the grain is extracted using the Canny edge detection algorithm, the grain thickness is identified, and the shape factor calculation formula is used to calculate the shape factor of the grain side profile image; based on the grain side profile image, the RGB channel value of the image is converted into a single grayscale value using the weighted average method; if the electric slide equipped with the optical camera moves and the illumination angle changes, the grayscale value compensation unit is used to compensate the grayscale value of each pixel in the grayscale image of the grain side profile according to the illumination angle; based on the compensated grayscale value, the grayscale image of the grain side profile is obtained, and the grayscale co-occurrence matrix is ​​extracted; based on the grayscale co-occurrence matrix of the grain side profile image, the homogeneity is calculated based on the homogeneity calculation formula, and based on the contrast The contrast is calculated using the degree calculation formula, and the homogeneity and contrast are used as the texture features of the grain side profile image; if the grain material is a polycrystalline material, the grayscale image of the grain side profile is binarized, and the area with grayscale below the preset grayscale threshold is marked as the grain boundary area; the grain boundary is refined into a single pixel width, and the distance from the skeleton to the nearest non-grain boundary pixel in the image is calculated to obtain the local grain boundary width of each pixel, and the width values ​​of all grain boundary pixels are averaged to obtain the average grain boundary width; based on the average grain boundary width and the actual pixel size, the physical width is converted to obtain the grain boundary width; based on the number of pixels in the grain boundary area and the actual area of ​​each pixel, the physical area is converted to obtain the grain boundary area; the shape factor, texture features, grain boundary width and grain boundary area of ​​the grain side profile image are standardized and combined into a feature vector to constitute the grain side feature data, which is stored in the grain image monitoring database.

[0058] For example, during the surface quality inspection of the grain, the grain is moved to the side / end face inspection station, and the electric slide is equipped with an optical camera to perform position correction according to the position compensation value of the grain. Through system control, the electric slide accurately moves the camera into position and focuses, successfully acquiring the side image of the grain. Next, the Canny edge detection algorithm is used to extract the side profile image of the grain, and based on this profile information, the system calculates the thickness of the grain. It is known that the side profile circumference of the grain is 180mm and the area is 40mm². Therefore, the shape factor S≈24.3 is calculated through the shape factor calculation formula, indicating that the shape complexity of the side profile of the grain is high. When processing the side image of the grain, the RGB channel values ​​of the image are converted into a single grayscale value by the weighted averaging method, and the color information of the image is converted into grayscale information. If the optical camera mounted on the electric slide causes the illumination angle to change during movement, the grayscale value of each pixel will be compensated according to the illumination angle to reduce the impact of illumination changes on image quality. If the illumination angle change causes the grayscale value of certain areas in the image to decrease by 5%, the system will automatically correct the grayscale image according to this change to obtain a compensated grayscale image. The compensated grayscale image is used to extract the grayscale co-occurrence matrix. According to the homogeneity calculation formula and the contrast calculation formula, the system calculates the homogeneity H and contrast C of the image respectively. After calculation, the homogeneity H is 0.78 and the contrast C is 0.65. These two features represent the texture information of the grain surface, indicating that the changes on the grain surface are relatively smooth, but still have a certain contrast. If the grain material is a polycrystalline material, the grayscale image is binarized, and according to the preset grayscale threshold of 0.5, all areas below the threshold are marked as grain boundary areas. Through image processing, the grain boundary area is refined into a single pixel width, and the distance from the skeleton to the nearest non-grain boundary pixel is calculated to obtain the local grain boundary width of each pixel. The average value of the grain boundary width of the image is measured to be 10μm. By converting this width with the pixel size of the image, the actual physical width of the grain boundary is 0.1mm. In the binarized image, the number of pixels in the grain boundary area is 500, and each pixel represents an area of ​​0.01mm², so the total area of ​​the grain boundary area is 5mm². The shape factor, texture characteristics, grain boundary width and grain boundary area of ​​the grain are standardized, and these features are combined into a feature vector to constitute the grain side feature data and stored in the grain image monitoring database.

[0059] The grayscale value of each pixel in the grain side profile grayscale image is compensated according to the illumination angle.

[0060] For each pixel in the grain side profile image, calculate the average grayscale value of the neighborhood around the pixel, and use the illumination compensation formula based on the illumination angle , calculate the lighting compensation factor ,in, is the original grayscale value of the pixel (x,y) in the image, is the average grayscale value of the neighborhood around the pixel, which is used to measure the local light intensity. is a constant to prevent division by zero errors and is obtained by fitting historical data. Is a coefficient that controls the intensity of the impact of light changes and is obtained by fitting historical data. It is the angle between the normal of the pixel and the illumination direction, indicating the change in illumination angle and reflecting the change in surface illumination conditions. According to the illumination compensation factor, the formula is used. , perform illumination compensation on each pixel in the grain side profile image to obtain the compensated grayscale value .

[0061] For example, during the processing of the grain side profile image, the original grayscale value of each pixel has been obtained, where the coordinates of a specific pixel point are (x, y), and its original grayscale value is =150, calculate the average gray value of the neighborhood around the pixel =140. According to the light compensation formula , calculate the illumination compensation factor illumination compensation factor ,in, is the original grayscale value of the pixel (x,y) in the image, is the average grayscale value of the neighborhood around the pixel, which is used to measure the local light intensity. is a constant to prevent division by zero errors and is obtained by fitting historical data. Is a coefficient that controls the intensity of the impact of light changes and is obtained by fitting historical data. It is the angle between the normal line of the pixel and the direction of illumination, which indicates the change in the angle of illumination and reflects the change in the surface illumination conditions. Substituting the known value into the illumination compensation formula, we get the illumination compensation factor. =1.1885. According to the light compensation factor, use the formula Perform illumination compensation on each pixel in the grain side profile image to obtain the compensated grayscale value =178.275. Therefore, the pixel grayscale value after illumination compensation is 178.275.

[0062] Step S105 : constructing a die side surface quality assessment model based on the die side surface feature data, material, process type, and specification to identify the die side surface quality.

[0063] Through the grain image monitoring database, historical grain side feature data, materials, process types and specifications are obtained, the grain side quality is marked, and the random forest algorithm is used for model training to build a grain side quality assessment model. The process types include but are not limited to casting, forging, cutting, welding and heat treatment processes, and the materials include but are not limited to metal materials, ceramic materials and polycrystalline materials. The grain side quality includes high quality, medium and low quality. According to the real-time acquired grain side feature data, process type and specifications, the grain side quality assessment model is used to identify the grain side quality.

[0064] For example, historical grain side feature data, materials, process types, and specifications are extracted from the grain image monitoring database, and the grain side quality is marked. Sample 1 has a process type of casting, a material type of metal material, a high-quality grain quality, a shape factor of 22.3, an average reflectivity of 0.84, a homogeneity of 0.79, a contrast of 0.72, a grain boundary width of 0.12 mm, and a grain boundary area of ​​5.0 mm². Sample 2 has a process type of forging, a material type of metal material, a high-quality grain quality, a shape factor of 21.5, an average reflectivity of 0.82, a homogeneity of 0.77, a contrast of 0.70, a grain boundary width of 0.11 mm, and a grain boundary area of ​​4.8 mm². Sample 3 was machined using a metal material, with medium grain quality, a shape factor of 20.9, an average reflectivity of 0.75, a homogeneity of 0.70, a contrast of 0.65, a grain boundary width of 0.10 mm, and a grain boundary area of ​​4.5 mm². Sample 4 was welded using a ceramic material, with medium grain quality, a shape factor of 24.0, an average reflectivity of 0.78, a homogeneity of 0.72, a contrast of 0.68, a grain boundary width of 0.13 mm, and a grain boundary area of ​​4.6 mm². Sample 5 was heat treated using a polycrystalline material, with poor grain quality, a shape factor of 26.1, an average reflectivity of 0.68, a homogeneity of 0.65, a contrast of 0.62, a grain boundary width of 0.15 mm, and a grain boundary area of ​​5.2 mm². Sample 6 has a casting process, a ceramic material, a poor grain quality, a shape factor of 27.5, an average reflectivity of 0.60, a homogeneity of 0.60, a contrast of 0.58, a grain boundary width of 0.16 mm, and a grain boundary area of ​​5.5 mm². Based on these sample data, a random forest algorithm was used for model training to construct a grain side quality assessment model. The grain side quality is classified into high quality, medium quality, and poor quality. During the production process, the side profile image feature data of a grain to be inspected is obtained in real time. If the material type of the grain is also a metal material, the process type is cutting, and the specifications are consistent with those in the training data, the obtained grain side profile feature values ​​in the real-time detection include a shape factor of 22.5, an average reflectivity of 0.78, a homogeneity value of the grayscale co-occurrence matrix of 0.70, a contrast value of 0.65, a grain boundary width of 0.11 mm, and a grain boundary area of ​​4.2 mm². These real-time acquired feature data are input into the trained grain side quality assessment model for prediction, and the side quality of the grain is determined to be medium.

[0065] Step S106 , performing blanking processing on the grains according to the grain surface quality assessment results and the grain side surface quality assessment results.

[0066] The grain blanking processing method is determined based on the grain surface quality assessment results and the grain side quality assessment results; if the grain surface quality assessment results or the grain side quality are poor quality, the grain is removed from the inspection table through the NG blanking head and placed in order on the blanking wide crystal ring; if the grain surface quality assessment results and the grain side quality are not poor quality, the grain is removed from the inspection table through the OK blanking head and placed on the OK blanking carrier; if the position compensation value of the grain is greater than the preset compensation value threshold, the grain is removed from the inspection table through the waste discharge head and placed on the re-inspection carrier for re-grain inspection.

[0067] For example, the surface quality assessment results and side quality assessment results of the die are obtained, and the die is blanked. If the surface quality assessment result or the side quality of the die is poor, the die is removed from the inspection table using the NG blanking header and placed in an orderly manner into the blanking wide crystal ring, ensuring that the die that does not meet the quality requirements will not continue to enter the production process, thereby avoiding affecting the quality of subsequent processing. If the surface quality assessment result and the side quality of the die are both not poor, the die is removed from the inspection table using the OK blanking header and placed into the OK blanking carrier. The high-quality die will be transferred to the downstream process for further processing to ensure that the product quality meets the standard. If the position compensation value of a die is 1.2mm, and the system's preset compensation threshold is 1.0mm, because the die's position compensation value exceeds the threshold, the die will be removed from the inspection table using a waste discharge head and placed on the re-inspection carrier for re-inspection. This ensures that the quality and specifications of the die with large position errors can be confirmed after re-inspection, preventing unqualified die from continuing to enter the production process. Therefore, during this inspection and processing process, although the surface quality of the die was assessed as medium, the side quality was poor and the position compensation value exceeded the threshold. Ultimately, the die was processed through the NG blanking head and waste discharge head, ensuring quality compliance in production.

[0068] The above description is merely a preferred embodiment of the present application and an illustration of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to the technical solutions formed by the specific combination of the above-mentioned technical features, but also encompasses other technical solutions formed by any combination of the above-mentioned technical features or their equivalents without departing from the concept of this application. For example, a technical solution formed by replacing the above-mentioned features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A wafer six-side detection method, characterized in that: The method comprises: The image of the grain to be inspected is acquired through the loading correction camera. The edge points in the edge contour of the grain are converted into geometric parameters using the Hough transform algorithm to determine the position of the grain to be inspected and obtain the position compensation value of the grain. The surface contour image of the grain is extracted by the Canny edge detection algorithm. Based on the surface contour image of the grain, the shape factor, reflectivity gradient characteristics and texture characteristics of the grain surface contour image are obtained to construct the grain surface feature data; Based on the grain surface profile image feature data, material, process type and specifications, a grain surface quality assessment model is constructed to identify the grain surface quality; The side profile of the grain is extracted using the Canny edge detection algorithm. Based on the grain side profile image, the shape factor, texture characteristics, grain boundary width and grain boundary area of ​​the grain side profile image are obtained to construct the grain side feature data; Based on the grain side profile image feature data, material, process type and specifications, a grain side quality assessment model is constructed to identify the grain side quality; Based on the results of the grain surface quality assessment and the grain side quality assessment, the grains are cut into pieces; The process of obtaining the texture features of the grain side profile image based on the grain side profile image includes: Based on the grain side profile image, the RGB channel values ​​of the image are converted into a single grayscale value using a weighted average method. If the electric slide carrying the optical camera moves and the illumination angle changes, a grayscale value compensation unit is used to compensate the grayscale value of each pixel in the grain side profile grayscale image according to the illumination angle. Based on the compensated grayscale values, a grayscale image of the grain side profile is obtained, and a grayscale co-occurrence matrix is ​​extracted. Based on the grayscale co-occurrence matrix of the grain side profile image, homogeneity is calculated based on a homogeneity calculation formula, and contrast is calculated based on a contrast calculation formula, and homogeneity and contrast are used as texture features of the grain side profile image. The method of compensating the grayscale value of each pixel in the grayscale image of the grain side profile according to the illumination angle includes: For each pixel in the grain side profile image, calculate the average grayscale value of the neighborhood around each pixel, and combine it with the illumination angle to use the illumination compensation formula , calculate the lighting compensation factor ,in, is the original grayscale value of the pixel (x,y) in the image, is the average grayscale value of the neighborhood around the pixel, which is used to measure the local light intensity. is a constant to prevent division by zero errors and is obtained by fitting historical data. Is a coefficient that controls the intensity of the impact of light changes and is obtained by fitting historical data. It is the angle between the normal of the pixel and the illumination direction, indicating the change in illumination angle and reflecting the change in surface illumination conditions. According to the illumination compensation factor, the formula is used. , perform illumination compensation on each pixel in the grain side profile image to obtain the compensated grayscale value .

2. The method according to claim 1, wherein The method includes obtaining an image of the grain to be inspected by a loading correction camera, converting edge points in the edge profile of the grain into geometric parameters using a Hough transform algorithm, determining the position of the grain to be inspected, and obtaining a position compensation value of the grain, including: The die is placed at the die inspection station by manual loading, and the optical camera at the die inspection station is used to obtain the die position information; the die position is corrected by the loading XYR platform, and the die is lifted up by the ejector module; The grain is placed on the loading transfer module through the first loading head and moved to the picking position of the second loading head; the loading correction camera at the picking position of the second loading head is used to obtain the image of the grain to be detected, the edge contour of the grain is obtained using the Canny edge detection algorithm, and the edge points in the edge contour of the grain are converted into geometric parameters through the Hough transform algorithm to determine the position of the grain to be detected; the initial front / back detection station image position of grains of the same specification at the correct imaging position is obtained through the historical detection records of the grain; according to the correct imaging position of the front / back detection station of the grain, the initial value of the correct imaging position of the side / end detection station is set, and the initial value of the correct imaging position of the front / back and side / end is recorded; by comparing the initial value of the correct imaging position of the front / back and side / end with the position of the currently detected grain, the position compensation value of the grain is obtained.

3. The method according to claim 1, wherein The surface contour image of the grain is extracted by the Canny edge detection algorithm, and based on the surface contour image of the grain, the shape factor, reflectivity gradient characteristics and texture characteristics of the grain surface contour image are obtained to construct the grain surface feature data, including: The second loading head picks up the grain and places it on the inspection table. The optical camera at the front / back inspection station of the grain is used to obtain the front / back image of the grain. The Canny edge detection algorithm is used to extract the surface contour image of the grain. Based on the surface contour image of the grain, the perimeter and area of ​​the grain surface contour image are determined, and the shape factor calculation formula is used. , calculate the shape factor of the grain surface profile image, where S is the shape factor of the grain, is the perimeter of the grain surface contour, The method comprises the following steps: obtaining the brightness or color distribution data of the grain surface profile image, calculating the reflectivity of each area by the change of the image brightness, obtaining the spatial distribution data of the surface reflectivity, and drawing a reflectivity distribution diagram; performing statistical analysis on the spatial distribution data of the surface reflectivity of the grain surface profile image, obtaining the reflectivity gradient characteristics of the grain surface, including the average reflectivity, the maximum reflectivity and the standard deviation of the gradient direction; converting the RGB channel values ​​of the image into a single grayscale value according to the grain surface profile image using the weighted average method, and obtaining the grayscale image of the grain surface profile; extracting the grayscale co-occurrence matrix according to the grayscale image of the grain surface profile, and calculating the homogeneity based on the formula , calculate the homogeneity H, and based on the contrast calculation formula , calculate the contrast C, and use homogeneity and contrast as texture features of the grain surface profile image, where i is the gray value of a pixel in the image, which is used as the reference pixel, and j is the gray value of the adjacent pixel paired with pixel i in a specific direction and distance, is the value of the i, j position in the gray-level co-occurrence matrix, indicating the frequency of occurrence of the gray-level value i and j pair; the shape factor, reflectivity gradient characteristics and texture characteristics of the grain surface profile image are standardized and combined into a feature vector to constitute the grain surface feature data, which is stored in the grain image monitoring database.

4. The method according to claim 1, wherein The method of constructing a grain surface quality assessment model based on grain surface feature data, material, process type and specifications to identify grain surface quality includes: Through the grain image monitoring database, historical grain surface feature data, materials, process types and specifications are obtained, the grain surface quality is marked, and the random forest algorithm is used for model training to build a grain surface quality assessment model. Process types include casting, forging, cutting, welding and heat treatment processes, materials include metal materials, ceramic materials and polycrystalline materials, and grain surface quality includes high quality, medium and low quality; based on the real-time acquired grain surface feature data, process type and specifications, the grain surface quality assessment model is used to identify the grain surface quality.

5. The method according to claim 1, wherein The side profile of the grain is extracted by the Canny edge detection algorithm, and based on the grain side profile image, the shape factor, texture characteristics, grain boundary width and grain boundary area of ​​the grain side profile image are obtained to construct the grain side feature data, including: If the die on the inspection table moves to the side / end inspection station, the electric slide carrying the optical camera is controlled to move to the position and focus the image according to the position compensation value of the die, and the side / end image of the die is obtained. The Canny edge detection algorithm is used to extract the side profile of the die, identify the die thickness, and use the shape factor calculation formula to calculate the shape factor of the die side profile image; If the grain material is a polycrystalline material, the grayscale image of the grain side profile is binarized, and the area with grayscale lower than the preset grayscale threshold is marked as the grain boundary area; the grain boundary is refined into a single pixel width, and the distance from the skeleton to the nearest non-grain boundary pixel in the image is calculated to obtain the local grain boundary width of each pixel, and the width values ​​of all grain boundary pixels are averaged to obtain the average grain boundary width; based on the average grain boundary width and the actual pixel size, the physical width is converted to obtain the grain boundary width; based on the number of pixels in the grain boundary area and the actual area of ​​each pixel, the physical area is converted to obtain the grain boundary area; the shape factor, texture characteristics, grain boundary width and grain boundary area of ​​the grain side profile image are standardized and combined into a feature vector to constitute the grain side feature data, which is stored in the grain image monitoring database.

6. The method according to claim 1, wherein The method of constructing a die side quality assessment model based on die side feature data, material, process type, and specifications to identify die side quality includes: Through the grain image monitoring database, historical grain side feature data, materials, process types and specifications are obtained, the grain side quality is marked, and the random forest algorithm is used for model training to build a grain side quality assessment model. Process types include casting, forging, cutting, welding and heat treatment processes, materials include metal materials, ceramic materials and polycrystalline materials, and grain side quality includes high quality, medium and low quality; based on the real-time acquired grain side feature data, process type and specifications, the grain side quality assessment model is used to identify the grain side quality.

7. The method according to claim 1, wherein The step of performing blanking processing on the grains according to the grain surface quality assessment results and the grain side quality assessment results comprises: Determine the grain cutting method based on the grain surface quality assessment results and the grain side quality assessment results; If the surface quality assessment result of the grain or the side quality of the grain is poor, the grain is removed from the inspection table by the NG blanking head and placed in order on the wide grain blanking ring; if the surface quality assessment result of the grain and the side quality of the grain are not poor, the grain is removed from the inspection table by the OK blanking head and placed on the OK blanking carrier; if the position compensation value of the grain is greater than the preset compensation value threshold, the grain is removed from the inspection table by the waste discharge head and placed on the re-inspection carrier for re-grain inspection.

8. A wafer six-side inspection system, which is implemented based on a wafer six-side inspection method according to any one of claims 1 to 7, characterized in that: The system includes the following modules: The grain loading and deflection correction module is used to obtain the grain image to be inspected through the loading and deflection correction camera, convert the edge points in the edge contour of the grain into geometric parameters using the Hough transform algorithm, determine the position of the grain to be inspected, and obtain the position compensation value of the grain; The surface feature data acquisition module is used to extract the surface contour image of the grain using the Canny edge detection algorithm, and based on the surface contour image of the grain, obtain the shape factor, reflectivity gradient characteristics and texture characteristics of the grain surface contour image to construct the grain surface feature data; The grain surface quality module is used to build a grain surface quality assessment model and identify the grain surface quality based on the grain surface profile image feature data, material, process type and specifications; The side feature data acquisition module is used to extract the side profile of the grain using the Canny edge detection algorithm, and based on the grain side profile image, obtain the shape factor, texture characteristics, grain boundary width and grain boundary area of ​​the grain side profile image to construct the grain side feature data; The die side quality module is used to build a die side quality assessment model and identify the die side quality based on the die side profile image feature data, material, process type and specifications; The grain blanking processing module is used to perform grain blanking processing on the grains according to the grain surface quality assessment results and the grain side quality assessment results.

Images