A method and system for measuring the dimensions of sheet metal based on image recognition

Abstract

The application relates to the technical field of image recognition, and particularly discloses a plate size measurement method and system based on image recognition, which comprises the following steps: collecting an image of a single local area, and obtaining the length of an aperture in the image; dividing the image into multiple segments according to surface feature information, obtaining the center position of a circle corresponding to an edge point set of a single segment; determining whether an abnormality exists in the corresponding local area based on the length of the aperture and the center position; obtaining a region value, generating a coordinate point of a single water-cooling plate based on the region value, clustering the coordinate point, obtaining a clustering cluster, and reworking the water-cooling plate according to the clustering cluster. The application realizes full-size online intelligent detection of a water-cooling plate and optimization of reworking of the same kind.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically to a method and system for measuring the dimensions of sheet metal based on image recognition. Background Technology

[0002] The automotive manufacturing industry, particularly the body assembly inspection process, currently faces multiple challenges. Traditional assembly inspection methods have many limitations and can no longer meet the development needs of the modern automotive industry, including but not limited to long tooling manufacturing cycles, high costs, lack of flexibility, and high dependence on personnel.

[0003] Against this backdrop, full-size measurement technology has emerged. As a new generation of measurement technology to replace traditional assembly gauges, it features non-contact measurement, digital virtual gauges, and full-size data acquisition. Furthermore, it integrates a powerful data processing and analysis system, enabling real-time monitoring and analysis of minute deviations in the manufacturing process and providing early warnings of quality trends.

[0004] When measuring water-cooled plates using full-size measurement technology, most defective workpieces are reworked uniformly without considering the differences between different defective workpieces. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for measuring the dimensions of sheet metal based on image recognition, thereby solving the aforementioned technical problems.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A method for measuring the dimensions of sheet metal based on image recognition includes the following steps:

[0008] The area to be detected on a single water-cooled plate is divided into several local regions. Images of a single local region are acquired, and the length of the aperture is obtained by a visual image processing algorithm based on a dual-strategy aperture measurement method.

[0009] The image is segmented into multiple segments based on surface feature information, including pixel-level grayscale distribution and texture description. The edge point set of a single segment is obtained, and the center position of the circle corresponding to the edge point set is obtained by a circle fitting method based on convex hull operation.

[0010] When the deviation between the center position and the preset standard position is greater than the preset deviation value and / or the difference between the length of the aperture and the preset standard length is greater than the preset difference threshold, it is determined that the current segment is abnormal and the corresponding local area is abnormal.

[0011] Based on whether there are anomalies in a local area, the regional value of the local area is obtained. Based on the regional value, the coordinate point of a single water-cooled plate is generated. The coordinate points within the preset monitoring period are clustered to obtain clusters. At the end of the monitoring period, the same rework process is performed on the water-cooled plates corresponding to the coordinate points in the same cluster.

[0012] As a further aspect of the present invention: segmenting an image into multiple segments based on surface feature information includes:

[0013] Step 1: Divide the image into grids and generate feature vectors (A1, A2) for each grid based on surface feature information. A1 and A2 represent gray-level gradient statistics and texture pattern frequency, respectively.

[0014] Step 2: Obtain the similarity between any two grids based on feature vectors, construct a similarity matrix, and use a graph-based clustering algorithm to group grids in the similarity matrix with a similarity greater than a preset similarity threshold into a cluster;

[0015] Obtain the Euclidean distance between the center vector of the cluster and any eigenvector in the cluster;

[0016] Step 3: When the total Euclidean distance is less than or equal to the preset total Euclidean distance threshold, the current cluster is treated as a segment.

[0017] As a further aspect of the present invention: segmenting an image into multiple segments based on surface feature information further includes:

[0018] When the total Euclidean distance is greater than the preset total Euclidean distance threshold, the current cluster is taken as the split cluster, the feature vectors in the split cluster are combined into a feature vector set, the direction a with the largest Euclidean distance is obtained, and the binary feature partitioning is performed along the direction a, dividing the feature vector set into two subsets.

[0019] Treat the subset as a cluster and determine whether it is a fragment.

[0020] As a further aspect of the present invention: the process of obtaining the set of edge points of a single segment includes:

[0021] Edge extraction methods based on sliding window edge point outlier correction and region difference edge point outlier correction are used to extract the edge point set of a single segment, and the edge point set with lower confidence is discarded.

[0022] As a further aspect of the present invention: generating the coordinate points of a single water-cooled plate includes:

[0023] The local regions are numbered. When region i has an anomaly, the corresponding region value Xi = 1; when region i does not have an anomaly, the corresponding region value Xi = 0.

[0024] Generate coordinate points (X1, X2, ..., Xm), where m represents the total number of points in the local region.

[0025] As a further aspect of this invention, the clustering method is k-means or DBSCAN.

[0026] As a further aspect of the present invention: obtaining the length of the aperture includes:

[0027] The image is segmented by thresholding to obtain a binary image. The breaks in the hole region are filled by dilation operation, and small noise is removed by erosion operation to obtain a coherent binary representation of the hole region.

[0028] The edge of the hole region is extracted based on the binary representation of the hole region, and the edge curve is smoothed to obtain the minimum bounding rectangle of the edge curve. The dimensions of the long and short sides of the minimum bounding rectangle are converted into actual physical dimensions according to a preset scaling factor.

[0029] The direction of the long side of the smallest bounding rectangle is defined as the normal vector of the hole region. A vertical dividing line is constructed based on the centroid coordinates and normal vector direction of the hole region. The vertical dividing line divides the hole region into two non-overlapping sub-regions.

[0030] Perform connected component analysis on the difference region between the binary image and the vertical segmentation line to obtain the area of ​​the connected region. Sort the areas in descending order and obtain the connected regions corresponding to the top two areas in the sort.

[0031] Perform least-squares circle fitting on the edge contour of the connected region to obtain two target circles. Obtain the circle parameters of the target circles, including the center coordinates and radius.

[0032] The major axis dimension of the waist-shaped hole profile is reconstructed based on the circle parameters and the predetermined geometric reconstruction relationship, which is the length of the hole diameter.

[0033] A sheet metal dimension measurement system based on image recognition, characterized in that it includes:

[0034] Aperture acquisition module: Divide the area to be detected on a single water-cooled plate into several local regions, acquire images of a single local region, and obtain the length of the aperture using a visual image processing algorithm based on a dual-strategy aperture measurement method;

[0035] Center acquisition module: The image is segmented into multiple segments based on surface feature information, including pixel-level grayscale distribution and texture description. The edge point set of a single segment is obtained, and the center position of the circle corresponding to the edge point set is obtained by a circle fitting method based on convex hull operation.

[0036] Judgment module: When the deviation between the center position and the preset standard position is greater than the preset deviation value and / or the difference between the length of the aperture and the preset standard length is greater than the preset difference threshold, the current segment is judged to be abnormal, and the corresponding local area is judged to be abnormal.

[0037] Rework Module: Based on whether there are anomalies in a local area, obtain the regional value of the local area, generate the coordinate point of a single water-cooled plate based on the regional value, cluster the coordinate points within the preset monitoring period to obtain clusters, and perform the same rework process on the water-cooled plates corresponding to the coordinate points in the same cluster at the end of the monitoring period.

[0038] The beneficial effects of this invention compared to the prior art are as follows:

[0039] This invention achieves rapid identification of local anomalies by adaptively segmenting water-cooled plate images according to surface features, extracting edges through multi-strategy fusion, and combining convex hull circle fitting for precise hole location. It utilizes dual threshold judgments of hole center deviation and hole diameter error. High-dimensional coordinate points are then generated using region numbers, and clustering algorithms are used to classify workpieces within the monitoring period in a patterned manner. This allows for automatic matching of the same rework strategy to workpieces with the same failure mechanism. This process requires no physical fixtures, does not rely on human experience, and possesses high adaptability and real-time performance. It improves the accuracy and stability of dimensional inspection while avoiding excessive rework of normal areas, thus significantly reducing production downtime and secondary processing costs, and comprehensively optimizing the quality control chain of the vehicle body assembly. Furthermore, the algorithm parameters can be updated online through self-learning, ensuring consistent inspection over long production cycles. Attached Figure Description

[0040] The invention will now be further described with reference to the accompanying drawings.

[0041] Figure 1 This is a flowchart illustrating a method for measuring the dimensions of sheet metal based on image recognition according to the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Please see Figure 1 As shown, this invention is a method for measuring the dimensions of sheet metal based on image recognition, comprising the following steps:

[0044] Based on the process drawings of the water-cooled plate, the entire plate is divided into several non-overlapping small blocks according to a regular grid on the offline calibration interface. The boundary of each small block is aligned with the geometric center of the theoretical hole position, so that a local area usually covers only one hole to be tested.

[0045] In the online inspection process, the automated handling device delivers the water-cooled plate to the fixed station, and the high-resolution area camera on the top sequentially completes the top view acquisition of each local area. The system simultaneously turns on the ring shadowless light source when the shutter is triggered to ensure that the grayscale distribution of each area is consistent.

[0046] After obtaining the original image, the length of the aperture is determined from the image.

[0047] By employing gridded local partitioning and dual-strategy measurement, complex large-format images can be decomposed into small tiles of consistent scale, reducing the interference of background texture on threshold segmentation and lowering the computational load of the algorithm in video memory.

[0048] In a preferred embodiment of the present invention, obtaining the length of the aperture includes:

[0049] The grayscale image of the original image is input into an adaptive thresholding algorithm, such as using Otsu's iteration or local mean thresholding, to automatically divide the grayscale distribution into foreground and background and output a binary image. A morphological dilation operator is then called, typically using 3×3 or 5×5 cross-shaped structural elements, to quickly fill in the discontinuous pixels in the hole outline caused by light spots or scratches. An erosion operator of the same size is used to perform an erosion operation on the entire binary image to eliminate isolated noise points that have been amplified by dilation, ensuring that the remaining connected regions contain only complete hole shapes. The Canny-Sobel combined operator is used to extract subpixel-level edges in the connected regions, and the edge coordinate sequence is smoothed by a five-point median filter. Then, the minimum bounding rectangle is searched based on the rotating caliper method, and the lengths of the long and short sides of the rectangle in pixel coordinates are read.

[0050] Then, using the pixel-physical scaling factor pre-determined by the external calibration board, the pixel length is directly converted into a millimeter-level physical dimension. Using the direction vector of the long side of the rectangle as the normal vector of the hole region, a vertical dividing line is established at the centroid of the hole, and the difference set between the line and the binary image is calculated using the linear equation. The remaining area of ​​the difference set is then analyzed for connected components and automatically sorted according to area. The top two connected components are selected as candidate sub-regions at both ends of the waist-shaped hole. Least square circle fitting is called for the edge of each sub-region, and the center and radius of the two circles are iteratively obtained. The theoretical outline of the complete waist-shaped hole is reconstructed according to the geometric relationship between the direction of the line connecting the center and the direction of the long axis of the hole. Finally, the long axis dimension of the hole is analytically calculated from the relationship between the distance between the centers of the two circles and the addition and subtraction of the radius. This length is bound to the workpiece's unique barcode and written into the database. If the fitting residual is abnormal, the backup edge integration algorithm is automatically triggered to re-estimate and output.

[0051] It is worth noting that the process of threshold segmentation followed by morphological repair can quickly obtain the closed hole shape when the lighting is uneven or the plate surface is reflective, avoiding model oscillation caused by fitting directly on a noisy background. Quantizing pixels to physical size using the minimum bounding rectangle utilizes the global scaling factor of the camera-station calibration, making different batches of workpieces comparable at the same scale. Using vertical dividing lines and connected component filtering to decompose complex waist-shaped holes into two approximate circles simplifies the degrees of freedom of subsequent fitting, allowing circle fitting to converge quickly using the least squares closed-form solution, reducing computational load, and creating conditions for the same algorithm framework to be compatible with both circular and waist-shaped holes. Finally, the geometric reconstruction process of back-deriving the major axis dimension from the parameters of the two circles can directly output the inspection quantity corresponding to traditional gauges, thus ensuring that the algorithm results are consistent with the workshop process documents.

[0052] The image is segmented into multiple segments based on surface feature information. The grayscale sub-image corresponding to each segment is read, and the initial edges are extracted using a combination of edge gradient operator and threshold. To avoid rough jagged edges interfering with subsequent geometric fitting, a fixed-length window is slid along the edge chain to perform median smoothing on the coordinate sequence. All smoothed pixel coordinates are transformed to the workstation calibration coordinate system and stored in the point cloud queue. Then, a polar angle sorting scanning method is used to perform convex hull operation on the point cloud to obtain a set of minimal closure points arranged in counterclockwise order. This set of points is used as candidate fitting points. When the number of convex hull points is less than the set threshold, the adjacent original edge points with the largest curvature are automatically added to maintain the integrity of the curve. Subsequently, the least squares circle fitting module is called to iteratively solve for the center position and radius of the circle using the convex hull point set as input. The initial value of the circle center is approximately given by the centroid of the edge points to accelerate convergence. After fitting, the absolute position of the circle center in the workstation coordinate system is output and written into the segment attribute table for subsequent error determination.

[0053] By performing convex hull and then circle fitting, key contour features can be preserved while removing outliers, avoiding noise or burrs from deviating from the overall circle center, thus improving the reliability of circle center positioning. The convex hull ensures that the line connecting any two points in the selected point set is inside the curve, making the circle fitting more consistent with the actual hole wall curvature distribution. The least squares method balances all errors in an overall sense, making the results insensitive to single-point anomalies and providing a stable basis for judging the hole center offset.

[0054] In one specific embodiment, segmenting an image into multiple segments based on surface feature information includes:

[0055] According to the physical dimensions given in the process drawings, the entire image is uniformly divided into several grid blocks with consistent rows and columns. For example, a 2,000-pixel square image is cut into 40x40 small grids, each grid roughly corresponding to the projection range of a rivet or weld spot on the board surface. The pixel grayscale gradient histogram is extracted for each grid as A1, and the frequency of occurrence of local binary texture patterns is extracted as A2. The two sets of statistics are concatenated end-to-end to form a low-dimensional dense feature vector. After traversing all grids, the cosine similarity is used to measure the similarity between any two grid feature vectors, and the similarity is written into the weight edges of a symmetric matrix to form a graph structure. Then, a graph-based clustering algorithm is called; a common practice is to perform feature division on the weight matrix. Solve the problem and perform connected component search in the low-dimensional spectral space using K-means or by directly using edge weight thresholding. Grids with similarity higher than a preset threshold are grouped into the same cluster. For each cluster, calculate the arithmetic mean of the feature vectors to obtain the center vector, and calculate the sum of the Euclidean distances between the center and all vectors in the cluster. If this total distance does not exceed an empirical threshold, the cluster is directly regarded as a segment. If the total distance is too large, the cluster is regarded as a split cluster. The feature vector set is projected onto the principal component direction or the maximum distance direction 'a' and bisected according to the median value of the projection. The vector set is hard-cut into two subsets, and new clusters are constructed for each subset. The calculation of the center and total distance is repeated until the internal distances of all clusters meet the segment determination conditions.

[0056] Introducing a regular grid in the early stages of image processing can transform large-format scenes into a controllable number of small blocks, making the subsequent computational load linearly related to the region content. This avoids missegmentation of local textures by global thresholds and facilitates subsequent connectivity clustering based on spatial adjacency. The feature vector formed by combining gray-level gradients and texture frequencies can simultaneously capture brightness changes and surface textures, enabling the identification of regions with the same color but different textures or the same texture but different colors. Using a similarity matrix to construct the graph and performing clustering in the graph space fully utilizes the similarity strength between grids expressed by edge weights, ensuring that the segmentation results conform to the true boundaries without excessive fragmentation. Setting a total Euclidean distance threshold to check the consistency within clusters can prevent excessively large clusters from masking local differences. Recursively splitting along the direction of maximum discreteness can also separate heterogeneous sub-regions in a timely manner, ensuring that the final fragment presents a local image with a single texture and a coherent gray-level distribution. This provides a clean and structurally clear input foundation for subsequent edge point extraction, circle center fitting, and even anomaly detection, thereby enhancing the system's adaptability to complex board surface defects and its positioning accuracy as a whole.

[0057] When the deviation between the center position and the preset standard position is greater than the preset deviation value and / or the difference between the length of the aperture and the preset standard length is greater than the preset difference threshold, it is determined that the current segment is abnormal and the corresponding local area is abnormal.

[0058] In another preferred embodiment of the present invention, generating the coordinate points of a single water-cooled plate includes:

[0059] The local regions are numbered. When region i has an anomaly, the corresponding region value Xi = 1; when region i does not have an anomaly, the corresponding region value Xi = 0.

[0060] Generate coordinate points (X1, X2, ..., Xm), where m represents the total number of points in the local region;

[0061] After inspecting a water-cooled plate, the abnormal Boolean results of each region are written into a one-dimensional array of length m according to the pre-agreed local area numbering order. The array element takes the value "1" to represent that there is a hole center offset or hole diameter abnormality in the region, and "0" to represent that the inspection is passed. Then, this string of values ​​is directly used as the feature coordinate points of the water-cooled plate and saved to the production database.

[0062] Using shifts as the monitoring cycle, all coordinate points generated within that cycle are queried periodically. The DBSCAN unsupervised clustering method based on Hamming distance is called and a minimum cluster capacity threshold is set. Workpieces with a Hamming distance less than or equal to the pre-set radius are grouped into the same cluster (other clustering methods, such as K-means, can be used, and no restrictions are placed here).

[0063] After clustering, a unique Closer-ID is generated for each cluster, and the most frequently occurring abnormal region combinations are retrieved within the cluster and mapped to the corresponding process routes in the rework knowledge base, such as "three-axis milling on the upper plate table - grinding - re-inspection" or "automatic welding - deburring - re-inspection". When the shift ends, the system pushes a rework list to the manufacturing execution system. Each Closer-ID in the list carries a set of water-cooled plate barcodes and a unified rework path. On-site operators or robots perform the same rework on the workpieces in the same cluster according to the instructions. After the rework is completed, the barcode is scanned again to send back the status and the work order is closed.

[0064] Using region Boolean vectors to generate coordinate points can transform dimensional inspection results into discrete features with fixed lengths and uniform formats, facilitating efficient indexing and computation in the database. Clustering based on Hamming distance can naturally measure the similarity of two boards in abnormal distribution, causing problems caused by defects from the same source or wear of the same tool to tend to cluster together. Binding the clustering results with the preset rework process not only avoids operators from judging the defect location and deciding the repair method for each piece, but also ensures that workpieces with the same failure mechanism are treated consistently, thereby reducing secondary errors caused by chaotic rework plans. It also allows for quick location of the corresponding defect pattern and repair record based on Closer-ID during subsequent quality traceability, providing organized data support for continuous improvement of the production line.

[0065] A sheet metal dimension measurement system based on image recognition, comprising:

[0066] Aperture acquisition module: Divide the area to be detected on a single water-cooled plate into several local regions, acquire images of a single local region, and obtain the length of the aperture using a visual image processing algorithm based on a dual-strategy aperture measurement method;

[0067] Center acquisition module: The image is segmented into multiple segments based on surface feature information, including pixel-level grayscale distribution and texture description. The edge point set of a single segment is obtained, and the center position of the circle corresponding to the edge point set is obtained by a circle fitting method based on convex hull operation.

[0068] Judgment module: When the deviation between the center position and the preset standard position is greater than the preset deviation value and / or the difference between the length of the aperture and the preset standard length is greater than the preset difference threshold, the current segment is judged to be abnormal, and the corresponding local area is judged to be abnormal.

[0069] Rework Module: Based on whether there are anomalies in a local area, obtain the regional value of the local area, generate the coordinate point of a single water-cooled plate based on the regional value, cluster the coordinate points within the preset monitoring period to obtain clusters, and perform the same rework process on the water-cooled plates corresponding to the coordinate points in the same cluster at the end of the monitoring period.

[0070] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. An image recognition-based plate size measurement method, characterized by, The method comprises the following steps: The area to be detected on the single water cooling plate is divided into a plurality of local areas, an image of a single local area is collected, and a visual image processing algorithm based on a double-strategy aperture measurement method is used to obtain the length of the aperture; The image is segmented into a plurality of segments based on surface feature information, the surface feature information including pixel-level gray scale distribution and texture description, an edge point set of a single segment is obtained, and a convex hull operation-based circular fitting method is used to obtain the center position of a circle corresponding to the edge point set; When the deviation between the center position and a preset standard position is greater than a preset deviation value and the difference between the length of the aperture and a preset standard length is greater than a preset difference threshold value, it is determined that the current segment is abnormal and that the corresponding local area is abnormal; A region value of the local area is obtained based on whether the local area is abnormal, coordinate points of the single water cooling plate are generated based on the region value, the coordinate points in a preset monitoring period are clustered to obtain a cluster, and the same rework process is performed on the water cooling plates corresponding to the coordinate points in the same cluster at the end of the monitoring period; The length of the aperture comprises: The image is threshold segmented to obtain a binary image, the aperture region is filled by an expansion operation, and small noise points are removed by an erosion operation to obtain a connected aperture region binary representation; The aperture region edge is extracted based on the aperture region binary representation, and the edge curve is smoothed to obtain a minimum circumscribed rectangle of the edge curve, and the size of the long side and the short side of the minimum circumscribed rectangle is converted into actual physical size according to a preset scale factor; The long side direction of the minimum circumscribed rectangle is defined as the normal vector of the aperture region, and a vertical segmentation line is constructed based on the centroid coordinates of the aperture region and the direction of the normal vector, and the vertical segmentation line divides the aperture region into two non-overlapping sub-regions; Connected region analysis is performed on the difference set region of the binary image and the vertical segmentation line to obtain the area of the connected domain, and the areas are sorted in descending order to obtain the connected domains corresponding to the top two areas in the sorting; Least square circle fitting is performed on the edge profile of the connected domain to obtain two target circles, and the circle parameters of the target circles are obtained, including the center coordinates and the radius; The long axis size of the waist-shaped aperture profile is reconstructed based on the circle parameters and a predetermined geometric reconstruction relationship, and the length of the aperture is obtained.

2. The image recognition-based plate size measurement method according to claim 1, characterized in that, Segmenting the image into a plurality of segments based on the surface feature information comprises: Step 1: The image is grid divided, and a feature vector (A1, A2) of a single grid is generated based on the surface feature information, A1 and A2 representing gray scale gradient statistics and texture mode frequency, respectively; Step 2: The similarity between any two grids is obtained based on the feature vector, a similarity matrix is constructed, and a graph-based clustering algorithm is used to divide the grids with a similarity greater than a preset similarity threshold value in the similarity matrix into a cluster; The Euclidean distance between the center vector of the cluster and any one feature vector in the cluster is obtained; Step 3: When the total Euclidean distance is less than or equal to a preset total Euclidean distance threshold value, the current cluster is taken as a segment.

3. The image recognition-based plate size measurement method according to claim 1, wherein Segmenting the image into a plurality of segments based on the surface feature information further comprises: When the total Euclidean distance is greater than a preset total Euclidean distance threshold, the current cluster is taken as a split cluster, a feature vector set is composed of feature vectors in the split cluster, a direction a with the maximum Euclidean distance is obtained, a binary feature division is performed along the direction a, and the feature vector set is divided into two subsets; The subset is taken as a cluster, and whether it is a segment is judged.

4. The image recognition-based plate size measurement method according to claim 1, wherein The process of obtaining the edge point set of a single segment includes: The edge point set of a single segment is extracted based on the edge extraction method of edge point abnormal value correction of a sliding window and the edge extraction method of edge point abnormal value correction of region difference, and the edge point set with a smaller confidence is discarded.

5. The image recognition-based plate size measurement method according to claim 1, wherein Generating the coordinate point of a single water cooling plate includes: The local area is numbered, when the area numbered i is abnormal, the corresponding area value Xi=1; when the area numbered i is not abnormal, the corresponding area value Xi=0; The coordinate point (X1, X2, …, Xm) is generated, and m represents the total number of local areas.

6. The image recognition-based plate size measurement method according to claim 1, wherein The clustering method is k-means or DBSCAN.

7. An image recognition-based board size measurement system, characterized by, It includes: The aperture acquisition module: the detection area on the single water cooling plate is divided into a plurality of local areas, the image of a single local area is collected, and the length of the aperture is obtained based on the visual image processing algorithm of the aperture measurement method of the double strategy; The center acquisition module: the image is segmented into a plurality of segments according to the surface feature information, the surface feature information includes pixel-level gray distribution and texture description, the edge point set of a single segment is obtained, and the center position of the circle corresponding to the edge point set is obtained based on the circle fitting method of convex hull operation; The judgment module: when the deviation between the center position and the preset standard position is greater than the preset deviation value, and the difference between the length of the aperture and the preset standard length is greater than the preset difference threshold, it is determined that the current segment is abnormal, and it is determined that the corresponding local area is abnormal; The rework module: the area value of the local area is obtained based on whether the local area is abnormal, the coordinate point of a single water cooling plate is generated based on the area value, the coordinate points in a preset monitoring period are clustered to obtain a cluster, and the same rework process is performed on the water cooling plates corresponding to the coordinate points in the same cluster at the end of the monitoring period; The length of the aperture includes: The image is threshold segmented to obtain a binary image, the broken holes in the hole region are filled by inflation operation, and then the small noise points are removed by corrosion operation to obtain a connected hole region binary representation; Based on the hole region binary representation, the hole region edge is extracted, and the edge curve is smoothed to obtain the minimum circumscribed rectangle of the edge curve, and the size of the long side and the short side of the minimum circumscribed rectangle is converted into the actual physical size according to the preset scale factor; The long side direction of the minimum circumscribed rectangle is defined as the normal vector of the hole region, and the vertical division line is constructed based on the centroid coordinates of the hole region and the normal vector direction, which divides the hole region into two non-overlapping sub-regions; The difference set region of the binary image and the vertical division line is executed to perform connected region analysis, the area of the connected domain is obtained, and the areas are sorted in descending order to obtain the connected domains corresponding to the first two areas in the sorting; Performing least square circle fitting on the edge contour of the connected domain to obtain two target circles, and obtaining circle parameters of the target circles, the circle parameters including a center coordinate and a radius; Reconstructing a long axis size of the waist hole contour, i.e., a length of the hole diameter, based on the circle parameters and a predetermined geometric reconstruction relationship.

Images