A lithium battery pole piece size detection method based on machine vision

Abstract

This invention discloses a machine vision-based method for detecting the size of lithium battery electrode sheets. The method includes acquiring an image of the lithium battery electrode sheet and roughly locating the edges; extracting edge distribution features containing edge gradient directions and position coordinates, matching them with a pre-stored template to determine the specifications and model, and setting the search area using detection parameters; acquiring candidate edge points within the search area, and employing a dual strategy to filter and eliminate false edge points: when the maximum cluster proportion reaches a preset threshold, neighboring points are filtered using the maximum cluster anchor point; if the threshold is not reached, the optimal point is selected based on gradient filtering conditions and a preset side direction; and calculating the final edge segment and size based on the true edge points using sub-pixel precision calculation and iterative fitting. This invention achieves adaptive registration for specifications and highly robust extraction of complex edges, improving detection accuracy and automation.

Description

Technical Field

[0001] This invention belongs to the field of image processing, specifically relating to a method for detecting the size of lithium battery electrode sheets based on machine vision. Background Technology

[0002] Currently, in the rolling and coating production of strip-shaped workpieces such as lithium battery electrodes, real-time tracking and detection of coating dimensions are crucial for adjusting processes and improving product yield. Existing dimensional detection solutions typically rely on manual sampling or sensor-based automated measurement devices. Manual sampling is inefficient, inaccurate, and lacks real-time performance, failing to provide timely feedback on production deviations for process adjustments. Existing automated inspection equipment often struggles to accommodate multiple product specifications, requiring cumbersome manual parameter modifications and switching operations when switching between different specifications. Furthermore, strip-shaped workpieces are prone to irregularities such as stacking, adhesion, interlocking, or breakage at the edges during production, easily introducing false edge points during image edge extraction. Existing visual inspection methods lack robustness when dealing with complex edges, making it difficult to achieve high-precision edge positioning and dimensional measurement. Therefore, a high-precision visual inspection solution that can adapt to multi-specification product switching and has high robustness in filtering out complex false edges is urgently needed. Summary of the Invention

[0003] To address the shortcomings of traditional manual sampling or existing sensor solutions, this application discloses a machine vision-based method for detecting the size of lithium battery electrode sheets, which automates and enables real-time detection of the process, and is specifically optimized for special defects and diverse specifications in lithium battery electrode sheet production.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] A machine vision-based method for detecting the size of lithium battery electrode sheets includes the following steps:

[0006] Step 1: Acquire images of lithium battery electrode sheets during the rolling process using a high-resolution camera, and locate the approximate edge positions of the aluminum foil, copper foil, and ceramic edge in the electrode sheet images;

[0007] Step 2: Based on the rough edge position, determine the model of the current electrode by matching the edge point registration algorithm;

[0008] Step 3: Based on the matched electrode model, call the corresponding detection parameters and set a custom caliper tool search area near the rough edge position;

[0009] Step 4: Within the search area of ​​the caliper tool, the edge points are initially classified using the clustering method, and false edge points are eliminated using the edge filtering method to obtain a set of candidate edge points.

[0010] Step 5: Perform sub-pixel precision calculation on the candidate edge points to improve the edge point positioning accuracy and obtain a sub-pixel edge point set;

[0011] Step 6: Based on the sub-pixel edge point set, edge fitting is performed by a two-step iterative fitting method: the first fitting obtains an initial line segment model, defective points are eliminated based on the deviation between the points and the initial line segment model, and the remaining points are used for a second fitting to obtain a high-precision final edge line segment.

[0012] Step 7: Calculate the dimensions of each component of the lithium battery electrode based on the final edge segment.

[0013] Furthermore, in step 2, the edge point registration algorithm specifically includes:

[0014] Vertical projection is performed on each column of pixels in the ROI image to obtain a one-dimensional projection array;

[0015] The sign of the element value in the filtered array is used as the first row element of the two-dimensional feature vector, and the corresponding array index is used as the second row element to generate a two-dimensional feature vector describing the edge distribution.

[0016] The two-dimensional feature vector of the current electrode is matched with the two-dimensional feature vectors of various electrode models pre-stored in the database, and the model with the highest matching score is taken as the matching result.

[0017] Furthermore, the length of the one-dimensional projection array is the same as the width of the ROI image, and the array element values... Represented as: ,in It is the image height. The coordinates in the image are The grayscale value of the pixel at the location;

[0018] The matching score s is represented as:

[0019]

[0020] In the formula, and These represent the values ​​in the first row and i-th column of the feature vector of the current electrode and a certain type of electrode in the database, respectively, representing the edge polarity; and represents the value of the second row and i-th column of the feature vectors of the two, respectively, representing the horizontal coordinates of the edge.

[0021] Furthermore, in step 3, the clustering method is a horizontal coordinate-based clustering method, which includes setting the cluster neighborhood radius. Clustering of candidate edge points yields clusters;

[0022] When the proportion of the point set of the largest cluster reaches a preset threshold, the median horizontal coordinate of the point set within that cluster is taken. As edge anchor points; iterate through all calipers in the caliper group, and select the candidate edge points of each caliper that satisfy the condition that the horizontal coordinates are within the specified range. The point with the smallest distance within the range is selected as the edge point after filtering;

[0023] Otherwise, for each set of candidate edge points of the caliper, determine its maximum gradient value. ,by Multiply by the ratio threshold to use as the filtering threshold, and remove edge points that do not meet the conditions; then, according to the leftmost or rightmost mode in the caliper group detection parameters, select the edge point with the smallest or largest horizontal coordinate among the filtered candidate points as the best edge point of the caliper.

[0024] Furthermore, in step 5, sub-pixel calibration is performed using Gaussian approximation, specifically: for the horizontal coordinates of pixel-level edge points... Get its value in the one-dimensional projection array. and its adjacent points , ;use , , Gaussian function fitting is performed on the three points to calculate the sub-pixel offset. :

[0025] ,

[0026] pixel-level edge point coordinates With offset Add them together to obtain the edge point positions with sub-pixel precision.

[0027] Furthermore, in step 6, the first fitting uses the least squares method to fit the linear equation; the method for eliminating bad points is as follows: calculate the distance between all edge points and the line obtained from the first fitting, sort them from near to far, discard the point set in the bottom A% of the sorted points, and use the point set in the top B% of the sorted points for the second fitting.

[0028] Furthermore, in step 7, based on the matched electrode model, the testing requirements for that model are obtained from the knowledge base. Based on the final edge line segment, the distance and position deviation between specific line segments are statistically analyzed to calculate the size of each component and determine whether it is qualified.

[0029] Beneficial effects:

[0030] (1) This invention performs specification registration by extracting edge distribution features including edge gradient direction and edge position coordinates, and automatically identifies product model and calls corresponding detection parameters by using the signature features of edge morphology. This avoids the cumbersome manual switching and parameter modification of traditional detection equipment when facing products with multiple specifications, and improves the automation level and production efficiency of the detection process.

[0031] (2) The invention innovatively introduces a dual-strategy screening mechanism based on the maximum cluster ratio. When the edge regularity and cluster ratio meet the criteria, neighboring points are screened through cluster anchor points, and discrete pseudo-edges are filtered out by utilizing spatial clustering. When the edge is broken and irregular, or the cluster ratio does not meet the criteria, the system automatically switches to a screening strategy based on gradient threshold and preset side direction, using gradient strength and geometric boundary constraints to locate the best edge point. This mechanism forms a logical closed loop for different edge morphologies, effectively overcoming the problem of extraction failure under pseudo-edge interference in traditional methods, and significantly improving the robustness of edge localization.

[0032] (3) Based on eliminating false edges, this invention further breaks through the physical pixel resolution limitation by calculating the sub-pixel precision, and eliminates defective points with large residuals by fitting through a second iteration, thereby eliminating the bias effect of occasional noise on the line segment model, thus realizing extremely high precision edge line segment reconstruction and size calculation, meeting the strict detection requirements of lithium battery electrode for small size deviations. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the region of interest in an electrode image according to an embodiment of the present invention;

[0034] Figure 2 This is a schematic diagram of the caliper assembly according to an embodiment of the present invention;

[0035] Figure 3 This is a schematic diagram of edge fitting according to an embodiment of the present invention;

[0036] Figure 4 This is a schematic diagram of line segment measurement according to an embodiment of the present invention;

[0037] Figure 5 This is a flowchart of the detection method according to an embodiment of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the present invention clearer, 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.

[0039] like Figure 5As shown in the figure, this embodiment provides a method for detecting the size of lithium battery electrode sheets based on machine vision. The method includes the following steps:

[0040] Step 1: Acquire images of lithium battery electrodes during the rolling process using a high-resolution camera, and locate the approximate edge positions of the aluminum foil, copper foil, and ceramic edge in the electrode images.

[0041] This embodiment is applicable to strip-shaped workpieces, that is, workpieces with a continuous extended shape and a multi-layered or double-sided edge structure in cross-section. It is not limited to lithium battery electrode sheets, but can also be metal foil strips, coated films, composite tapes, etc. After acquiring images of the strip-shaped workpiece during the rolling or coating process using a high-resolution camera, preliminary image processing algorithms (such as threshold segmentation or coarse edge detection operators) are used to locate the approximate positions of edges such as aluminum foil, copper foil, ceramic edges, or coating boundaries in the images.

[0042] Steps 2-3: Based on the rough edge position, extract the edge distribution features including the edge gradient direction and edge position coordinates. Match the edge distribution features with the pre-stored template to determine the current electrode specification. Based on the specification, call the corresponding detection parameters to set the caliper tool search area near the rough edge position.

[0043] Here, the edge gradient direction refers to the direction of the brightness change at the edge (e.g., from bright to dark or from dark to bright), and the edge position coordinates refer to the horizontal or vertical coordinate position of the edge in the image. These two factors combined constitute the edge distribution features. In the electrode model database, different specifications of electrodes (such as electrodes with different widths and coating area distributions) correspond to a pre-stored template describing the edge distribution. By calculating the matching degree between the edge distribution features of the current image and each pre-stored template, the specification corresponding to the pre-stored template with the highest matching degree is determined as the specification of the current strip workpiece. Once the specification is determined, the system automatically calls the detection parameters bound to that model and sets a search area near the approximate edge position. This search area is typically represented in the visual inspection logic as a series of rectangular regions of interest arranged along the edge (i.e., the caliper tool commonly used in the industry). Specifically,

[0044] like Figure 1 As shown, the black rectangular area represents the main image of the lithium battery electrode (the edge area of ​​aluminum foil, copper foil, or ceramic), and the red rectangular box marks the region of interest (ROI). The ROI is located in the edge transition area of ​​the electrode (such as the boundary between aluminum foil and ceramic edge), covering the width direction of the electrode laterally, and being cropped vertically along the direction of electrode movement (rolling direction).

[0045] A Region of Interest (ROI) is set within the red box area of ​​the image, and this ROI image is cropped for edge detection. The cropped image is smaller, and interference from redundant areas is reduced.

[0046] Based on this, the pixels of each column of the ROI image are vertically projected, which is defined as follows:

[0047] (1)

[0048] In the formula These are the grayscale projection values ​​of column x. It is the image height. The coordinates in the image are The pixel grayscale value at a given location. After vertical projection, a one-dimensional array is generated from the two-dimensional ROI image, with the length of the array matching the width of the ROI image. Thresholding and non-maximum suppression methods are used to filter points in the output array whose absolute value is greater than a threshold, and then the maximum points within each region are selected. The sign of the elements in the filtered output array is used as the first row of the two-dimensional feature vector, and the index is used as the second row. This results in a two-dimensional feature vector describing the edge distribution, where the first row represents the polarity (direction of brightness change) of the edge, with 1 representing bright to dark and -1 representing dark to bright. The second row represents the horizontal coordinates of the edge in the image.

[0049] In the electrode model database, different electrode models correspond to a two-dimensional feature vector describing the edge distribution, which is used for electrode model registration to obtain the relevant detection parameters for that product. This paper proposes a registration calculation method:

[0050] (2)

[0051] In the formula, s represents the matching score. This represents the value (polarity) of the first row and i-th column of the two-dimensional feature vector of the current electrode edge. This represents the value (polarity) of the first row and i-th column of the two-dimensional feature vector of an electrode of a certain specification. When the polarities are the same, multiplying the two results in a positive number, indicating that the polarity directions of the two edges are consistent. This represents the value (horizontal coordinate) of the second row and i-th column of the current two-dimensional feature vector of the electrode. This represents the value (horizontal coordinate) in the second row and i-th column of the two-dimensional feature vector of a certain type of electrode. The product of the two coordinates is then subtracted from the square of their average. The greater the distance between the two edge coordinates, the larger the value; the reciprocal of the value indicates a smaller value. This reflects that the greater the distance between the horizontal coordinates of two edges, the weaker their correlation.

[0052] By using formula (2) to register the two-dimensional feature vector of the current electrode model with the two-dimensional feature vector of the electrode in the database, the matching degree between the current electrode model and the electrode models in the database can be obtained. After calculating the two-dimensional feature vector of the current electrode with the two-dimensional feature vectors of all models in the database, multiple similarity scores s are obtained. The electrode model corresponding to the highest score s is selected, thus achieving electrode model matching.

[0053] like Figure 2 As shown, the blue rectangle represents the search area of ​​the custom caliper tool, arranged parallel to the edge of the electrode (the light-dark boundary line on the left side of the image), covering a local area along the edge normal direction. The length (vertical) and width (horizontal) of the caliper set are dynamically adjusted according to the detection parameters of the electrode model. The caliper set uses a block-based search strategy to transform global edge detection into parallel processing of multiple local areas, improving detection efficiency. Simultaneously, the geometric parameters of the calipers (such as spacing and dimensions) can be adaptively adjusted according to the electrode model, adapting to the edge distribution characteristics of electrodes of different specifications.

[0054] Step 4: Obtain candidate edge points within the search area and use a dual strategy to filter and exclude false edge points: when the maximum cluster ratio of candidate edge points reaches a preset threshold, use the anchor point of the maximum cluster to filter neighboring points as real edge points; when the maximum cluster ratio does not reach the preset threshold, select the best point as a real edge point based on gradient filtering conditions and preset side direction.

[0055] After obtaining a series of candidate edge points within the search area through gradient calculation or convolution operations, due to irregularities such as stacking, bonding, embedding, or breakage that may occur during the electrode production process, these candidate edge points often contain a large number of pseudo-edge points (such as noise, stacked points, or broken edge points). This embodiment innovatively introduces a dual-strategy screening mechanism based on the maximum cluster ratio. When the maximum cluster ratio reaches a preset threshold, it indicates that the current edge shape is relatively regular and the real edge points are highly clustered in space. At this time, the first strategy is adopted: using the anchor point of the maximum cluster to screen neighboring points, and using spatial clustering to filter out discrete pseudo-edge points that deviate from the baseline. When the maximum cluster ratio does not reach the preset threshold, it indicates that the current edge is broken or irregular and cannot form a high-ratio cluster. At this time, it automatically switches to the second strategy: screening the best point based on gradient screening conditions and preset side direction, using the geometric boundary constraints of the real edge grayscale jump (high gradient intensity) and the electrode edge necessarily being at the leftmost or rightmost extreme position to lock the best edge point. Specifically,

[0056] Based on the candidate edge point set of the caliper group, clustering is performed using horizontal coordinates, and the clustering neighborhood radius is set. In this embodiment, the following settings are provided. This yields clusters. When the largest cluster comprises more than 60% of the total points, the median horizontal coordinate of the points within that cluster is taken. Use these as edge anchor points. Iterate through all candidate calipers in the caliper group, and for each candidate caliper edge point, select the one whose horizontal coordinates satisfy the condition... The point with the smallest distance within the caliper set is selected as the edge point and forms the final set of edge points for the caliper set.

[0057] In edge detection of images with damaged or highly irregular edges, the largest cluster cannot exceed 60% after clustering these edge point sets. Therefore, a second strategy is adopted for edge point selection. First, for each caliper's candidate edge point set, the maximum gradient in the caliper is determined by the element with the largest absolute value in the one-dimensional array. Furthermore, a proportional threshold is set; in this paper, it is set to 0.75, i.e. The threshold operation is applied to the candidate edge points of the calipers to remove edge points that do not meet the criteria. Then, based on the leftmost or rightmost pattern in the caliper group detection parameters, the edge point with the smallest or largest horizontal coordinate among these second-filtered candidate points is selected as the best edge point in the caliper. The above steps are repeated to traverse all calipers in the caliper group, forming the final set of caliper group edge points.

[0058] Step 5: Perform sub-pixel precision calculation and iterative fitting based on the real edge points to determine the final edge segments, and calculate the dimensions of each component of the strip workpiece based on the final edge segments.

[0059] To overcome the limitations of physical pixel resolution, sub-pixel precision calculations are performed on real edge points to improve edge point localization accuracy, resulting in a sub-pixel edge point set. This sub-pixel precision calculation can be achieved using the Gaussian approximation method, which involves fitting a Gaussian curve to the grayscale values ​​of the edge point and its neighboring pixels, using the extreme points of the curve as sub-pixel level position coordinates, thereby improving the localization accuracy from the integer pixel level to the sub-pixel level.

[0060] The one-dimensional Gaussian function is as follows:

[0061] (3)

[0062] when The time is a standard Gaussian function. In the set of edge points, the horizontal coordinates of the edge points are... This corresponds to the index in the one-dimensional projection array obtained in step 1). Indicates that the index is The value of the array element at that location. (Used) , , These three points are fitted with a Gaussian function to obtain the model parameter μ, which represents the sub-pixel position. δ, where

[0063] (4)

[0064] Calculate the sub-pixel offset and pixel edges The summation yields the edge points with sub-pixel precision.

[0065] Steps 6-7 involve performing an initial line segment fitting based on the sub-pixel edge point set to obtain an initial line segment model. Defective points are then eliminated based on the deviation of each edge point from the initial line segment model. The remaining edge points are then used for a second fitting to obtain the final edge line segment. For example, the initial fitting can use the least squares method to obtain the initial straight line equation. Then, the distance (i.e., deviation) from each edge point to the line is calculated. Points with the largest deviations (such as those in the bottom 20% of residuals) are considered defective points due to noise bias and are eliminated. Finally, the remaining high-quality edge points are used for a second fitting to obtain the accurate final edge line segment.

[0066] like Figure 3 The diagram illustrates the principle of edge fitting. The boundary between the black area on the left and the white area on the right represents the physical edge of the electrode. To determine the accurate location of this edge, this application uses a densely arranged set of calipers (blue rectangles) to scan the area. The gray-level extreme points extracted within each caliper frame constitute the original dataset. By clustering and performing secondary fitting on these densely distributed data points, local noise interference can be effectively filtered out (i.e., jagged edges are eliminated), thereby restoring the true straight edge of the electrode.

[0067] Based on the final edge segments, the system calculates the dimensions of each component of the electrode (such as aluminum foil width, electrode body length, copper foil width, etc.) according to the detection requirements corresponding to the currently registered specifications and models, based on the statistical analysis of the distance and positional deviations between specific segments, and determines whether they are qualified. This embodiment utilizes sub-pixel calculation and a second-order iterative fitting. The mechanism is that sub-pixel calculation eliminates the quantization error caused by pixel discretization, while the second-order iterative fitting removes the bias effect of occasional strong noise points on the overall segment model. The combination of these two methods achieves extremely high-precision edge segment reconstruction and dimension measurement.

[0068] like Figure 4As shown, multiple colored line segments are marked in the electrode image, each corresponding to a component edge of the electrode (such as the ceramic edge, aluminum foil boundary, or copper foil boundary). The values ​​next to the line segments (such as "141.42mm" and "138.91mm") are the measured results of the component dimensions. The yellow baseline is the central axis of the electrode, used for symmetry calculation. By visualizing the measurement results, the deviation between the manually measured values ​​and the algorithm values ​​of this embodiment can be intuitively verified. Different colored line segments in the figure can distinguish different components, facilitating the rapid location of out-of-tolerance areas.

[0069] The embodiments described above are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A method for detecting the size of lithium battery electrode sheets based on machine vision, characterized in that, Includes the following steps: Step 1: Acquire images of lithium battery electrode sheets during the rolling process using a high-resolution camera, and locate the approximate edge positions of the aluminum foil, copper foil, and ceramic edge in the electrode sheet images; Step 2: Based on the rough edge position, determine the model of the current electrode by matching the edge point registration algorithm; Step 3: Based on the matched electrode model, call the corresponding detection parameters and set a custom caliper tool search area near the rough edge position; Step 4: Within the search area of ​​the caliper tool, the edge points are initially classified using the clustering method, and false edge points are eliminated using the edge filtering method to obtain a set of candidate edge points. Step 5: Perform sub-pixel precision calculation on the candidate edge points to improve the edge point positioning accuracy and obtain a sub-pixel edge point set; Step 6: Based on the sub-pixel edge point set, edge fitting is performed by a two-step iterative fitting method: the first fitting obtains an initial line segment model, defective points are eliminated based on the deviation between the points and the initial line segment model, and the remaining points are used for a second fitting to obtain a high-precision final edge line segment. Step 7: Calculate the dimensions of each component of the lithium battery electrode based on the final edge segment.

2. The lithium battery electrode size detection method based on machine vision according to claim 1, characterized in that, In step 2, the edge point registration algorithm is specifically as follows: Vertical projection is performed on each column of pixels in the ROI image to obtain a one-dimensional projection array; The sign of the element value in the filtered array is used as the first row element of the two-dimensional feature vector, and the corresponding array index is used as the second row element to generate a two-dimensional feature vector describing the edge distribution. The two-dimensional feature vector of the current electrode is matched with the two-dimensional feature vectors of various electrode models pre-stored in the database, and the model with the highest matching score is taken as the matching result.

3. The lithium battery electrode size detection method based on machine vision according to claim 2, characterized in that, The length of the one-dimensional projection array is the same as the width of the ROI image, and the array element values... Represented as: ,in It is the image height. The coordinates in the image are The grayscale value of the pixel at the location; The matching score s is represented as: , In the formula, and These represent the values ​​in the first row and i-th column of the feature vector of the current electrode and a certain type of electrode in the database, respectively, representing the edge polarity; and represents the value of the second row and i-th column of the feature vectors of the two, respectively, representing the horizontal coordinates of the edge.

4. The lithium battery electrode size detection method based on machine vision according to any one of claims 1-3, characterized in that, In step 3, the clustering method is based on horizontal coordinates, including setting the cluster neighborhood radius. Clustering of candidate edge points yields clusters; When the proportion of the point set of the largest cluster reaches a preset threshold, the median horizontal coordinate of the point set within that cluster is taken. As edge anchor points; iterate through all calipers in the caliper group, and select the candidate edge points of each caliper that satisfy the condition that the horizontal coordinates are within the specified range. The point with the smallest distance within the range is selected as the edge point after filtering; Otherwise, for each set of candidate edge points of the caliper, determine its maximum gradient value. ,by Multiply by the proportional threshold as the filtering threshold to remove edge points that do not meet the conditions; then, based on the leftmost or rightmost mode in the caliper group detection parameters, select the edge point with the smallest or largest horizontal coordinate among the filtered candidate points as the best edge point of the caliper.

5. The lithium battery electrode size detection method based on machine vision according to claim 1, characterized in that, In step 5, sub-pixel calibration is performed using Gaussian approximation, specifically: for the horizontal coordinates of pixel-level edge points... Get its value in the one-dimensional projection array. and its adjacent points , ;use , , Gaussian function fitting is performed on the three points to calculate the sub-pixel offset. : ， pixel-level edge point coordinates With offset Add them together to get the edge point positions with sub-pixel precision.

6. The lithium battery electrode size detection method based on machine vision according to claim 1, characterized in that, In step 6, the first fitting uses the least squares method to fit the linear equation; the method for eliminating bad points is as follows: calculate the distance between all edge points and the line obtained from the first fitting, sort them from near to far, discard the point set in the bottom A% of the sorted points, and use the point set in the top B% of the sorted points for the second fitting.

7. The lithium battery electrode size detection method based on machine vision according to claim 1, characterized in that, In step 7, based on the matched electrode model, the testing requirements for that model are obtained from the knowledge base. Based on the final edge line segment, the distance and position deviation between specific line segments are statistically analyzed to calculate the size of each component and determine whether it is qualified.

Images