Semiconductor laser cleaning method based on visual guidance

Abstract

A semiconductor laser cleaning method based on visual guidance relates to the technical field of semiconductor component laser cleaning, and comprises the following steps: acquiring three-dimensional geometric data and surface image data of a to-be-cleaned workpiece, and constructing a three-dimensional model of the to-be-cleaned workpiece; performing point cloud analysis on the three-dimensional model, and marking the gradient, Gaussian curvature, average curvature and texture feature set of each point in the three-dimensional model as key features; the three-dimensional model is converted into a topology undirected graph, key points and background points are marked in the topology undirected graph, weight marks of all undirected edges in the topology undirected graph are set, and nodes in the topology undirected graph are divided into the key points and the background points; path planning is conducted on key points in the three-dimensional model, a laser cleaning path is obtained, a laser parameter control model is constructed, and laser parameters of all the key points in the laser cleaning path are obtained; and real-time monitoring and feedback correction are performed on the laser cleaning operation of the to-be-cleaned workpiece and the laser parameter control model, so that the cleaning quality and efficiency are improved.

Description

Technical Field

[0001] The invention relates to the technical field of laser cleaning of semiconductor parts, and in particular to a semiconductor laser cleaning method based on vision guidance. Background Art

[0002] CN116078751A "A pulse laser cleaning control device and cleaning control method" includes: after the host computer module establishes a connection with the control unit module, the host computer module sends control command parameters to the control unit module; based on the received control command parameters, the control unit module controls the laser generation module to perform corresponding laser operations by exchanging data with the laser generation module.

[0003] CN117798137A "An intelligent laser cleaning control system" includes an acquisition module that collects three-dimensional point cloud data of the bogie frame in real time; a processing module generates the three-dimensional coordinates, deflection angle and surface color data of the bogie frame according to the three-dimensional point cloud data, and obtains a laser output instruction based on the surface color data processing; a transportation module transports the bogie frame to a cleaning station according to the three-dimensional coordinates of the frame, and a composite laser cleaning device is provided at the cleaning station; a position adjustment module adjusts the three-dimensional spatial position of the composite laser cleaning device according to the three-dimensional coordinates of the frame and the deflection angle; a first emitting unit outputs a semiconductor laser to the bogie frame according to the laser output instruction to reduce the bonding force between the metal material of the bogie frame and the surface attachments; a second emitting unit outputs a pulsed laser to the bogie frame according to the laser output instruction to separate the surface attachments from the metal material.

[0004] The existing technology usually uses manual teaching, and can only perform simple plane scanning planning, control the robot to perform indiscriminate scanning from left to right and from top to bottom, and can only clean the top view of the workpiece. It cannot clean other surfaces when turning over. It has high limitations and cannot dynamically adjust the laser parameters. Summary of the invention

[0005] In order to solve the above technical problems, the object of the present invention is to provide a semiconductor laser cleaning method based on vision guidance, comprising the following steps:

[0006] Step s1: collecting three-dimensional geometric data and surface image data of the workpiece to be cleaned through a three-dimensional scanning terminal, and constructing a three-dimensional model of the workpiece to be cleaned according to the three-dimensional geometric data;

[0007] Step s2: performing point cloud analysis on the three-dimensional model to obtain the gradient, Gaussian curvature and mean curvature of each point in the three-dimensional model, performing texture analysis on the surface image data of the workpiece to be cleaned to obtain the texture feature set of each point in the three-dimensional model, and marking the gradient, Gaussian curvature, mean curvature and texture feature set of each point in the three-dimensional model as key features;

[0008] Step s3: Convert the 3D model into a topological undirected graph, mark the key points and background points in the topological undirected graph, set the weight annotations for each undirected edge in the topological undirected graph, and divide the nodes in the topological undirected graph into key points and background points;

[0009] Step s4: Perform path planning on the key points in the 3D model, obtain the laser cleaning path, construct a laser parameter control model, and obtain the laser parameters for each key point in the laser cleaning path;

[0010] Step s5: Perform real-time monitoring and feedback correction on the laser cleaning operation of the workpiece to be cleaned and the laser parameter control model.

[0011] Furthermore, the process of obtaining the gradient, Gaussian curvature, and mean curvature of each point in the 3D model by performing point cloud analysis on the 3D model includes:

[0012] Obtain the point cloud data included in the 3D model, preprocess the point cloud data, obtain the normal vector of each point in the preprocessed point cloud data, define a neighborhood for each point, fit a local plane within the neighborhood of each point, and based on the local planes fitted for each point, obtain the gradient of each point, and at the same time obtain the Gaussian curvature and mean curvature of each point.

[0013] Furthermore, the process of obtaining the texture feature set of each point in the 3D model by performing texture analysis on the surface image data of the workpiece to be cleaned includes:

[0014] Preprocess the surface image data of the workpiece to be cleaned, obtain the gray-level co-occurrence matrix of each pixel point in the preprocessed surface image data, obtain the texture feature set of each pixel point according to the gray-level co-occurrence matrix of each pixel point, match each pixel point in the surface image data with each point in the point cloud data on the surface of the 3D model, obtain the corresponding relationship between each pixel point and each point in the point cloud data, and obtain the texture feature set of each point in the point cloud data according to the corresponding relationship between each pixel point and each point in the point cloud data.

[0015] Furthermore, the process of converting the 3D model into a topological undirected graph includes:

[0016] Obtain the key features corresponding to each point in the point cloud data included in the 3D model, convert the 3D model into a topological undirected graph, use each point in the point cloud data included in the 3D model as a node in the topological undirected graph, and import the key features corresponding to each point in the point cloud data included in the 3D model into the node corresponding to each point;

[0017] Mark key points and background points in the topological directed graph, construct undirected edges between adjacent nodes in the topological undirected graph, and at the same time construct undirected edges between each node in the topological undirected graph and the key points, and between each node and the background points.

[0018] Furthermore, set the weight annotations of each undirected edge in the topological undirected graph. The process of dividing the nodes in the topological undirected graph into key points and background points includes:

[0019] Compare the key features between adjacent nodes to obtain the key feature similarity between adjacent nodes. According to the key feature similarity between adjacent nodes, set the weight labels of the undirected edges between adjacent nodes. At the same time, compare the key features of each node with the key features of the key points and background points respectively to obtain the key feature similarity between each node and the key points and the key feature similarity between each node and the background points. According to the key feature similarity between each node and the key points, set the weight labels of the undirected edges between each node and the key points. According to the key feature similarity between each node and the background points, set the weight labels of the undirected edges between each node and the background points;

[0020] Construct an energy function according to the weight labels of each undirected edge in the topological undirected graph, use the graph cut algorithm to segment the topological undirected graph based on the energy function, and divide the nodes in the topological undirected graph into key points and background points according to the segmentation result.

[0021] Furthermore, the process of marking key points and background points in the topological directed graph includes:

[0022] Obtain the standard texture parameters of the workpiece to be cleaned, match the texture feature set of each point in the topological directed graph with the standard texture parameters to obtain the texture parameter similarity between each point and the standard texture parameters, screen out the point corresponding to the minimum texture parameter similarity, mark the point as a key point, screen out the point corresponding to the maximum texture parameter similarity, and mark the point as a background point.

[0023] Furthermore, the process of path planning for the key points in the three-dimensional model, obtaining the laser cleaning path, and constructing the laser parameter control model to obtain the laser parameters of each key point in the laser cleaning path includes:

[0024] Use the global path planning algorithm to perform path planning on the key points in the three-dimensional model to obtain the laser cleaning path. Construct a laser parameter control model based on deep learning, input the key features of each key point in the laser cleaning path and the key features of the background points adjacent to each key point into the laser parameter control model, and obtain the laser parameters of each key point in the laser cleaning path according to the output of the laser parameter control model.

[0025] Furthermore, the process of real-time monitoring and feedback correction for the laser cleaning operation and the laser parameter control model of the workpiece to be cleaned includes:

[0026] Perform laser cleaning operation on the workpiece to be cleaned according to the laser cleaning path and the laser parameters of each key point in the laser cleaning path. After the laser cleaning operation is completed, obtain the texture feature sets of each key point in the laser cleaning path and the texture feature sets of the background points adjacent to each key point. Match the texture feature sets of each key point and the texture feature sets of the background points adjacent to each key point with the standard texture parameters to obtain the texture parameter similarity of each key point and each background point. Compare the texture parameter similarity of each key point and each background point with the preset texture parameter similarity threshold. If the texture parameter similarity of each key point and each background point is greater than or equal to the texture parameter similarity threshold, the laser cleaning is completed;

[0027] If the texture parameter similarity of a key point or a background point is less than the texture parameter similarity threshold, mark the key point or the background point as an abnormal point, and feedback the key features and laser parameters of the abnormal point to the relevant operator. The relevant operator performs manual annotation of the laser parameters for the abnormal point, and performs a reactivated cleaning operation on the workpiece to be cleaned according to the laser parameters of the manually annotated abnormal point. At the same time, supplement the key features of the abnormal point and the manually annotated laser parameters to the training set of the laser parameter control model, retrain the laser parameter control model, and output the laser parameter control model after the retraining is completed.

[0028] Compared with the prior art, the beneficial effects of the present invention are:

[0029] 1. Improve the cleaning efficiency and accuracy: By obtaining the geometric data of the workpiece through three-dimensional scanning and comprehensively analyzing it in combination with the surface image data, the area to be cleaned can be identified more accurately. Utilizing features such as gradients, Gaussian curvatures, and mean curvatures obtained from point cloud analysis, as well as texture feature sets, makes the laser cleaning path planning more precise. This not only reduces the ineffective cleaning area but also ensures the cleaning quality.

[0030] 2. Realize adaptive parameter setting: Based on key features (including gradients, Gaussian curvatures, mean curvatures, and texture features), a laser parameter control model is constructed, which can automatically adjust laser parameters such as power density, pulse width, and spot size according to different surface characteristics. This adaptability helps to avoid problems such as overheating damage or incomplete cleaning, thereby improving the cleaning quality and safety.

[0031] 3. Enhanced complex structure processing ability: For workpieces with complex geometries or irregular surfaces, traditional laser cleaning methods often struggle to achieve ideal results. However, by converting the 3D model into a topological undirected graph and optimizing the path planning for key points, this invention effectively solves this problem, enabling even surfaces with rich details or unevenness to be efficiently cleaned.

[0032] 4. Improved real-time monitoring and feedback correction ability: During the cleaning process, the system can monitor the cleaning status of each key point in real time and dynamically adjust the laser parameters according to the actual cleaning effect. If certain areas do not meet the expected standards, they will be marked as abnormal points and the operator will be notified to intervene. At the same time, the training set will be updated to optimize the model performance. This closed-loop control system ensures that the entire cleaning process is always in the best state.

[0033] 5. Reduced need for manual intervention: Through highly automated vision guidance technology and intelligent algorithms, the influence of human factors on the cleaning results is greatly reduced. Operators only need to make a small number of manual adjustments when necessary, thus saving a large amount of time and labor costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a schematic diagram of a vision-guided semiconductor laser cleaning method according to an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0035] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0036] As Figure 1 shown, a vision-guided semiconductor laser cleaning method includes the following steps:

[0037] Step S1: Collect the 3D geometric data and surface image data of the workpiece to be cleaned through a 3D scanning terminal, and construct a 3D model of the workpiece to be cleaned according to the 3D geometric data;

[0038] Step S2: Perform point cloud analysis on the 3D model to obtain the gradient, Gaussian curvature, and mean curvature of each point in the 3D model, perform texture analysis on the surface image data of the workpiece to be cleaned to obtain the texture feature set of each point in the 3D model, and mark the gradient, Gaussian curvature, mean curvature, and texture feature set of each point in the 3D model as key features;

[0039] Step S3: Convert the three-dimensional model into a topological undirected graph, mark the key points and background points in the topological undirected graph, set the weight annotations of each undirected edge in the topological undirected graph, and divide the nodes in the topological undirected graph into key points and background points;

[0040] Step S4: Perform path planning on the key points in the three-dimensional model, obtain the laser cleaning path, construct a laser parameter control model, and obtain the laser parameters of each key point in the laser cleaning path;

[0041] Step S5: Perform real-time monitoring and feedback correction on the laser cleaning operation of the workpiece to be cleaned and the laser parameter control model.

[0042] It should be further noted that in the specific implementation process, the process of performing point cloud analysis on the three-dimensional model to obtain the gradient, Gaussian curvature, and mean curvature of each point in the three-dimensional model includes:

[0043] Obtain the point cloud data contained in the three-dimensional model, and perform preprocessing on the point cloud data. The preprocessing includes performing point cloud cleaning to remove noise points or outliers to ensure the data quality for subsequent processing, using the uniform downsampling method to reduce the number of points in the point cloud to reduce the computational complexity while maintaining the main geometric characteristics of the model, obtain the normal vector of each point in the preprocessed point cloud data. In the present invention, the NormalEstimation class in the PCL library is used for normal estimation. Define a neighborhood for each point, and the points within this neighborhood will be used for subsequent gradient calculations. The selection of the neighborhood size depends on the point cloud density and the required level of detail. The specific steps will not be elaborated here. Fit a local plane within the neighborhood of each point. The specific process of fitting a local plane within the neighborhood of each point is to calculate the average position of all points within the neighborhood, that is, the centroid C. Taking the centroid C as the origin, calculate the position vector of each point p i relative to the centroid C, and then use these vectors to construct a 3x3 covariance matrix Σ. N represents the total number of points within the neighborhood. Perform eigenvalue decomposition on the covariance matrix Σ to obtain three eigenvalues λ1, λ2, λ3 and the corresponding unit eigenvectors v1, v2, v3. If v1 is the eigenvector corresponding to the smallest eigenvalue, then the eigenvector corresponding to the smallest eigenvalue v1 is the normal direction of the plane to be found. Using the normal vector v3 and the centroid C, the equation of the local plane ax + by + cz = d can be expressed, where a, b, c are the components of the eigenvector corresponding to the smallest eigenvalue, and d is a constant term calculated by substituting the coordinates of any point on the plane. Based on the local planes fitted for each point, obtain the gradient of each point, and at the same time obtain the Gaussian curvature and mean curvature of each point. Among them, the Gaussian curvature and mean curvature of each point are obtained using the eigenvalues of the covariance matrix. Gaussian curvature = λ1 × λ2, where λ1, λ2 are the two smallest eigenvalues of the covariance matrix Σ, and mean curvature = (λ1 + λ2) / 2.

[0044] It should be further noted that in the specific implementation process, the process of performing texture analysis on the surface image data of the workpiece to be cleaned and obtaining the texture feature set of each point in the three-dimensional model includes:

[0045] Preprocess the surface image data of the workpiece to be cleaned. It should be noted that there may be noise problems in the surface image data of the workpiece to be cleaned. Therefore, in an embodiment of the present invention, after obtaining the surface image data, it is necessary to perform image preprocessing operations on it to improve the quality of the surface image data and obtain the surface image data for subsequent analysis. Improving the image quality through image preprocessing operations is a well-known image processing means for those skilled in the art. In the embodiment of the invention, noise in the image is removed by the Gaussian filtering method, and then the surface image data is equalized to improve the contrast in the image, and the surface image data is grayscaled to convert the color image into a grayscale image. The grayscaling process is usually achieved by averaging the weights of the color channels. Among them, denoising and equalization processing, semantic segmentation network, and grayscaling are well-known technical means for those skilled in the art, and the specific steps will not be elaborated. Obtain the gray-level co-occurrence matrix of each pixel point in the preprocessed surface image data. Specific details are as follows: Quantize the gray-level of the surface image data into discrete gray-levels, divide the gray-value range into several levels. For example, an 8-bit image can be divided into 16, 32, 64 levels, and define the parameters required for the gray-level co-occurrence matrix, including distance (d) and direction (θ), which are determined according to actual application requirements. The specific steps will not be elaborated. According to the gray-level co-occurrence matrix of each pixel point, obtain the texture feature set of each pixel point. Match each pixel point in the surface image data with each point in the point cloud data on the surface of the three-dimensional model to obtain the corresponding relationship between each pixel point and each point in the point cloud data. According to the corresponding relationship between each pixel point and each point in the point cloud data, obtain the texture feature set of each point in the point cloud data. The texture features in the texture feature set include but are not limited to contrast, correlation, energy, homogeneity, entropy, etc.

[0046] It should be further noted that in the specific implementation process, the process of converting the three-dimensional model into a topological undirected graph includes:

[0047] Obtain the key features corresponding to each point in the point cloud data included in the three-dimensional model, convert the three-dimensional model into a topological undirected graph, use each point in the point cloud data included in the three-dimensional model as the nodes of the topological undirected graph, and import the key features corresponding to each point in the point cloud data included in the three-dimensional model into the nodes corresponding to each point;

[0048] Mark key points and background points in the topological directed graph, construct undirected edges between adjacent nodes in the topological undirected graph, and at the same time construct undirected edges between each node in the topological undirected graph and the key points, as well as undirected edges between each node and the background points.

[0049] It should be further noted that in the specific implementation process, setting the weight annotation of each undirected edge in the topological undirected graph, and the process of dividing the nodes in the topological undirected graph into key points and background points includes:

[0050] Compare the key features between adjacent nodes to obtain the key feature similarity SIM between adjacent nodes. According to the key feature similarity SIM between adjacent nodes, set the weight label D of the undirected edge between adjacent nodes, where, δ represents the conversion coefficient. At the same time, compare the key features of each node with the key features of the key points and background points respectively to obtain the key feature similarity between each node and the key points and the key feature similarity between each node and the background points. According to the key feature similarity between each node and the key points, set the weight label of the undirected edge between each node and the key points. According to the key feature similarity between each node and the background points, set the weight label of the undirected edge between each node and the background points;

[0051] Construct an energy function according to the weight labels of each undirected edge in the topological undirected graph, and use the graph cut algorithm to segment the topological undirected graph based on the energy function. The present invention uses the minimum cut algorithm in the graph cut algorithm for segmentation, continuously cutting off the path from the source point to the sink point until the sum of the weights of all the cut edges is the smallest (i.e., the energy function is minimized). Subsequently, mark the nodes connected to the source point at this time as key points, and mark the nodes connected to the sink point as key points. Divide the nodes in the topological undirected graph into key points and background points according to the segmentation result.

[0052] It should be further noted that the core of the graph cut algorithm is to find a minimum cut, that is, a segmentation method that cuts off the path from the source point to the sink point, and the cost of this cut (i.e., the sum of the weights of all the cut edges) is the smallest. Once the energy function is defined, it can be converted into a minimum cut problem on the graph. Then, by solving the minimum cut problem, a cut that divides the graph into two parts can be found, and this cut corresponds to the optimal segmentation of the image. Among them, the calculation formula of the energy function is:

[0053] E(A)=∑ p∈P D (p，a) +∑ p∈P D (p，b) +γ∑ (p，q)∈N D (p，q) ;

[0054] Among them, A represents a partition of a topological undirected graph, P is the set of all points, N is the set of all adjacent points, and D (p，a) represents the weight label of the undirected edge between point p and key point a, and D (p，b) represents the weight label of the undirected edge between point p and background point b, and D (p，a) and D (p，b) is used to measure the cost of assigning a certain node to a key point or a background point, which is calculated based on gradient, Gaussian curvature, mean curvature, and texture feature set. For example, if the mean curvature of a point is very similar to the mean curvature of a known key point, then the cost of labeling it as a key point is small, ensuring that each point is correctly assigned to the corresponding category, and D (p，q) represents the weight label of the undirected edge between point p and point q, and D (p，q) is used to encourage the property of consistency within the same region, such as the consistency of gradient, texture, or curvature. If two adjacent nodes belong to the same region, it is expected that there is a small difference between them, thereby reducing the value of the energy function, ensuring the spatial consistency of the segmentation result, encouraging adjacent pixels or feature points to have similar labels, so as to avoid generating unnatural segmentation boundaries, and thus helping to maintain the continuity and smoothness of the region.

[0055] It should be further noted that in the specific implementation process, the process of marking key points and background points in a topological directed graph includes:

[0056] Obtain the standard texture parameters of the workpiece to be cleaned, perform feature matching between the texture feature set of each point in the topological directed graph and the standard texture parameters, obtain the texture parameter similarity of each point and the standard texture parameters, screen out the point corresponding to the minimum texture parameter similarity, mark the point as a key point, screen out the point corresponding to the maximum texture parameter similarity, and mark the point as a background point.

[0057] It should be further noted that in the specific implementation process, the process of performing path planning on the key points in the three-dimensional model, obtaining the laser cleaning path, and constructing a laser parameter control model to obtain the laser parameters of each key point in the laser cleaning path includes:

[0058] Use global path planning algorithms (such as A* algorithm, RRT-Connect, etc.) to perform path planning on key points in the 3D model, obtain the laser cleaning path, construct a laser parameter control model based on deep learning, and based on the idea of imitation learning, establish a virtual working environment on the computer to simulate the entire cleaning process. Obtain the corresponding optimal simulated laser parameters under different simulated key features of key points in the laser cleaning path and different simulated key features of background points adjacent to the key points. Use the different simulated key features of key points and the corresponding optimal simulated laser parameters of background points adjacent to the key points as the training set and the test set. Input the training set into the laser parameter control model for training until the loss function training is stable, save the model parameters, test the laser parameter control model with the test set until it meets the preset requirements, output the laser parameter control model, input the key features of each key point in the laser cleaning path and the key features of background points adjacent to each key point into the laser parameter control model, and output the laser parameters of each key point in the laser cleaning path according to the laser parameter control model.

[0059] It should be further noted that in the specific implementation process, the process of real-time monitoring and feedback correction of the laser cleaning operation of the workpiece to be cleaned and the laser parameter control model includes:

[0060] Perform laser cleaning operations on the workpiece to be cleaned according to the laser cleaning path and the laser parameters of each key point in the laser cleaning path. After the laser cleaning operation is completed, obtain the texture feature set of each key point in the laser cleaning path and the texture feature set of background points adjacent to each key point. Match the texture feature sets of each key point and the texture feature sets of background points adjacent to each key point with the standard texture parameters to obtain the texture parameter similarity of each key point and each background point. Compare the texture parameter similarity of each key point and each background point with the preset texture parameter similarity threshold. If the texture parameter similarity of each key point and each background point is greater than or equal to the texture parameter similarity threshold, the laser cleaning is completed;

[0061] If there is a key point or a background point whose texture parameter similarity is less than the texture parameter similarity threshold, mark the key point or the background point as an abnormal point, feedback the key features and laser parameters of the abnormal point to the relevant operator, and let the relevant operator perform manual annotation of the laser parameters for the abnormal point. Perform a reactivated cleaning operation on the workpiece to be cleaned according to the laser parameters of the manually annotated abnormal point. At the same time, supplement the key features of the abnormal point and the manually annotated laser parameters to the training set of the laser parameter control model, retrain the laser parameter control model, and output the laser parameter control model after retraining.

[0062] The above embodiments are only used to illustrate the technical method of the present invention and not to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical method of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A semiconductor laser cleaning method based on vision guidance, characterized in that: The following steps are involved: Step s1: collecting three-dimensional geometric data and surface image data of the workpiece to be cleaned through a three-dimensional scanning terminal, and constructing a three-dimensional model of the workpiece to be cleaned according to the three-dimensional geometric data; Step s2: performing point cloud analysis on the three-dimensional model to obtain the gradient, Gaussian curvature and mean curvature of each point in the three-dimensional model, performing texture analysis on the surface image data of the workpiece to be cleaned to obtain the texture feature set of each point in the three-dimensional model, and marking the gradient, Gaussian curvature, mean curvature and texture feature set of each point in the three-dimensional model as key features; Step s3: converting the three-dimensional model into a topological undirected graph, marking key points and background points in the topological undirected graph, setting weight labels for each undirected edge in the topological undirected graph, and dividing the nodes in the topological undirected graph into key points and background points; Step s4: perform path planning for key points in the three-dimensional model, obtain the laser cleaning path, build a laser parameter control model, and obtain the laser parameters of each key point in the laser cleaning path; Step s5: Real-time monitoring and feedback correction of the laser cleaning operation of the workpiece to be cleaned and the laser parameter control model.

2. The semiconductor laser cleaning method based on vision guidance according to claim 1, characterized in that: The process of performing point cloud analysis on a 3D model and obtaining the gradient, Gaussian curvature and mean curvature of each point in the 3D model includes: The point cloud data contained in the 3D model is obtained, the point cloud data is preprocessed, the normal vector of each point in the preprocessed point cloud data is obtained, a neighborhood is defined for each point, a local plane is fitted in the neighborhood of each point, and the gradient of each point is obtained based on the local plane fitted at each point, and the Gaussian curvature and the mean curvature of each point are obtained.

3. The semiconductor laser cleaning method based on vision guidance according to claim 2, characterized in that: The process of performing texture analysis on the surface image data of the workpiece to be cleaned and obtaining the texture feature set of each point in the three-dimensional model includes: The surface image data of the workpiece to be cleaned is preprocessed to obtain the grayscale co-occurrence matrix of each pixel in the preprocessed surface image data, and the texture feature set of each pixel is obtained according to the grayscale co-occurrence matrix of each pixel. Each pixel in the surface image data is matched with each point in the point cloud data of the three-dimensional model surface to obtain the correspondence between each pixel and each point in the point cloud data, and the texture feature set of each point in the point cloud data is obtained according to the correspondence between each pixel and each point in the point cloud data.

4. The semiconductor laser cleaning method based on vision guidance according to claim 3 is characterized in that: The process of converting a 3D model into a topological undirected graph includes: Obtaining key features corresponding to each point in the point cloud data contained in the three-dimensional model, converting the three-dimensional model into a topological undirected graph, using each point in the point cloud data contained in the three-dimensional model as a node of the topological undirected graph, and importing key features corresponding to each point in the point cloud data contained in the three-dimensional model into the node corresponding to each point; Key points and background points are marked in the topological directed graph, undirected edges are constructed between adjacent nodes in the topological undirected graph, and undirected edges are constructed between each node and the key points in the topological undirected graph, as well as between each node and the background points.

5. The semiconductor laser cleaning method based on vision guidance according to claim 4, characterized in that: The process of setting weight labels for each undirected edge in the topological undirected graph and dividing the nodes in the topological undirected graph into key points and background points includes: Compare the key features between adjacent nodes to obtain the key feature similarity between adjacent nodes, set the weight label of the undirected edge between adjacent nodes according to the key feature similarity between adjacent nodes, and compare the key features of each node with the key features of the key points and the background points respectively to obtain the key feature similarity between each node and the key points and the key feature similarity between each node and the background points, set the weight label of the undirected edge between each node and the key points according to the key feature similarity between each node and the key points, and set the weight label of the undirected edge between each node and the background points according to the key feature similarity between each node and the background points; An energy function is constructed according to the weight labels of each undirected edge in the topological undirected graph. The topological undirected graph is segmented using the graph cut algorithm based on the energy function. The nodes in the topological undirected graph are divided into key points and background points according to the segmentation results.

6. The semiconductor laser cleaning method based on vision guidance according to claim 5, characterized in that: The process of marking key points and background points in a topological directed graph includes: The standard texture parameters of the workpiece to be cleaned are obtained, and the texture feature set of each point in the topological directed graph is feature matched with the standard texture parameters, and the texture parameter similarity of each point and the standard texture parameters is obtained. The points corresponding to the minimum texture parameter similarity are screened out, and the points are marked as key points. The points corresponding to the maximum texture parameter similarity are screened out, and the points are marked as background points.

7. The semiconductor laser cleaning method based on vision guidance according to claim 6, characterized in that: The process of performing path planning for key points in the three-dimensional model, obtaining the laser cleaning path, building a laser parameter control model, and obtaining the laser parameters of each key point in the laser cleaning path includes: A global path planning algorithm is used to plan the paths of the key points in the three-dimensional model to obtain the laser cleaning path. A laser parameter control model is constructed based on deep learning. The key features of each key point in the laser cleaning path and the key features of the background points adjacent to each key point are input into the laser parameter control model. The laser parameters of each key point in the laser cleaning path are output according to the laser parameter control model.

8. The semiconductor laser cleaning method based on vision guidance according to claim 7, characterized in that: The process of real-time monitoring and feedback correction of the laser cleaning operation of the workpiece to be cleaned and the laser parameter control model includes: A laser cleaning operation is performed on the workpiece to be cleaned according to the laser cleaning path and the laser parameters of each key point in the laser cleaning path. After the laser cleaning operation is completed, a texture feature set of each key point in the laser cleaning path and a texture feature set of background points adjacent to each key point are obtained, and the texture feature set of each key point and the texture feature set of background points adjacent to each key point are feature matched with standard texture parameters to obtain the texture parameter similarity of each key point and each background point, and the texture parameter similarity of each key point and each background point is compared with a preset texture parameter similarity threshold. If the texture parameter similarity of each key point and each background point is greater than or equal to the texture parameter similarity threshold, the laser cleaning is completed. If there are key points or background points whose texture parameter similarity is less than the texture parameter similarity threshold, the key points or background points are marked as abnormal points, and the key features and laser parameters of the abnormal points are fed back to the relevant operators, who manually mark the laser parameters of the abnormal points, and reactivate the cleaning operation for the workpiece to be cleaned according to the laser parameters of the manually marked abnormal points. At the same time, the key features of the abnormal points and the manually marked laser parameters are added to the training set of the laser parameter control model, the laser parameter control model is retrained, and the retrained laser parameter control model is output.

Images