A machine vision-based silicone rubber product detection method and system

Abstract

This invention relates to the field of silicone rubber product inspection technology, specifically to a machine vision-based method and system for silicone rubber product inspection. The method includes: acquiring a three-dimensional theoretical digital model of the silicone rubber product; performing mesh analysis on the smooth surface; extracting the surface normal vectors of each discrete mesh point; and offsetting the surface by a preset recognition distance along the outward normal direction to obtain multiple initial measurement space coordinates; using a shortest path graph theory optimization algorithm to perform topological recombination and redundancy removal on each initial measurement space coordinate to generate an optimized spatial scanning trajectory; driving a multi-axis industrial robotic arm equipped with a laser displacement sensor to move uniformly along the optimized spatial scanning trajectory to acquire a three-dimensional depth topography data point cloud; synchronously controlling a vision camera system equipped with a multi-directional adjustable closed light source to perform continuous exposure imaging to acquire a two-dimensional multi-illumination image data array; fusing the three-dimensional depth topography data point cloud with the two-dimensional multi-illumination image data array across modalities to construct a unified format dataset; and then inputting the unified format dataset into a deep convolutional neural network model to output a defect classification probability determination result. This invention can solve the problems of insufficient defect feature acquisition, poor stability in surface region detection, low detection efficiency, and insufficient classification accuracy.

Description

Technical Field

[0001] This invention relates to the field of silicone rubber product testing technology, and in particular to a method and system for testing silicone rubber products based on machine vision. Background Technology

[0002] Silicone rubber products are widely used in consumer goods, electronic components, and industrial parts. These products often have smooth surfaces, significant local curvature variations, and a reflective appearance. For these products, minor surface defects such as dents, scratches, and indentations are typically small in size and have indistinct boundaries. Under conventional visual inspection conditions, they are easily affected by factors such as high surface reflectivity, variations in surface morphology, and limited imaging angles, leading to unstable image representation of defect areas. This, in turn, results in missed detections, false detections, or inconsistent inspection results.

[0003] Among existing inspection methods, one approach relies primarily on ordinary cameras or manual visual inspection to determine surface quality, but it is poorly suited for products with high gloss or curved surfaces. Another approach, while incorporating laser measurement, image recognition, or deep learning classification techniques, still generally suffers from insufficient utilization of two-dimensional appearance and three-dimensional morphology information, weak correlation between different acquired data, and a difficulty in balancing inspection efficiency and accuracy for smooth curved surfaces. Especially for silicone rubber products with both smooth curved surfaces and highly reflective surfaces, how to reliably acquire effective defect information and improve the reliability of automated inspection remains a pressing technical problem that needs to be solved in this field. Summary of the Invention

[0004] In view of the above technical problems, the present invention provides a machine vision-based method and system for detecting silicone rubber products, in order to solve the problems of insufficient acquisition of defect features, poor detection stability of curved areas, low detection efficiency, and insufficient classification and recognition accuracy in the existing technology when detecting silicone rubber products with smooth curved surfaces and high reflectivity.

[0005] Other features and advantages of this disclosure will become apparent from the following detailed description, or may be learned in part from practice of this disclosure.

[0006] According to one aspect of the present invention, a machine vision-based method for inspecting silicone rubber products is proposed, the method comprising: A three-dimensional theoretical digital model of a silicone rubber product is obtained. The smooth surface of the three-dimensional theoretical digital model is analyzed by surface meshing. The surface normal vectors of each discrete mesh point are extracted. Based on the surface normal vectors, a preset identification distance is offset along the outer normal direction. Multiple initial measurement space coordinates that maintain an orthogonal attitude with the smooth surface are calculated. The shortest path graph theory optimization algorithm is used to perform global topology reorganization and redundant region removal on all the initial measurement spatial coordinates to generate a continuous optimized spatial scanning trajectory that does not overlap or reuse in space. A multi-axis industrial robotic arm equipped with a laser displacement sensor is driven to control the laser displacement sensor to perform uniform conformal motion along the optimized spatial scanning trajectory at a preset detection speed while maintaining a perpendicular incident relationship with the smooth curved surface. During the motion, three-dimensional depth topography data point cloud representing the physical depth of the smooth curved surface is collected. A vision camera system with synchronous control and fixed settings and equipped with a multi-directional adjustable closed light source continuously exposes and images the smooth area of ​​the silicone rubber product under different illumination angles and different light source spatial projection vector states to obtain a two-dimensional multi-illumination image data array. Perform cross-modal fusion at the feature level on the acquired 3D depth topography point cloud and the 2D multi-light image data array to construct a unified format dataset; The unified format dataset is input into a pre-established deep convolutional neural network model for forward propagation and high-dimensional feature extraction. Through nonlinear mapping of cascaded layers within the network, the corresponding defect classification probability determination result is obtained at the prediction output.

[0007] Furthermore, the generation of the optimized spatial scan trajectory specifically includes: A graph structure model is constructed in the system's computing memory, with all the initial measurement space coordinates as discrete topological nodes, the actual physical movement paths between each node as undirected data edges, and the expected robot motion time cost corresponding to each undirected data edge at the preset detection speed as a weighted cost. Initialize the global state matrix of the graph theory algorithm, specify the traversal starting point, set the cumulative path cost of the starting point to zero, initialize the cumulative path cost of the remaining nodes to the maximum value, and mark all the discrete topological nodes as unprocessed. The unprocessed node with the lowest current cumulative path cost is taken as the central reference node. Each adjacent unprocessed node directly connected to the central reference node is traversed, and the sum of the current cumulative path cost of the central reference node and the weighted cost of the undirected data edge connected to the corresponding adjacent unprocessed node is calculated. When the calculated summation value is less than the cumulative path cost currently recorded by the adjacent unprocessed node, the cumulative path cost of the adjacent unprocessed node is updated with the summation value, and the corresponding predecessor node mapping relationship is written into the predecessor relationship record memory. After all the adjacent edges of the central reference node have been calculated, the central reference node is marked as processed. Then, among the remaining unprocessed nodes, the node with the smallest cumulative path cost is selected as the central reference node for the next iteration, until all the initial measurement space coordinates are marked as processed. Read the predecessor relationship records in the predecessor relationship record memory, perform reverse backtracking from the end point to the starting point, and construct the optimized spatial scanning trajectory with the minimum expected robot motion time cost and elimination of repeated redundant coverage scanning nodes.

[0008] Furthermore, before performing the cross-modal fusion, the three-dimensional depth topography point cloud data is processed as follows: Based on the continuous motion trajectory of the laser displacement sensor in the time series and the depth measurement feedback parameters obtained from the reflected echo analysis, a spatial two-dimensional mapping structured data matrix is ​​constructed; wherein, the spatial two-dimensional mapping structured data matrix uses the horizontal motion axis as the first dimension x-coordinate, the vertical motion axis as the second dimension y-coordinate, and the depth value measured by the corresponding grid cell as the matrix height element; For the missing regions in the spatial two-dimensional mapped structured data matrix caused by mechanical vibration or sampling discontinuity, interpolation calculations are performed using the surrounding effective values ​​to smoothly compensate and fit the missing depth values, so as to convert the discrete collected scattered points into a topologically continuous three-dimensional geometric entity data structure. The spatial two-dimensional mapping structured data matrix, after smoothing compensation fitting, is input into the three-dimensional graphics rendering engine to generate a three-dimensional surface visualization simulation image with visual brightness and darkness feedback attributes. The three-dimensional surface visualization simulation image is used as the image representation result of the three-dimensional depth topography data point cloud, and is used to perform the cross-modal fusion with the two-dimensional multi-light image data array.

[0009] Furthermore, when the deep convolutional neural network model is pre-built, it performs data augmentation processing on the uniform format dataset used for training. This data augmentation processing includes brightness shift adjustment, scale scaling transformation, contrast stretching transformation, and channel gain scaling processing, wherein: For the brightness offset adjustment, a preset deviation value is superimposed on each channel element of the original pixel matrix of the image samples in the unified format dataset to simulate background illumination changes; For the scaling transformation, the spatial distribution relationship of the original feature point coordinates of the image samples is adjusted according to a preset scaling factor in order to reconstruct the image boundary resolution; For the contrast stretching transformation, the statistical mean of the pixel data of the entire image is calculated, and the difference between bright and dark pixels is scaled proportionally based on the statistical mean to enhance the gray-scale transition characteristics of the boundary of the tiny scratch. For the channel gain scaling process, a predefined multiplier is applied to the feature pixels in each channel dimension within the data matrix of the unified format dataset to change the response intensity corresponding to the local lighting environment.

[0010] Furthermore, when performing continuous exposure imaging on the smooth area of ​​the silicone rubber product to obtain the two-dimensional multi-illumination image data array, the specific steps include: In the first stage, the multi-directional adjustable closed light source is controlled to enter the diffuse coverage projection mode, and the visual camera system is continuously triggered to perform multiple exposure acquisitions according to a preset time interval, so as to suppress the halo overflow distortion caused by the bright area of ​​silicone rubber through time-series integration superposition. In the second stage, the emission projection vector of the multi-directional adjustable closed light source is switched to a posture with a first preset oblique angle relative to the vertical reference center axis of the vision camera system, and continuous exposure imaging is performed according to the same acquisition rhythm as in the first stage, so as to evaluate the micro-concave-convex texture distribution of the smooth area in the horizontal direction by utilizing the small shadow gradient formed under tangential lighting conditions. In the third stage, the emission projection vector of the multi-directional adjustable closed light source is deflected to a position with a second preset oblique angle relative to the system horizontal reference bottom surface of the vision camera system, and continuous exposure imaging is performed according to the acquisition parameters corresponding to the second stage to analyze the surface topological undulation changes of the smooth area in the vertical direction. During the data aggregation stage, the image arrays acquired in the first, second, and third stages are subjected to spatial registration and fusion processing to form the two-dimensional multi-illumination image data array that suppresses ambient background light diffusion interference and material specular reflection interference.

[0011] Furthermore, the step of inputting the unified format dataset into a pre-established deep convolutional neural network model for forward propagation and high-dimensional feature extraction, and obtaining the corresponding defect classification probability determination result at the prediction output, specifically includes: Pixel-level normalization is performed on the multi-channel composite data in the input unified format dataset. The original pixel values ​​are linearly scaled according to the full scale of the sensor grayscale, so that the feature elements of each pixel converge to a unified value range, thereby eliminating the dimensional differences between different sensor channels. The normalized data is input into the input layer of the deep convolutional neural network model, and convolution operations are performed through a group of convolutional layers set in a deep series to extract the edge, texture and high-order shape features of the image. Pooling layers are set at intervals along the forward propagation direction of the deep convolutional neural network model, and local region aggregation downsampling is performed on the feature map of the convolution output to reduce the feature dimension and enhance the robustness of the feature to local displacement perturbations. In the backend dimensionality reduction stage, the multidimensional feature tensors after convolution and pooling are converted into one-dimensional feature vectors through an unfolding layer, which are then used as input for subsequent classification decisions.

[0012] Furthermore, after converting the multidimensional feature tensor into a one-dimensional feature vector, the one-dimensional feature vector is input as a whole into the tail fully connected layer decision module. Each neuron in the fully connected layer decision module receives a weighted signal from the outputs of all associated neurons in the previous layer, and generates a node output by superimposing a bias term. A nonlinear activation function control unit is connected in series at the output of the fully connected layer decision module and the output of each preceding convolutional layer. The nonlinear activation function control unit located in the intermediate hidden layer employs a modified linear activation function to retain positive responses and suppress negative responses. The nonlinear activation function control unit located in the final prediction output layer uses a normalized probability distribution function to convert the one-dimensional feature vector into a posterior confidence probability vector corresponding to each defect category, and outputs the final defect classification result according to the principle of maximizing probability.

[0013] According to another aspect of the present invention, a machine vision-based inspection system for silicone rubber products is provided, comprising: The coordinate generation module is used to obtain a three-dimensional theoretical digital model of silicone rubber products, perform surface meshing analysis on the smooth surface of the three-dimensional theoretical digital model, extract the surface normal vector of each discrete mesh point, and offset the surface normal vector along the outward normal direction by a preset recognition distance, and calculate multiple initial measurement space coordinates that are orthogonal to the smooth surface. The scanning trajectory optimization module is used to perform global topological reorganization and redundant region removal on all the initial measurement spatial coordinates using the shortest path graph theory optimization algorithm, so as to generate a continuous optimized spatial scanning trajectory that does not overlap or reuse in space. A three-dimensional depth data acquisition module is used to drive a multi-axis industrial robotic arm equipped with a laser displacement sensor. The module controls the laser displacement sensor to move at a preset speed along the optimized spatial scanning trajectory while maintaining a perpendicular incident relationship with the smooth surface. During the movement, the module acquires three-dimensional depth topography data point cloud that characterizes the physical depth of the smooth surface. A two-dimensional multi-light image acquisition module is used to synchronously control a fixed visual camera system equipped with a multi-directional adjustable closed light source, and to continuously expose and image the smooth area of ​​the silicone rubber product under different lighting angles and different light source spatial projection vector states to acquire a two-dimensional multi-light image data array. The cross-modal fusion data construction module is used to perform feature-level cross-modal fusion on the acquired 3D depth topography point cloud and the 2D multi-light image data array to construct a unified format dataset. The defect identification module is used to input the unified format dataset into a pre-established deep convolutional neural network model for forward propagation and high-dimensional feature extraction, and obtain the corresponding defect classification probability judgment result at the prediction output through nonlinear mapping of cascaded layers within the network.

[0014] The technical solution of the present invention has the following beneficial effects: By taking into account the inspection characteristics of smooth curved surfaces of silicone rubber products and comprehensively utilizing surface morphology information and appearance image information under multiple illumination conditions, it is possible to reduce the interference of highly reflective surfaces on defect detection to a certain extent and improve the characterization ability of small defects in curved areas. Optimizing the detection trajectory helps reduce the time overhead caused by repeated scanning and improves the continuity and efficiency of the detection process. By unifying the organization of data from different sources and inputting it into a deep convolutional neural network model for feature extraction and defect classification, it is beneficial to enhance the automatic identification capability of defect categories such as dents and scratches, and improve the accuracy and stability of detection results. Attached Figure Description

[0015] Figure 1 This is a flowchart of a machine vision-based method for inspecting silicone rubber products, as described in the embodiments of this specification. Figure 2 This is a structural block diagram of a machine vision-based silicone rubber product inspection system as described in the embodiments of this specification. Detailed Implementation

[0016] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this disclosure more comprehensive and complete, and to fully convey the concept of the example embodiments to those skilled in the art. The described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a full understanding of embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced with one or more of the specific details omitted, or other methods, components, systems, steps, etc., can be employed. In other instances, well-known technical solutions are not shown or described in detail to avoid obscuring various aspects of this disclosure.

[0017] Furthermore, the accompanying drawings are merely illustrative of this disclosure. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor systems and / or microcontroller systems.

[0018] This invention provides a machine vision-based method for inspecting silicone rubber products. (Refer to...) Figure 1 The diagram shown is a flowchart illustrating a machine vision-based inspection method for silicone rubber products according to an embodiment of the present invention. The method specifically includes the following steps S101-S106: In step S101, a three-dimensional theoretical digital model of the silicone rubber product is obtained, the smooth surface of the three-dimensional theoretical digital model is analyzed by surface meshing, the surface normal vector of each discrete mesh point is extracted, and the surface normal vector is offset by a preset recognition distance along the outer normal direction, and multiple initial measurement space coordinates that maintain an orthogonal attitude with the smooth surface are calculated.

[0019] A three-dimensional theoretical digital model is used to characterize the overall shape and contour of the product under inspection, the surface transition relationships, and the spatial geometry of each local region. Due to the smooth surface and continuously varying curvature of silicone rubber products, defects are difficult to visualize stably when inspected using a fixed viewing angle because surface reflection and surface morphology changes can cause problems. Therefore, it is necessary to pre-determine suitable spatial pose information for measurement from the geometric model. To this end, surface meshing analysis is performed on the smooth surface of the three-dimensional theoretical digital model, discretizing the continuous surface into multiple mesh regions, and extracting the surface normal vector at each discrete mesh point. The surface normal vector reflects the local orientation of the surface at the corresponding mesh point and can serve as the geometric basis for setting the subsequent measurement direction. To reduce the influence of light reflection from the smooth surface and maintain a matching relationship between the measurement direction and the local surface morphology, a perpendicular correspondence is established between the measuring device and the surface sample. By making the measurement direction close to the surface normal direction, the impact of light reflection on the stability of subsequent inspection data is reduced.

[0020] After obtaining the surface normal vectors of each discrete grid point, a recognition distance offset is performed along the outward normal direction of each grid point to obtain spatial coordinate points located outside the surface of the workpiece. The recognition distance is used to limit the working interval between the measuring device and the workpiece surface to ensure measurement accuracy and safety. In one implementation, the recognition distance can be set according to the working requirements of the displacement sensor; for example, the coordinate points can be arranged at a fixed recognition distance above the surface to maintain a stable spatial correspondence between the acquired coordinates and the surface. The numerical control machining module can automatically generate a motion path adapted to the surface geometry based on the shape of the three-dimensional theoretical digital model, and consider the spatial avoidance relationship between the measuring device and the measured surface when generating coordinates, thereby avoiding interference between the measuring head and the workpiece. Each spatial coordinate obtained corresponds to a local measurement position on the surface, and the measurement posture at this position is orthogonal to the local normal direction of the surface, thus forming multiple initial measurement spatial coordinates. These initial measurement spatial coordinates are essentially a set of candidate measurement poses derived from the surface geometry information. Their function is to provide basic coordinate input for subsequent trajectory optimization and conformal surface measurement, and to ensure that the subsequent acquisition area corresponds as closely as possible to a spatial position with theoretically minimal reflection influence.

[0021] In practical applications, this invention can be deployed at online inspection stations on industrial production lines. After the silicone rubber product to be inspected enters the inspection area via a conveyor mechanism, it is first positioned at a predetermined workpiece location. Subsequently, a vision acquisition unit completes the imaging of the frontal area, and then a multi-axis industrial robotic arm drives a laser displacement sensor to perform conformal measurement on the curved surface area.

[0022] In step S102, the shortest path graph theory optimization algorithm is used to perform global topological reorganization and redundant region removal on all the initial measurement spatial coordinates to generate a continuous optimized spatial scanning trajectory that does not overlap or reuse in space.

[0023] The initial measurement spatial coordinates form candidate measurement positions. While these positions correspond to areas with low reflection on the curved surface, overlapping measurement areas, duplicate adjacent paths, and unreasonable overall movement order often exist between different coordinates. If scanning is performed directly according to the generated order, the acquisition device will repeatedly travel back and forth between multiple local areas, resulting in duplicate acquisition and invalid movement, significantly increasing the detection time. Therefore, all initial measurement spatial coordinates are transformed from a simple set of spatial points into discrete topological nodes in a graph structure. The executable physical movement relationships between nodes are then transformed into edge relationships to optimize the scanning order from a global perspective. After the graph structure is established, each edge no longer merely represents the geometric connectivity between two nodes but further characterizes the execution cost required for the acquisition device to move from one measurement coordinate to another. This execution cost can be obtained from the movement time between three-dimensional coordinates, essentially reflecting the time overhead of the robotic arm movement under preset detection speed conditions. This constructed graph structure transforms the curved surface scanning problem into a path optimization problem in a weighted graph, allowing subsequent trajectory planning to consider both spatial coverage integrity and execution efficiency.

[0024] Specifically, generating the optimized spatial scanning trajectory includes: constructing a graph structure model in the system's computing memory, using all the initial measured spatial coordinates as discrete topological nodes, the actual physical movement paths between each node as undirected data edges, and using the expected robot motion time cost corresponding to each undirected data edge at the preset detection speed as a weighted cost; initializing the global state matrix of the graph theory algorithm, specifying the traversal starting point, setting the cumulative path cost of the starting point to zero, initializing the cumulative path cost of the remaining nodes to a maximum value, and marking all the discrete topological nodes as unprocessed; using the unprocessed node with the smallest current cumulative path cost as the central reference node, traversing each adjacent unprocessed node directly connected to the central reference node, and calculating the current cumulative path cost of the central reference node and the cost of the nodes connected to the corresponding adjacent nodes. The weighted cost of undirected data edges of unprocessed nodes is summed. When the summation is less than the cumulative path cost of the currently recorded adjacent unprocessed node, the cumulative path cost of the adjacent unprocessed node is updated with the summation, and the corresponding predecessor node mapping relationship is written into the predecessor relationship record memory. After all adjacent edges of the central reference node are calculated, the central reference node is marked as processed, and the node with the smallest cumulative path cost among the remaining unprocessed nodes is reselected as the central reference node for the next iteration round, until all the initial measurement space coordinates are marked as processed. The predecessor relationship record in the predecessor relationship record memory is read, and a reverse backtracking concatenation is performed from the traversal endpoint to the starting point to construct the optimized spatial scanning trajectory with the minimum global expected robotic arm motion time cost and without duplicate redundant coverage scanning nodes.

[0025] In the specific processing, a graph structure model is first established in the system's operating memory, denoted as: ; Where V represents the set of nodes corresponding to all initial measurement space coordinates, and E represents the set of edges between nodes. For any edge connecting node u and node v, a weight w(u,v) is assigned, which represents the time cost of moving from the current node to the adjacent node when the robot arm's detection speed is determined. To enable the graph theory optimization algorithm to perform a global solution, the cumulative path cost d[v] from the starting node to each node needs to be defined, as well as the predecessor node p[v] used to record the trajectory connection relationships. During initialization, the cumulative path cost of the starting node is set to zero, and the cumulative path cost of the remaining nodes is set to infinity, i.e.: ; In this initialization phase, all nodes are marked as unprocessed, ensuring that every node is on the same comparison benchmark before entering the solution process. This initialization method corresponds to the standard starting condition for shortest path graph theory optimization algorithms, and its purpose is to provide a unified state basis for the subsequent gradual relaxation of edge costs.

[0026] During the iteration process, the node with the smallest current cumulative path cost is selected from the unprocessed nodes as the central reference node at each step. After selection, the adjacent unprocessed nodes directly connected to the central reference node are traversed, and the new cumulative path cost to the adjacent node via the current node is calculated for each. For any adjacent node v, the distance update rule is as follows: ; Where u represents the current central reference node, d[u] represents the cumulative path cost from the starting node to the current node, and w(u,v) represents the edge cost of moving from the current node to an adjacent node. When the new path cost obtained through the current node is less than the original recorded value of the adjacent node, it is considered that a better connection method has been found. At this time, the new cumulative path cost is written to the adjacent node, and the current node is simultaneously recorded as the predecessor node of the adjacent node. The result of this processing is that each node will continuously retain the currently known optimal arrival path information throughout the solution process, while the predecessor relationship record storage stores the connection links required for global trajectory backtracking. After the calculation of all adjacent edges of the current node is completed, the current node is marked as processed. Then, the node with the smallest cumulative path cost is selected from the remaining unprocessed nodes to continue the same operation until all initial measurement space coordinates are processed. This process adopts a greedy solution method that expands node by node, retaining the local optimal path at each step and gradually approaching the overall optimal solution.

[0027] After all nodes have been processed, the predecessor relationship record memory has saved the connection relationships of each node under the optimal arrival conditions. At this point, using the traversal endpoint as the backtracking starting point, the discrete nodes can be connected into a continuous scanning path by tracing back to the starting node level by level along the predecessor relationship corresponding to each node. Since this path is based on the minimum cumulative motion time cost, and overlapping measurement areas have been screened and eliminated before trajectory generation, the generated spatial scanning trajectory has the characteristics of being continuous, having no repeated coverage, and having a reasonable movement order. The optimized spatial scanning trajectory obtained in this way can reduce the number of times the robotic arm repeatedly passes through the same scene, shorten the total detection time, and ensure that each effective detection area of ​​the curved surface is scanned in a more compact order. For smooth curved surface products, multiple candidate paths selected in the early stage based solely on the principle of minimum reflection may still lead to repeated shooting and overlapping measurement areas. However, after eliminating overlapping areas and combining the shortest path solution, the movement path of the robotic arm is further compressed, and the trajectory execution efficiency is significantly improved, thus providing a continuous, stable, and more efficient motion foundation for subsequent 3D depth data acquisition.

[0028] In one implementation, the initial measured spatial coordinates can first be filtered through regions with low reflection, and then regions with overlapping coverage relationships can be deleted to reduce the number of redundant nodes entering the graph structure model. After this processing, the graph theory optimization algorithm faces a set of candidate nodes that already meet the basic measurement validity requirements, and the subsequent solution focuses on the global optimization of the node access order. Through this processing method, not only is the transformation from candidate measured coordinates to the actual scanning trajectory completed, but the requirements of reflection suppression, region deduplication, and execution time compression are also taken into account during the trajectory generation process, enabling the optimized spatial scanning trajectory to meet the dual requirements of continuity and efficiency for subsequent conformal surface scanning.

[0029] In step S103, a multi-axis industrial robotic arm equipped with a laser displacement sensor is driven, and the laser displacement sensor is controlled to perform uniform conformal motion along the optimized spatial scanning trajectory at a preset detection speed while maintaining a perpendicular incident relationship with the smooth surface. During the motion, three-dimensional depth topography data point cloud representing the physical depth of the smooth surface is collected.

[0030] In this system, a multi-axis industrial robotic arm serves as the motion execution unit. A laser displacement sensor is mounted at the end of the robotic arm. Leveraging the robotic arm's multi-degree-of-freedom posture adjustment capabilities, the laser displacement sensor can continuously scan the smooth curved surface of a silicone rubber product along an optimized spatial scanning trajectory. Because the smooth surface exhibits both curvature variations and strong reflectivity, using a laser displacement sensor with a fixed posture alone can easily introduce noise due to diffuse and local reflection interference, affecting the stability of depth measurement results. To mitigate this impact, the laser displacement sensor is kept perpendicular to the local area of ​​the curved surface during the measurement process. This means the measurement optical path is aligned as closely as possible with the local normal direction of the curved surface, thereby reducing the disturbance of reflected light to the sampling signal and improving the effectiveness and consistency of the surface measurement. After receiving the optimized spatial scanning trajectory obtained in the previous step, the robotic arm sequentially passes through each measurement spatial coordinate, continuously adjusting its end-effector posture during movement to allow the laser displacement sensor to perform conformal scanning of the curved surface, thus achieving continuous coverage of the sides and smooth transition areas.

[0031] In terms of motion control, the preset detection speed is matched with the sampling capability of the laser displacement sensor. The amount of data that the sensor can collect per unit time is basically fixed. Therefore, if the detection speed is too fast, the measured shape will be compressed in the result; if the detection speed is too slow, the measured shape will be expanded. Only by controlling the movement speed of the robotic arm within the range corresponding to the sensor sampling frequency can we ensure that the surface topography data does not become significantly distorted.

[0032] During the data acquisition process, the laser displacement sensor continuously outputs depth measurement results corresponding to changes in surface height. The robotic arm's spatial motion trajectory provides a planar position reference, and the sensor's echo measurement results provide height information for the corresponding position. The combination of these two elements forms a three-dimensional depth topography point cloud characterizing the physical depth of the smooth surface. This three-dimensional depth topography point cloud can reflect local depressions, indentations, scratches, and other minute undulations on the surface, providing the raw data foundation for subsequent depth matrix construction, gap compensation, image representation, and fusion with two-dimensional multi-illumination image data. Because the optimized spatial scanning trajectory eliminates overlapping measurement areas, the robotic arm can reduce repeated traversal of the same scene during scanning, thereby ensuring coverage integrity while reducing unnecessary movement and improving the continuity and efficiency of surface depth data acquisition.

[0033] In step S104, a vision camera system with a fixed setting and equipped with a multi-directional adjustable closed light source is synchronously controlled to continuously expose and image the smooth area of ​​the silicone rubber product under different illumination angles and different light source spatial projection vector states, thereby acquiring a two-dimensional multi-illumination image data array.

[0034] The vision camera system is used to acquire two-dimensional appearance information of smooth areas on silicone rubber products. Because these products have significant reflective properties, glare, localized overexposure, and the obscuring of defect boundaries are easily generated under ordinary fixed lighting conditions. Therefore, the vision camera system works in conjunction with a multi-directional adjustable enclosed light source. This controlled lighting environment reduces the influence of stray light and improves the consistency of imaging in smooth areas. The enclosed light source creates a relatively stable illumination field within which the vision camera system performs continuous exposure acquisition on the workpiece. This allows the surface brightness distribution, subtle texture changes, and localized shadow changes under different lighting directions to be displayed separately, thus providing a two-dimensional image basis for subsequent defect identification.

[0035] When performing continuous exposure imaging on a smooth area of ​​the silicone rubber product to obtain the two-dimensional multi-illumination image data array, the specific steps include: In the first stage, the multi-directional adjustable enclosed light source is controlled to enter the diffuse coverage projection mode, and the vision camera system is continuously triggered to perform multiple exposure acquisitions according to a preset time interval, so as to suppress the halo overflow distortion caused by the bright area of ​​silicone rubber through time-series integration superposition; wherein, the continuous exposure imaging can be carried out in stages according to different illumination modes.

[0036] In diffuse overlay projection mode, the light source projects light in multiple directions, primarily to acquire the overall appearance contour of smooth areas and reduce the strong reflection concentration phenomenon caused by unidirectional illumination. The visual camera system can continuously perform multiple exposure acquisitions within a short time interval. The image information acquired in succession forms a superposition relationship in time, thereby reducing the halo spillover distortion caused by bright areas and making the large-scale appearance features of smooth areas more stable.

[0037] In the second stage, the emission projection vector of the multi-directional adjustable closed light source is switched to a position with a first preset oblique angle relative to the vertical reference axis of the vision camera system. Continuous exposure imaging is then performed at the same acquisition rate as in the first stage to utilize the minute shadow gradients formed under tangential illumination to evaluate the microscopic uneven texture distribution of the smooth region in the horizontal direction. Under tangential illumination, the light source projection direction forms an oblique angle relative to the vertical reference direction of the vision camera system. In this case, the light acts obliquely on the surface of the smooth region, and the minute surface unevenness forms a weak but continuous shadow gradient in local areas, thus making it more effective in revealing subtle texture undulations in the horizontal direction. By performing continuous imaging at the same exposure rate as the previous stage, the response differences of different locations on the surface to tangential illumination can be acquired, making minute indentations, shallow depressions, or weak texture disturbances easier to distinguish in the image grayscale.

[0038] In the third stage, the emission projection vector of the multi-directional adjustable closed light source is deflected to a position with a second preset oblique angle relative to the horizontal reference surface of the vision camera system. Continuous exposure imaging is then performed according to the acquisition parameters corresponding to the second stage to analyze the surface topological undulations of the smooth region in the vertical direction. Under oblique illumination in another direction, the light source projection direction forms an angle relative to the horizontal reference direction of the vision camera system, used to reveal surface topological changes in a different direction than in the previous stage. Since the reflection and shadow manifestations of minor defects in the smooth region are not the same in different directions, relying solely on a single oblique illumination can easily miss local fine features. Therefore, acquiring continuous exposure images from another direction helps to supplement the undulation information in the vertical direction, allowing scratch directions, local depression boundaries, and surface abnormalities to be further revealed from another angle.

[0039] In the data aggregation stage, the image arrays acquired in the first, second, and third stages undergo spatial registration and fusion processing to form the two-dimensional multi-illumination image data array that suppresses ambient background light diffusion interference and material specular reflection interference. Specifically, in this stage, the image arrays acquired under the diffuse overlay projection mode and the image arrays acquired in the two oblique illumination stages undergo spatial registration and fusion processing, ensuring that image information of the same smooth area under different illumination conditions corresponds to a unified pixel space. After registration, the brightness variations, shadow gradients, and detailed textures formed by illumination from different directions can be superimposed into the same image representation system, thus forming the two-dimensional multi-illumination image data array. This two-dimensional multi-illumination image data array simultaneously retains overall appearance information and local directional texture information, and can suppress ambient background light diffusion interference and material specular reflection interference to a certain extent, improving the identifiability of minor defects in smooth areas.

[0040] In step S105, cross-modal fusion at the feature level is performed on the acquired three-dimensional depth topography point cloud data and the two-dimensional multi-light image data array to construct a unified format dataset.

[0041] Before performing the cross-modal fusion, the three-dimensional depth topography point cloud is processed as follows: A spatial two-dimensional mapping structured data matrix is ​​constructed based on the continuous motion trajectory of the laser displacement sensor over time and the depth measurement feedback parameters obtained from reflected echo analysis. This spatial two-dimensional mapping structured data matrix uses the horizontal motion axis as the first dimension (x-coordinate) and the vertical motion axis as the second dimension (y-coordinate), with the depth values ​​measured by the corresponding grid cells used as the matrix height element. For missing regions in the spatial two-dimensional mapping structured data matrix caused by mechanical vibration or sampling discontinuities, interpolation calculations are performed using surrounding valid values ​​to smooth and compensate for the missing depth values, converting discrete collected points into a topologically continuous three-dimensional geometric entity data structure. The spatial two-dimensional mapping structured data matrix after smoothing and compensation fitting is input into a three-dimensional graphics rendering engine to generate a three-dimensional surface visualization simulation image with visual brightness and darkness feedback attributes. This three-dimensional surface visualization simulation image is used as the image representation result of the three-dimensional depth topography point cloud for cross-modal fusion with the two-dimensional multi-illumination image data array.

[0042] Before entering the unified data construction stage, the 3D depth topography point cloud and the 2D multi-illumination image data array undergo structuring processing. This transforms the surface measurement results from discrete sampling points into a regular data format that facilitates image representation and subsequent fusion. The laser displacement sensor outputs height information continuously acquired along the robotic arm's trajectory during scanning. Therefore, by combining the robotic arm's motion path and depth measurement feedback parameters during scanning, a spatial 2D mapping structured data matrix can be established, indexed by planar position and containing height values. The horizontal direction is defined as the transverse coordinate, the vertical direction as the longitudinal coordinate, and the depth measurement result at each position serves as the height value of the corresponding unit.

[0043] During matrix construction, due to mechanical micro-vibrations, local reflection instability, or sampling discontinuities, some regions may have missing cells with unmeasured data. To ensure the continuity of the surface topography representation, interpolation is used to compensate for these missing regions, restoring the discrete sampling results to a continuous surface. Linear interpolation or spline interpolation can be used to smoothly fit the surrounding areas with existing valid height values, allowing the height information of the missing regions to be completed according to the neighborhood variation trend. After this processing, the originally scattered depth sampling points are organized into a continuous three-dimensional surface data structure, which can more completely reflect the geometric undulation characteristics of the smooth surface and reduce the impact of data fragmentation on subsequent fusion.

[0044] The spatial two-dimensional mapped structured data matrix, after interpolation compensation, is input into a graphics rendering environment to generate a three-dimensional surface visualization simulation image that can be directly used for visual processing. During rendering, three-dimensional graphics software or visualization tools can be used to display the height distribution in the matrix in stereo, and to present the local undulations of the surface through changes in brightness and darkness, thus transforming the three-dimensional depth topography point cloud into a visually expressive image. After this processing, the three-dimensional depth topography point cloud no longer exists merely as a set of geometric coordinates, but forms surface image data with a similar expressive structure to two-dimensional images, facilitating unified processing with the two-dimensional multi-illumination image data array acquired by the vision camera system.

[0045] After completing the 3D data visualization representation, the 3D surface visualization simulation image is used as a depth information carrier and enters the cross-modal fusion stage together with the 2D multi-illumination image data array. The 2D multi-illumination image data array mainly reflects the appearance texture, brightness changes, and local shadow information of smooth areas under different controlled lighting conditions, while the 3D surface visualization simulation image mainly reflects the spatial undulation and depth change information of curved areas. The two are complementary in terms of information. By organizing these two types of images accordingly, a sample can simultaneously contain appearance features and morphological features, thereby constructing a unified format dataset. The samples in this unified format dataset can be further divided into normal, dent, and scratch categories according to defect status. Data from the vision camera system and laser displacement sensor are combined, so that the 2D appearance data and 3D morphological data of the same category are correspondingly merged into the same data organization system, so as to extract joint features and perform defect identification in the future.

[0046] In one implementation, the visual image and laser measurement image corresponding to the same product sample can be merged during dataset construction, so that each sample contains both appearance information of the frontal view area and depth information of the curved surface area. The resulting unified format dataset not only reduces the problem of insufficient representation of high reflectivity and smooth curved surface defects by a single modality, but also gives defects such as dents and scratches a more complete representation at the data level, providing stable input for high-dimensional feature extraction in the subsequent recognition stage.

[0047] In step S106, the unified format dataset is input into a pre-established deep convolutional neural network model for forward propagation and high-dimensional feature extraction. Through nonlinear mapping of cascaded layers within the network, the corresponding defect classification probability determination result is obtained at the prediction output.

[0048] The unified format dataset serves as input data for a deep convolutional neural network model, entering a pre-established classification and recognition process. This model jointly learns 2D appearance information acquired by a visual camera system and 3D shape representation information acquired by a laser displacement sensor. Through multi-layer feature extraction and nonlinear mapping, it outputs the probability results corresponding to the defect categories. During the training phase, the unified format dataset can be divided into training, validation, and test sets according to a predetermined ratio for parameter learning, training process monitoring, and classification performance verification. Data categories can be organized as healthy samples, dented samples, and scratched samples, allowing the model to establish classification boundaries around actual quality assessment targets. For smooth curved surfaces and highly reflective surfaces, individual samples are prone to fluctuations in illumination, scale, and local brightness. Therefore, data augmentation is performed on the training data during the model building phase to improve the model's adaptability to changes in actual working conditions and reduce the risk of overfitting.

[0049] The deep convolutional neural network model, when pre-built, performs data augmentation processing on the unified format dataset used for training. This data augmentation processing includes brightness offset adjustment, scale transformation, contrast stretching transformation, and channel gain scaling. Specifically: for brightness offset adjustment, a preset deviation value is superimposed on each channel element of the original pixel matrix of the image samples in the unified format dataset to simulate background illumination changes; for scale transformation, the spatial distribution of the original feature point coordinates of the image samples is adjusted according to a preset scaling factor to reconstruct the image boundary resolution; for contrast stretching transformation, the statistical mean of the pixel data of the entire image is calculated, and the difference between bright and dark pixels is scaled proportionally based on this statistical mean to enhance the grayscale transition features of minute scratch boundaries; for channel gain scaling, a predefined multiplier is applied to the feature pixels in each channel dimension within the data matrix of the unified format dataset to change the response intensity corresponding to the local illumination environment.

[0050] Data augmentation is used to increase the variety of training samples, making the model less reliant on overly simplistic surface patterns. Brightness shift adjustment simulates changes in background illumination by altering the overall brightness of the image; its expression is: ; in, Indicates position The original pixel value at that location, This represents the brightness adjustment amount. Scale transformation, by altering the spatial sampling relationships of the image, allows the model to adapt to different viewing distances and changes in local region size; its expression is: ; in, This represents the scaling factor. Contrast stretching transformation is used to enhance the difference between light and dark areas, making the boundaries of small scratches more prominent; its expression is: ; in, This represents the statistical mean of image pixels. This represents the contrast scaling factor. Channel gain scaling simulates changes in local illumination intensity by applying a multiplier factor to pixel values; its expression is: ; in, This represents the pixel scaling factor. Through the above processing, the training samples exhibit richer variations in brightness, contrast, and response intensity.

[0051] The process of inputting the unified format dataset into a pre-established deep convolutional neural network model for forward propagation and high-dimensional feature extraction, and obtaining the corresponding defect classification probability judgment result at the prediction output, specifically includes: performing pixel-level normalization processing on the multi-channel composite data in the input unified format dataset, linearly scaling the original pixel values ​​according to the full-scale grayscale of the sensor, so that the feature elements of each pixel converge to a unified numerical range to eliminate the dimensional differences between different sensor channels; inputting the normalized data into the input layer of the deep convolutional neural network model, and performing convolution operations through a group of convolutional layers set in a deep series to extract the edge, texture, and high-order shape features of the image; setting pooling layers at intervals along the forward propagation direction of the deep convolutional neural network model, and performing local region aggregation downsampling on the feature map output by the convolution to reduce the feature dimension and enhance the robustness of the features to local displacement perturbations; in the back-end dimensionality reduction stage, converting the multi-dimensional feature tensor after convolution and pooling operations into a one-dimensional feature vector through an unfolding layer as the input for subsequent classification judgment.

[0052] After converting the multidimensional feature tensor into a one-dimensional feature vector, the one-dimensional feature vector is input as a whole into the tail fully connected layer decision module. Each neuron in the fully connected layer decision module receives a weighted signal from the output of all related neurons in the previous layer, and generates a node output by superimposing a bias term. Nonlinear activation function control units are connected in series at the output of the fully connected layer decision module and the output of each preceding convolutional layer. The nonlinear activation function control unit located in the intermediate hidden layer uses a modified linear activation function to retain positive responses and suppress negative responses. The nonlinear activation function control unit located in the final prediction output layer uses a normalized probability distribution function to convert the one-dimensional feature vector into a posterior confidence probability vector corresponding to each defect category, and outputs the final defect classification result according to the principle of maximizing probability.

[0053] After inputting the deep convolutional neural network model with a uniformly formatted dataset, pixel-level normalization is performed first. Normalization maps the raw pixel values ​​of different channels to a uniform numerical range, making training more stable and reducing the impact of different input units on model learning. For image input, the normalization expression is: ; in, Represents the original pixel value. This represents the normalized pixel value, where x and y represent the pixel position, and c represents the channel index. The normalized multi-channel tensor is fed into the input layer to provide standardized input for subsequent convolution operations.

[0054] Convolutional layers are used to perform weight-sharing computations within local receptive fields, extracting edge, corner, texture, and higher-level shape features from the input tensor. The convolution operation can be represented as: ; in, Let represent the output value of the k-th feature map at position (i,j), where U and V represent the size of the convolution kernel, and C represents the number of input channels. Indicates the convolution kernel weights, This represents the bias term. Convolutional layers can use multiple convolutional kernels stacked layer by layer, allowing the model to learn higher-level semantic structures progressively from lower-level local structures. For custom convolutional networks, a structure consisting of multiple convolutional layers, pooling layers, and fully connected layers can be used. The convolutional kernel size can be 3x3, the pooling window can be 2x2, and the input image resolution can be set according to computational resources and recognition accuracy requirements.

[0055] After the convolutional output enters the pooling layer, the spatial size of the feature map is compressed, while the main response features are preserved. The output size of the pooling layer satisfies: ; in, Indicates the output size. Indicates the input dimension. Let P represent the pooling window size, P represent the padding parameter, and S represent the stride. Max pooling can be expressed as: ; This processing can retain the most significant response in the local region while reducing the feature dimension, which helps to enhance the robustness of the model to local displacement perturbations and reduce the subsequent computational burden.

[0056] The multidimensional feature tensors, after processing by convolutional and pooling layers, enter the unrolling stage and are reconstructed into one-dimensional feature vectors, enabling spatial features to be passed to the backend classification part. The unrolling layer performs the structural transformation between convolutional features and classification decisions, mapping and organizing the multidimensional features into linear vectors before inputting them into the fully connected layer. Each neuron in the fully connected layer receives a weighted input from all neurons in the previous layer, and after adding a bias term, outputs the node response, expressed as: ; Where F represents the neuron output, Indicates connection weights. Let b represent the input feature and b represent the bias term. This represents the activation function. The role of the fully connected layer is to comprehensively judge the multi-dimensional features extracted from the front end, and to associate the features learned from different locations and modalities to form a judgment result for a specific defect category.

[0057] In the nonlinear mapping stage, the intermediate hidden layers can use a modified linear activation function, which has the following form: ; This function preserves positive valid responses and suppresses negative invalid responses, thereby improving the training stability of deep networks. The final output prediction layer uses a normalized probability distribution function, which has the following form: ; in, This represents the input response value for the corresponding category. After normalization, the output results form the posterior confidence probability vector for each defect category. Then, based on the principle of maximizing probability, the category with the highest probability value is taken as the final defect classification result for the current input sample. In this way, after forward propagation, each sample in the uniform format dataset can obtain a probability judgment result corresponding to the healthy state, dented state, or scratch state.

[0058] In one implementation, the deep convolutional neural network model can employ traditional convolutional networks, visual geometric group networks, or residual network structures. Traditional convolutional networks can be configured with different numbers of convolutional layers and input resolutions to achieve defect classification with lower computational burden. Visual geometric group networks enhance local pattern learning capabilities through the stacking of consecutive small convolutional kernels. Residual networks improve the training stability of deep networks through residual connections and bottleneck structures, enabling deeper feature representation. For smooth curved surfaces and highly reflective surfaces, higher input resolution is generally beneficial for improving classification performance, while a lighter network structure helps control computational resource consumption; therefore, the appropriate network configuration can be selected based on detection accuracy and deployment environment requirements.

[0059] Through the above processing, the 3D surface visualization simulation image and the 2D multi-light image data are combined into a unified format dataset as joint input. During the training phase, the network learns the defect feature distribution under different modalities, and during the inference phase, it outputs the defect classification probability judgment result for the test sample. This preserves the 2D appearance texture information while introducing 3D morphological change information, which helps to improve the ability to identify small dents and shallow scratches on smooth curved silicone rubber products.

[0060] Based on the same line of thought, such as Figure 2 As shown, a machine vision-based inspection system for silicone rubber products is provided, comprising: The coordinate generation module 201 is used to obtain a three-dimensional theoretical digital model of the silicone rubber product, perform surface meshing analysis on the smooth surface of the three-dimensional theoretical digital model, extract the surface normal vector of each discrete mesh point, and offset the surface normal vector along the outward normal direction by a preset recognition distance, and calculate multiple initial measurement space coordinates that maintain an orthogonal attitude with the smooth surface. The scanning trajectory optimization module 202 is used to perform global topological reorganization and redundant region removal on all the initial measurement spatial coordinates using the shortest path graph theory optimization algorithm to generate a continuous optimized spatial scanning trajectory that does not overlap or reuse in space. The three-dimensional depth data acquisition module 203 is used to drive a multi-axis industrial robotic arm equipped with a laser displacement sensor, and control the laser displacement sensor to perform uniform conformal motion along the optimized spatial scanning trajectory at a preset detection speed under the condition of maintaining a perpendicular incident relationship with the smooth surface, and to acquire three-dimensional depth topography data point cloud representing the physical depth of the smooth surface during the motion. The two-dimensional multi-light image acquisition module 204 is used to synchronously control a fixed visual camera system equipped with a multi-directional adjustable closed light source, and to continuously expose and image the smooth area of ​​the silicone rubber product under different lighting angles and different light source spatial projection vector states to acquire a two-dimensional multi-light image data array. The cross-modal fusion data construction module 205 is used to perform feature-level cross-modal fusion on the acquired three-dimensional depth topography data point cloud and the two-dimensional multi-light image data array to construct a unified format dataset. The defect identification module 206 is used to input the unified format dataset into a pre-established deep convolutional neural network model for forward propagation and high-dimensional feature extraction, and obtain the corresponding defect classification probability judgment result at the prediction output end through nonlinear mapping of cascaded layers within the network.

[0061] This system, by taking advantage of the characteristics of smooth curved surfaces in silicone rubber products, comprehensively utilizes surface morphology information and appearance image information under multiple lighting conditions. This can, to a certain extent, reduce the interference caused by highly reflective surfaces on defect detection and improve the characterization ability of minute defects in curved areas. Optimizing the detection trajectory helps reduce the time overhead caused by repeated scanning, improving the continuity and efficiency of the detection process. By uniformly organizing data from different sources and inputting it into a deep convolutional neural network model for feature extraction and defect classification, it enhances the automatic identification capability of defect categories such as dents and scratches, improving the accuracy and stability of the detection results.

[0062] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to exemplary embodiments of this disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0063] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

[0064] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A machine vision-based method for inspecting silicone rubber products, characterized in that, The method includes: A three-dimensional theoretical digital model of a silicone rubber product is obtained. The smooth surface of the three-dimensional theoretical digital model is analyzed by surface meshing. The surface normal vectors of each discrete mesh point are extracted. Based on the surface normal vectors, a preset identification distance is offset along the outer normal direction. Multiple initial measurement space coordinates that maintain an orthogonal attitude with the smooth surface are calculated. The shortest path graph theory optimization algorithm is used to perform global topology reorganization and redundant region removal on all the initial measurement spatial coordinates to generate a continuous optimized spatial scanning trajectory that does not overlap or reuse in space. A multi-axis industrial robotic arm equipped with a laser displacement sensor is driven to control the laser displacement sensor to perform uniform conformal motion along the optimized spatial scanning trajectory at a preset detection speed while maintaining a perpendicular incident relationship with the smooth curved surface. During the motion, three-dimensional depth topography data point cloud representing the physical depth of the smooth curved surface is collected. A vision camera system with synchronous control and fixed settings and equipped with a multi-directional adjustable closed light source continuously exposes and images the smooth area of ​​the silicone rubber product under different illumination angles and different light source spatial projection vector states to obtain a two-dimensional multi-illumination image data array. Perform cross-modal fusion at the feature level on the acquired 3D depth topography point cloud and the 2D multi-light image data array to construct a unified format dataset; The unified format dataset is input into a pre-established deep convolutional neural network model for forward propagation and high-dimensional feature extraction. Through nonlinear mapping of cascaded layers within the network, the corresponding defect classification probability determination result is obtained at the prediction output.

2. The machine vision-based inspection method for silicone rubber products according to claim 1, characterized in that, The generation of the optimized spatial scan trajectory specifically includes: A graph structure model is constructed in the system's computing memory, with all the initial measurement space coordinates as discrete topological nodes, the actual physical movement paths between each node as undirected data edges, and the expected robot motion time cost corresponding to each undirected data edge at the preset detection speed as a weighted cost. Initialize the global state matrix of the graph theory algorithm, specify the traversal starting point, set the cumulative path cost of the starting point to zero, initialize the cumulative path cost of the remaining nodes to the maximum value, and mark all the discrete topological nodes as unprocessed. The unprocessed node with the lowest current cumulative path cost is taken as the central reference node. Each adjacent unprocessed node directly connected to the central reference node is traversed, and the sum of the current cumulative path cost of the central reference node and the weighted cost of the undirected data edge connected to the corresponding adjacent unprocessed node is calculated. When the calculated summation value is less than the cumulative path cost currently recorded by the adjacent unprocessed node, the cumulative path cost of the adjacent unprocessed node is updated with the summation value, and the corresponding predecessor node mapping relationship is written into the predecessor relationship record memory. After all the adjacent edges of the central reference node have been calculated, the central reference node is marked as processed. Then, among the remaining unprocessed nodes, the node with the smallest cumulative path cost is selected as the central reference node for the next iteration, until all the initial measurement space coordinates are marked as processed. Read the predecessor relationship records in the predecessor relationship record memory, perform reverse backtracking from the end point to the starting point, and construct the optimized spatial scanning trajectory with the minimum expected robot motion time cost and elimination of repeated redundant coverage scanning nodes.

3. The machine vision-based inspection method for silicone rubber products according to claim 1, characterized in that, Before performing the cross-modal fusion, the 3D depth topography point cloud data is processed as follows: Based on the continuous motion trajectory of the laser displacement sensor in the time series and the depth measurement feedback parameters obtained from the reflected echo analysis, a spatial two-dimensional mapping structured data matrix is ​​constructed; wherein, the spatial two-dimensional mapping structured data matrix uses the horizontal motion axis as the first dimension x-coordinate, the vertical motion axis as the second dimension y-coordinate, and the depth value measured by the corresponding grid cell as the matrix height element; For the missing regions in the spatial two-dimensional mapped structured data matrix caused by mechanical vibration or sampling discontinuity, interpolation calculations are performed using the surrounding effective values ​​to smoothly compensate and fit the missing depth values, so as to convert the discrete collected scattered points into a topologically continuous three-dimensional geometric entity data structure. The spatial two-dimensional mapping structured data matrix, after smoothing compensation fitting, is input into the three-dimensional graphics rendering engine to generate a three-dimensional surface visualization simulation image with visual brightness and darkness feedback attributes. The three-dimensional surface visualization simulation image is used as the image representation result of the three-dimensional depth topography data point cloud, and is used to perform the cross-modal fusion with the two-dimensional multi-light image data array.

4. The machine vision-based silicone rubber article inspection method of claim 1, wherein, When the deep convolutional neural network model is pre-built, it performs data augmentation processing on the uniform format dataset used for training. This data augmentation processing includes brightness shift adjustment, scale scaling transformation, contrast stretching transformation, and channel gain scaling, wherein: For the brightness offset adjustment, a preset deviation value is superimposed on each channel element of the original pixel matrix of the image samples in the unified format dataset to simulate background illumination changes; For the scaling transformation, the spatial distribution relationship of the original feature point coordinates of the image samples is adjusted according to a preset scaling factor in order to reconstruct the image boundary resolution; For the contrast stretching transformation, the statistical mean of the pixel data of the entire image is calculated, and the difference between bright and dark pixels is scaled proportionally based on the statistical mean to enhance the gray-scale transition characteristics of the boundary of the tiny scratch. For the channel gain scaling process, a predefined multiplier is applied to the feature pixels in each channel dimension within the data matrix of the unified format dataset to change the response intensity corresponding to the local lighting environment.

5. The machine vision-based silicone rubber article inspection method of claim 1, wherein, When performing continuous exposure imaging on a smooth area of ​​the silicone rubber product to obtain the two-dimensional multi-illumination image data array, the specific steps include: In the first stage, the multi-directional adjustable closed light source is controlled to enter the diffuse coverage projection mode, and the visual camera system is continuously triggered to perform multiple exposure acquisitions according to a preset time interval, so as to suppress the halo overflow distortion caused by the bright area of ​​silicone rubber through time-series integration superposition. In the second stage, the emission projection vector of the multi-directional adjustable closed light source is switched to a posture with a first preset oblique angle relative to the vertical reference center axis of the vision camera system, and continuous exposure imaging is performed according to the same acquisition rhythm as in the first stage, so as to evaluate the micro-concave-convex texture distribution of the smooth area in the horizontal direction by utilizing the small shadow gradient formed under tangential lighting conditions. In the third stage, the emission projection vector of the multi-directional adjustable closed light source is deflected to a position with a second preset oblique angle relative to the system horizontal reference bottom surface of the vision camera system, and continuous exposure imaging is performed according to the acquisition parameters corresponding to the second stage to analyze the surface topological undulation changes of the smooth area in the vertical direction. During the data aggregation stage, the image arrays acquired in the first, second, and third stages are subjected to spatial registration and fusion processing to form the two-dimensional multi-illumination image data array that suppresses ambient background light diffusion interference and material specular reflection interference.

6. The machine vision-based silicone rubber article inspection method of claim 1, wherein, The process of inputting the unified format dataset into a pre-established deep convolutional neural network model for forward propagation and high-dimensional feature extraction, and obtaining the corresponding defect classification probability determination result at the prediction output, specifically includes: Pixel-level normalization is performed on the multi-channel composite data in the input unified format dataset. The original pixel values ​​are linearly scaled according to the full scale of the sensor grayscale, so that the feature elements of each pixel converge to a unified value range, thereby eliminating the dimensional differences between different sensor channels. The normalized data is input into the input layer of the deep convolutional neural network model, and convolution operations are performed through a group of convolutional layers set in a deep series to extract the edge, texture and high-order shape features of the image. Pooling layers are set at intervals along the forward propagation direction of the deep convolutional neural network model, and local region aggregation downsampling is performed on the feature map of the convolution output to reduce the feature dimension and enhance the robustness of the feature to local displacement perturbations. In the backend dimensionality reduction stage, the multidimensional feature tensors after convolution and pooling are converted into one-dimensional feature vectors through an unfolding layer, which are then used as input for subsequent classification decisions.

7. The machine vision-based silicone rubber article detection method of claim 6, wherein, After converting the multidimensional feature tensor into a one-dimensional feature vector, the one-dimensional feature vector is input as a whole into the tail fully connected layer decision module. Each neuron in the fully connected layer decision module receives a weighted signal from the outputs of all associated neurons in the previous layer, and generates a node output by superimposing a bias term. A nonlinear activation function control unit is connected in series at the output of the fully connected layer decision module and the output of each preceding convolutional layer. The nonlinear activation function control unit located in the intermediate hidden layer employs a modified linear activation function to retain positive responses and suppress negative responses. The nonlinear activation function control unit located in the final prediction output layer uses a normalized probability distribution function to convert the one-dimensional feature vector into a posterior confidence probability vector corresponding to each defect category, and outputs the final defect classification result according to the principle of maximizing probability.

8. A machine vision based silicone rubber article inspection system, characterized by, include: The coordinate generation module is used to obtain a three-dimensional theoretical digital model of silicone rubber products, perform surface meshing analysis on the smooth surface of the three-dimensional theoretical digital model, extract the surface normal vector of each discrete mesh point, and offset the surface normal vector along the outward normal direction by a preset recognition distance, and calculate multiple initial measurement space coordinates that are orthogonal to the smooth surface. The scanning trajectory optimization module is used to perform global topological reorganization and redundant region removal on all the initial measurement spatial coordinates using the shortest path graph theory optimization algorithm, so as to generate a continuous optimized spatial scanning trajectory that does not overlap or reuse in space. A three-dimensional depth data acquisition module is used to drive a multi-axis industrial robotic arm equipped with a laser displacement sensor. The module controls the laser displacement sensor to move at a preset speed along the optimized spatial scanning trajectory while maintaining a perpendicular incident relationship with the smooth surface. During the movement, the module acquires three-dimensional depth topography data point cloud that characterizes the physical depth of the smooth surface. A two-dimensional multi-light image acquisition module is used to synchronously control a fixed visual camera system equipped with a multi-directional adjustable closed light source, and to continuously expose and image the smooth area of ​​the silicone rubber product under different lighting angles and different light source spatial projection vector states to acquire a two-dimensional multi-light image data array. The cross-modal fusion data construction module is used to perform feature-level cross-modal fusion on the acquired 3D depth topography point cloud and the 2D multi-light image data array to construct a unified format dataset. The defect identification module is used to input the unified format dataset into a pre-established deep convolutional neural network model for forward propagation and high-dimensional feature extraction, and obtain the corresponding defect classification probability judgment result at the prediction output through nonlinear mapping of cascaded layers within the network.

Images