An image recognition-based aluminum plate surface defect detection method and system

Abstract

The application discloses an aluminum plate surface defect detection method and system based on image recognition, and relates to the technical field of image processing. The method comprises the following steps: collecting the surface of an aluminum plate online by a high-speed industrial camera according to multi-band ring light source parameters to generate an original image data set; performing three-dimensional reconstruction based on original structure point data to construct a three-dimensional digital model of the aluminum plate surface; performing transfer learning on the three-dimensional digital model based on original surface pixel data, performing multi-scale defect detection according to a learning result, and generating a defect feature vector set; and feeding back process control instructions of the aluminum plate surface to a remote terminal after the defect feature vector set is traced back to the aluminum plate surface for process optimization. The technical problems of insufficient aluminum plate surface defect detection precision and difficulty in comprehensively and accurately identifying defects in the prior art are solved, and the technical effects of improving the accuracy and comprehensiveness of aluminum plate surface defect detection are achieved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and specifically to a method and system for detecting surface defects in aluminum plates based on image recognition. Background Technology

[0002] Aluminum sheets are crucial metallic materials in aerospace, rail transportation, building curtain walls, and automobile manufacturing, and their surface quality directly impacts subsequent processing and the performance of the finished product. During aluminum sheet production and processing, factors such as raw materials, processes, and equipment can easily lead to various defects on the surface, including scratches, pits, indentations, oxide spots, and inclusions. Current methods for detecting surface defects in aluminum sheets primarily rely on manual visual inspection or two-dimensional imaging with a single light source. These methods suffer from insufficient accuracy, poor ability to identify subtle and hidden defects, low efficiency, and significant susceptibility to lighting conditions and human experience in determining the results. Summary of the Invention

[0003] This application provides a method and system for detecting surface defects on aluminum plates based on image recognition, which solves the technical problems of insufficient detection accuracy and difficulty in comprehensively and accurately identifying defects on aluminum plates in the prior art.

[0004] The first aspect of this application provides a method for detecting surface defects in aluminum plates based on image recognition, the method comprising:

[0005] A high-speed industrial camera is activated according to the parameters of a multi-band ring light source to acquire data online on the surface of an aluminum plate, generating an original image dataset. The original image dataset includes original structural point data and original surface pixel data. Based on the original structural point data, a three-dimensional reconstruction is performed to construct a three-dimensional digital model of the aluminum plate surface. Based on the original surface pixel data, the three-dimensional digital model of the aluminum plate surface is subjected to transfer learning. Based on the learning results, multi-scale defect detection is performed to generate a defect feature vector set. The defect feature vector set is then used to optimize the process on the aluminum plate surface, generating process control instructions for the aluminum plate surface and feeding them back to a remote terminal.

[0006] A second aspect of this application provides an image recognition-based aluminum plate surface defect detection system, the system comprising:

[0007] Image acquisition module: Activates a high-speed industrial camera according to multi-band ring light source parameters to acquire images of the aluminum plate surface online, generating an original image dataset containing original structural point data and original surface pixel data; 3D reconstruction module: Performs 3D reconstruction based on the original structural point data to construct a 3D digital model of the aluminum plate surface; Defect detection module: Performs transfer learning on the 3D digital model of the aluminum plate surface based on the original surface pixel data, performs multi-scale defect detection based on the learning results, and generates a defect feature vector set; Optimization control module: Backtracks the defect feature vector set to the aluminum plate surface for process optimization, generates process control instructions for the aluminum plate surface, and feeds them back to the remote terminal.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0009] First, a high-speed industrial camera is activated according to the parameters of a multi-band ring light source to acquire data online on the aluminum plate surface, generating a raw image dataset containing raw structural point data and raw surface pixel data. Next, 3D reconstruction is performed based on the raw structural point data to construct a 3D digital model of the aluminum plate surface. Then, transfer learning is performed on the 3D digital model based on the raw surface pixel data, and multi-scale defect detection is performed based on the learning results, generating a defect feature vector set. Finally, the defect feature vector set is used to optimize the aluminum plate surface process, generating process control instructions for the aluminum plate surface and feeding them back to a remote terminal. This solves the technical problems of insufficient accuracy in aluminum plate surface defect detection and difficulty in comprehensively and accurately identifying defects in existing technologies, achieving the technical effect of improving the accuracy and comprehensiveness of aluminum plate surface defect detection. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A schematic flowchart of an image recognition-based method for detecting surface defects in aluminum plates is provided in an embodiment of this application.

[0012] Figure 2 This is a schematic diagram of an image recognition-based aluminum plate surface defect detection system provided in an embodiment of this application.

[0013] Figure labeling: Image acquisition module 11, 3D reconstruction module 12, defect detection module 13, optimization control module 14. Detailed Implementation

[0014] This application provides an image recognition-based method and system for detecting surface defects in aluminum plates, which solves the technical problems of insufficient detection accuracy and difficulty in comprehensively and accurately identifying defects in the prior art.

[0015] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0016] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0017] Example 1, as Figure 1 As shown, this application provides a method for detecting surface defects in aluminum plates based on image recognition, wherein the method includes:

[0018] The high-speed industrial camera is activated according to the parameters of the multi-band ring light source to collect data on the surface of the aluminum plate online, generating an original image dataset. The original image dataset includes original structural point data and original surface pixel data.

[0019] On the aluminum plate production line, the parameters of the multi-band ring light source (including infrared light source intensity data and visible light source color temperature threshold) are input to the light source controller. The light source controller then adjusts the brightness and switches the bands of each light source module of the ring light source to uniformly illuminate the aluminum plate surface under different wavelength illumination conditions, thereby enhancing the imaging contrast of different types of defects.

[0020] After the light source illumination stabilizes, a high-speed industrial camera (frame rate ≥ 500fps, resolution not less than 5 megapixels) is synchronously triggered with the production line conveyor system. Continuous image acquisition is performed at set sampling intervals via the external trigger interface of the industrial camera, and the acquired image data is cached in real time to the acquisition control unit. During the acquisition process, the system synchronously records the timestamp, camera attitude parameters, and light source parameters for each frame. Using a dual-channel data splitting method, image data containing geometric topography information is extracted as raw structural point data, and image data containing surface texture and color information is extracted as raw surface pixel data, thereby generating a raw image dataset with three-dimensional structure and multispectral texture information.

[0021] Furthermore, the process of constructing parameters for a multi-band ring light source includes the following methods:

[0022] Based on the material data of the aluminum plate, light reflection analysis is performed to obtain the target spectral reflectance curve; the infrared band absorption analysis is performed by traversing the target spectral reflectance curve to set the infrared light source intensity data; the visible light band reflection analysis is performed by traversing the target spectral reflectance curve to set the visible light source color temperature threshold; the light source controller is activated according to the infrared light source intensity data and the light source color temperature threshold to perform light surround adjustment and construct the multi-band ring light source parameters.

[0023] Specifically: First, the material data of the aluminum plate to be tested is acquired, including parameters such as alloy composition ratio, surface roughness, and oxide film thickness. This material data is then input into an optical simulation or actual reflection testing system to analyze the reflection characteristics of the aluminum plate surface under different wavelengths of light, obtaining a target spectral reflectance curve covering the infrared to visible light range. Second, the spectral data corresponding to the infrared band (wavelength range 0.78μm~2.5μm) in the target spectral reflectance curve is traversed, the absorptivity at each wavelength is calculated, and the optimal intensity value of the infrared light source is determined based on the signal-to-noise ratio requirements for detection, forming infrared light source intensity data. Then, the spectral data corresponding to the visible light band (wavelength range 380nm~780nm) in the target spectral reflectance curve is traversed, the reflectance distribution at different wavelengths is analyzed, and the light source color temperature threshold is determined based on the sensitivity of surface defects to specific color temperatures, forming the visible light source color temperature setting value. Finally, the infrared light source intensity data and the visible light source color temperature threshold are simultaneously input to the light source controller. The controller drives and adjusts the brightness of multiple light source modules of the ring light source according to the set parameters, realizing sequential or mixed excitation of multiple bands of infrared and visible light, and constructing multi-band ring light source parameters for aluminum plate defect detection.

[0024] Furthermore, a high-speed industrial camera is activated according to the multi-band ring light source parameters to perform online acquisition of the aluminum plate surface, generating an original image dataset. This original image dataset includes original structural point data and original surface pixel data. The method includes:

[0025] The surface of the aluminum plate is gridded, and a sampling interval threshold is set according to the gridding result. Based on the parameters of the multi-band ring light source, a high-speed industrial camera is triggered to collect images at equal intervals according to the sampling interval threshold, thereby obtaining multiple sets of images at equal intervals. Based on the multiple sets of images at equal intervals, dual-channel processing is performed to generate original structural point data and original surface pixel data.

[0026] Preferably, a two-dimensional coordinate system is established on the surface area of ​​the aluminum plate to be inspected within the detection control system. A meshing algorithm is then used to divide the surface area into meshes. The size of each mesh cell is determined based on the aluminum plate's dimensions and the minimum defect detection size; for example, the cell side length is set to 0.5 to 1 times the minimum defect size. This yields the meshed processing result used for data acquisition, and a sampling interval threshold is set accordingly. Next, the parameters of the multi-band ring light source are input to the light source controller and synchronized with the external trigger module of the high-speed industrial camera, ensuring stable illumination of the aluminum plate surface at the set wavelength, brightness, and color temperature. Subsequently, based on the sampling interval threshold, the detection control system triggers the high-speed industrial camera to acquire images at equal intervals, ensuring coverage of the entire mesh area and obtaining multiple equally spaced image sets. Each image set corresponds to a specific illumination wavelength and sampling position. Finally, the multiple equally spaced image sets are input into a dual-processing unit. One channel extracts geometric features to generate raw structural point data containing spatial coordinate information; the other channel extracts surface texture and color information to generate raw surface pixel data containing multi-band pixel information, thus forming a raw image dataset containing spatial structural information and surface spectral information.

[0027] Furthermore, the method for generating original structural point data by performing dual-path processing based on the multiple equally spaced image sets includes:

[0028] The multiple equally spaced image sets are subjected to feature analysis to generate aluminum plate surface feature parameters; the multiple equally spaced image sets are identified according to the aluminum plate surface feature parameters to determine multiple feature points; a three-dimensional coordinate system of the aluminum plate surface is constructed, and the multiple feature points are synchronized to the three-dimensional coordinate system of the aluminum plate surface to be converted into multiple three-dimensional spatial coordinate points to construct the original structural point data.

[0029] Specifically, multiple equally spaced image sets are input into the feature analysis module. Edge detection algorithms, morphology gradient operators, and multi-scale filtering methods are used to comprehensively analyze the morphological contours, edge changes, and differences in illumination reflection in the images, extracting feature information that characterizes the spatial morphology of the aluminum plate surface. This feature information is then quantified into aluminum plate surface feature parameters, including curvature values, concavity / convexity measurements, and edge direction vectors. Based on these surface feature parameters, the multiple equally spaced image sets are labeled frame by frame. Feature matching algorithms (such as SIFT or SURF) are used to establish a correspondence between feature points in adjacent images, thereby locking down multiple stable and reproducible feature points in each frame. A three-dimensional coordinate system for the aluminum plate surface is established. Multi-view geometric reconstruction technology is used to project these feature points into a unified three-dimensional space, and stereo matching and triangulation methods are used to calculate the three-dimensional spatial coordinates of each feature point. All calculated three-dimensional spatial coordinates are organized and stored in point cloud format to form original structural point data containing spatial morphological information of the aluminum plate surface.

[0030] Furthermore, the method for generating raw surface pixel data by performing dual-path processing based on the multiple equally spaced image sets includes:

[0031] The multiple equally spaced image sets are transmitted to the RGB channel for image analysis to generate multiple RGB pixel blocks; the multiple RGB pixel blocks are used as target separation points to separate the multiple equally spaced image sets and determine multiple separation pixel values; based on the multiple separation pixel values, image blur fusion is performed according to multi-band information to construct multimodal image data; the original surface pixel data is obtained by filling according to the pixel distribution of the multimodal image data.

[0032] Specifically, multiple equally spaced image sets are input into the RGB channels of the image processing module. Each frame undergoes color component separation and brightness normalization, decomposing the image into three color component matrices: R, G, and B. Based on an image segmentation algorithm, these component matrices are divided into multiple RGB pixel blocks according to a preset pixel block size (e.g., 8×8 pixels or 16×16 pixels). These generated RGB pixel blocks are used as target separation points. Color domain separation is then performed on the multiple equally spaced image sets, and different color features are categorized using color histogram clustering or K-means clustering algorithms. Pixel values ​​corresponding to each cluster center are extracted to obtain multiple separated pixel values. Based on these multiple separated pixel values ​​and multi-band imaging information (including visible light and infrared band data), images of different bands are blurred and fused using a multi-scale Gaussian blur and weighted fusion algorithm to construct multimodal image data that simultaneously preserves spectral features and color details. According to the pixel distribution characteristics of the multimodal image data, interpolation filling and boundary smoothing algorithms are used to fill in missing or incomplete pixel regions, generating original surface pixel data that covers the complete surface area and has multi-band spectral information.

[0033] Based on the original structural point data, a three-dimensional reconstruction is performed to construct a three-dimensional digital model of the aluminum plate surface.

[0034] Furthermore, based on the original structural point data, a three-dimensional reconstruction is performed to construct a three-dimensional digital model of the aluminum plate surface. The method includes:

[0035] The original structural point data is traversed, and outlier filtering is performed on the multiple three-dimensional spatial coordinate points to determine multiple data points. Based on the multiple data points, topology verification is performed, and feature analysis is performed on the original structural point data according to the verification results to generate point cloud distribution feature data. Dynamic reference surface fitting is performed based on the point cloud distribution feature data to generate a reference plane. Multi-resolution reconstruction is performed according to the reference plane to construct a three-dimensional digital model of the aluminum plate surface.

[0036] Specifically: First, the original structural point data is processed by traversing it. Multiple 3D spatial coordinate points are input into the outlier filtering module. A combination of statistical filtering and radius neighborhood filtering is used to remove outliers whose number of points in the neighborhood is below a preset threshold or whose distance from the neighborhood mean deviates from the standard deviation by a multiple, thus obtaining multiple data points with continuous distribution and no obvious noise. Second, the multiple data points are imported into the topology verification module to detect the spatial connectivity, normal vector consistency, and boundary integrity of the point cloud. Based on the topology verification results, feature analysis is performed on the point cloud data to extract point cloud distribution feature data, including point density distribution, curvature variation range, and normal vector direction. Then, based on the point cloud distribution feature data, dynamic reference plane fitting is performed using the least squares plane fitting method or the B-spline surface fitting method to generate a reference plane that reflects the overall surface trend of the aluminum plate. This reference plane is used as the basis for shape alignment in subsequent modeling. Finally, using the reference plane as a reference, a multi-resolution reconstruction algorithm (such as the Poisson surface reconstruction algorithm or the octree-based hierarchical reconstruction algorithm) is used to generate surface mesh models at different resolution levels in sequence. Through normal vector interpolation and surface smoothing optimization, a complete and high-precision three-dimensional digital model of the aluminum plate surface is obtained.

[0037] Based on the original surface pixel data, the three-dimensional digital model of the aluminum plate surface is subjected to transfer learning. Based on the learning results, multi-scale defect detection is performed to generate a defect feature vector set.

[0038] Furthermore, based on the original surface pixel data, transfer learning is performed on the three-dimensional digital model of the aluminum plate surface. Multi-scale defect detection is then performed based on the learning results to generate a defect feature vector set. The method includes:

[0039] The original surface pixel data is merged and aligned with the three-dimensional digital model of the aluminum plate surface through multiple channels. Based on the alignment result, transfer learning is performed on the three-dimensional digital model of the aluminum plate surface to construct a visual backbone network. Channel expansion is performed through the visual backbone network to obtain the learning result. Multi-scale feature extraction is performed based on the learning result to generate multi-scale surface texture features. Geometric features of the aluminum plate surface are extracted based on the three-dimensional digital model of the aluminum plate surface. The geometric features are fused with the multi-scale surface texture features across modes. Defect detection is performed based on the fusion result to generate a defect-sensitive feature map. The defect-sensitive feature map is traversed to locate defects and lock multiple defect regions. Structured feature processing is performed on the multiple defect regions to generate the defect feature vector set.

[0040] Specifically, the original surface pixel data is merged and spatially aligned with the 3D digital model of the aluminum plate surface through multi-channel data merging. This involves mapping the multi-band pixel matrix onto the corresponding surface patches according to 3D grid coordinates, and eliminating sampling gaps using bilinear interpolation or nearest-neighbor interpolation to obtain an aligned dataset that integrates geometric and spectral texture information. This aligned dataset is then input into a pre-trained visual backbone network (e.g., ResNet, DenseNet, or SwinTransformer) for transfer learning. By freezing some general feature extraction layers and fine-tuning the parameters of task-related layers, the network is adapted to the unique texture and shape features of the aluminum plate defects, thus constructing a visual backbone network capable of simultaneously processing geometric and texture information. Channel expansion is performed at the backbone network output, and learning results containing features from different receptive fields are obtained through multi-branch convolution or multi-scale feature fusion. The learning results utilize a multi-scale feature extraction module to extract features from shallow feature maps (capturing subtle texture changes), mid-level feature maps (identifying local morphological differences), and deep feature maps (analyzing global structural trends) to generate multi-scale surface texture features. Next, geometric features, including curvature distribution, normal vector changes, and concavity / convexity measurements, are extracted from the 3D digital model of the aluminum plate surface and fused with the multi-scale surface texture features across modalities. During the fusion process, geometric attention is used to highlight the texture response of structurally abnormal areas while suppressing interference features from the background, generating a defect-sensitive feature map. Finally, the defect-sensitive feature map is traversed, and multiple defect regions are located using region growing, connected component analysis, or attention-based region extraction methods. Each defect region undergoes structured feature processing to extract information including location coordinates, shape and size, orientation angle, and texture pattern, which are then combined to form a defect feature vector set.

[0041] Furthermore, the method for cross-modal fusion of the geometric features and the multi-scale surface texture features includes:

[0042] Geometric attention is calculated based on the geometric features to generate multiple geometric position curvature values; feature enhancement is performed on the geometric features according to the multiple geometric position curvature values ​​to construct a geometric feature map; the multi-scale surface texture features are analyzed to extract high curvature texture features and flat texture features; weights are assigned to the high curvature texture features and the flat texture features to generate multiple weight coefficients; the high curvature texture features and the flat texture features are calibrated according to the multiple weight coefficients to generate a texture feature map; the geometric feature map and the texture feature map are fused across modes to generate the fusion result.

[0043] Geometric attention calculation is performed based on geometric features extracted from a 3D digital model. Specifically, attention weight distributions are constructed on geometric attributes such as curvature, normal vector change, and surface concavity / convexity measurement. The geometric curvature value corresponding to each spatial location is calculated to reflect the importance of that location in terms of morphology. Next, feature enhancement is performed on the geometric features according to the geometric curvature value. This involves weighted amplification of feature responses in high-curvature regions and suppression of feature responses in flat regions, generating geometric feature maps that highlight structural anomalies. Then, multi-scale surface texture features are analyzed, categorizing them into high-curvature texture features and flat texture features based on their geometric curvature distribution. Descriptive information for each type of texture is extracted through texture statistical analysis (such as gray-level co-occurrence matrix and frequency domain energy distribution). Next, weights are assigned to high-curvature and flat texture features, generating multiple weight coefficients based on the sensitivity requirements of different regions for defect detection. Finally, feature calibration is performed on the high-curvature and flat texture features according to these weight coefficients to achieve a balance between response intensity and noise suppression, generating calibrated texture feature maps. Finally, the geometric feature map and the calibrated texture feature map are input into the cross-modal fusion module. Through feature splicing, weighted fusion or multi-layer attention mechanism, the geometric shape information and surface texture information are fused in a unified feature space to obtain the fusion result.

[0044] The defect feature vector set is traced back to the aluminum plate surface for process optimization, and process control instructions for the aluminum plate surface are generated and fed back to the remote terminal.

[0045] The location coordinates of the defect feature vector set are mapped back to a 3D digital model of the aluminum plate surface. Through coordinate transformation, these coordinates are converted into the 3D coordinates of the defect center point in the actual space of the aluminum plate surface, and then visually annotated on the surface model. Next, using the defect center point as a reference, the aluminum plate surface is divided into regions based on the defect size, shape, and distribution range, generating one or more areas to be repaired. Then, based on the defect type, texture pattern, and geometric morphology parameters in the defect feature vector, defect impact analysis is performed. Combined with the production process database, the impact coefficient of each defect on surface quality, structural strength, or functional performance is calculated. Based on this, process compensation calculations are performed according to the impact coefficients, such as adjusting rolling pressure, polishing time, spraying thickness, or heat treatment temperature, to obtain process parameter compensation values. Finally, the process parameter compensation values ​​are matched with the existing process flow to generate process control instructions containing process adjustment schemes. These instructions are then fed back to the production line control terminal via a remote communication module, achieving adaptive defect compensation and repair of the aluminum plate surface, forming a closed-loop control of defect detection and process optimization.

[0046] Furthermore, the defect feature vector set is traced back to the aluminum plate surface for process optimization, and process control instructions for the aluminum plate surface are generated and fed back to the remote terminal. The method includes:

[0047] Based on the defect feature vector set, coordinate transformation is performed back to the aluminum plate surface to determine the three-dimensional coordinates of the defect center point; the aluminum plate surface is divided according to the three-dimensional coordinates of the defect center point to generate the area to be repaired; defect influence analysis is performed based on the defect feature vector to determine the defect influence coefficient; the process is compensated according to the defect influence coefficient to generate process parameter compensation values; the process parameter compensation values ​​are matched with the process compensation to generate process control instructions; the process control instructions are fed back to the remote terminal to perform adaptive defect compensation on the aluminum plate surface.

[0048] First, the defect feature vector set is input into the backtracking and positioning module. Based on the defect location coordinates, size information, and spatial calibration parameters of the corresponding 3D digital model contained in the defect feature vector set, a coordinate system transformation is performed to map the relative coordinates in the feature vectors to the actual spatial coordinates of the aluminum plate surface, thereby determining the 3D coordinates of the center point of each defect. Second, using the defect center point as a reference, and combining the defect size and shape parameters, the aluminum plate surface is divided into regions. A buffer expansion algorithm or a rectangular / polygonal envelope method is used to generate the repair area covering the defect. Then, the defect type, depth, texture pattern, etc., recorded in the defect feature vector are matched with the defect-process influence model in the process database to perform defect influence analysis and calculate the influence coefficient of the defect on surface quality, structural performance, or functionality. Based on this influence coefficient, the process compensation calculation module is called to generate corresponding process parameter compensation values ​​according to the adjustable process parameters of the production line (such as rolling pressure, polishing speed, coating thickness, heat treatment temperature, etc.). Finally, the process parameter compensation values ​​are matched with the current production process formula to form a process control instruction containing specific adjustment instructions, execution sequence and safety constraints. This process control instruction is then fed back to the remote terminal of the production line through a remote communication interface, realizing adaptive defect compensation and repair of the aluminum plate surface and constructing a closed-loop control system of defect detection-process adjustment-real-time repair.

[0049] In summary, the embodiments of this application have at least the following technical effects:

[0050] First, a high-speed industrial camera is activated according to the parameters of a multi-band ring light source to acquire data online on the aluminum plate surface, generating a raw image dataset containing raw structural point data and raw surface pixel data. Next, 3D reconstruction is performed based on the raw structural point data to construct a 3D digital model of the aluminum plate surface. Then, transfer learning is performed on the 3D digital model based on the raw surface pixel data, and multi-scale defect detection is performed based on the learning results, generating a defect feature vector set. Finally, the defect feature vector set is used to optimize the aluminum plate surface process, generating process control instructions for the aluminum plate surface and feeding them back to a remote terminal. This solves the technical problems of insufficient accuracy in aluminum plate surface defect detection and difficulty in comprehensively and accurately identifying defects in existing technologies, achieving the technical effect of improving the accuracy and comprehensiveness of aluminum plate surface defect detection.

[0051] Example 2 is based on the same inventive concept as the image recognition-based aluminum plate surface defect detection method in the foregoing examples, such as... Figure 2 As shown, this application provides an image recognition-based aluminum plate surface defect detection system, wherein the system includes:

[0052] Image acquisition module 11: Activates a high-speed industrial camera according to multi-band ring light source parameters to acquire images of the aluminum plate surface online, generating an original image dataset containing original structural point data and original surface pixel data; 3D reconstruction module 12: Performs 3D reconstruction based on the original structural point data to construct a 3D digital model of the aluminum plate surface; Defect detection module 13: Performs transfer learning on the 3D digital model of the aluminum plate surface based on the original surface pixel data, performs multi-scale defect detection based on the learning results, and generates a defect feature vector set; Optimization control module 14: Backtracks the defect feature vector set to the aluminum plate surface for process optimization, generates process control instructions for the aluminum plate surface, and feeds them back to the remote terminal.

[0053] Furthermore, the image acquisition module 11 is used to perform the following methods:

[0054] Based on the material data of the aluminum plate, light reflection analysis is performed to obtain the target spectral reflectance curve; the infrared band absorption analysis is performed by traversing the target spectral reflectance curve to set the infrared light source intensity data; the visible light band reflection analysis is performed by traversing the target spectral reflectance curve to set the visible light source color temperature threshold; the light source controller is activated according to the infrared light source intensity data and the light source color temperature threshold to perform light surround adjustment and construct the multi-band ring light source parameters.

[0055] Furthermore, the image acquisition module 11 is used to perform the following methods:

[0056] The surface of the aluminum plate is gridded, and a sampling interval threshold is set according to the gridding result. Based on the parameters of the multi-band ring light source, a high-speed industrial camera is triggered to collect images at equal intervals according to the sampling interval threshold, thereby obtaining multiple sets of images at equal intervals. Based on the multiple sets of images at equal intervals, dual-channel processing is performed to generate original structural point data and original surface pixel data.

[0057] Furthermore, the image acquisition module 11 is used to perform the following methods:

[0058] The multiple equally spaced image sets are subjected to feature analysis to generate aluminum plate surface feature parameters; the multiple equally spaced image sets are identified according to the aluminum plate surface feature parameters to determine multiple feature points; a three-dimensional coordinate system of the aluminum plate surface is constructed, and the multiple feature points are synchronized to the three-dimensional coordinate system of the aluminum plate surface to be converted into multiple three-dimensional spatial coordinate points to construct the original structural point data.

[0059] Furthermore, the image acquisition module 11 is used to perform the following methods:

[0060] The multiple equally spaced image sets are transmitted to the RGB channel for image analysis to generate multiple RGB pixel blocks; the multiple RGB pixel blocks are used as target separation points to separate the multiple equally spaced image sets and determine multiple separation pixel values; based on the multiple separation pixel values, image blur fusion is performed according to multi-band information to construct multimodal image data; the original surface pixel data is obtained by filling according to the pixel distribution of the multimodal image data.

[0061] Furthermore, the three-dimensional reconstruction module 12 is used to perform the following methods:

[0062] The original structural point data is traversed, and outlier filtering is performed on the multiple three-dimensional spatial coordinate points to determine multiple data points. Based on the multiple data points, topology verification is performed, and feature analysis is performed on the original structural point data according to the verification results to generate point cloud distribution feature data. Dynamic reference surface fitting is performed based on the point cloud distribution feature data to generate a reference plane. Multi-resolution reconstruction is performed according to the reference plane to construct a three-dimensional digital model of the aluminum plate surface.

[0063] Furthermore, the defect detection module 13 is used to perform the following method:

[0064] The original surface pixel data is merged and aligned with the three-dimensional digital model of the aluminum plate surface through multiple channels. Based on the alignment result, transfer learning is performed on the three-dimensional digital model of the aluminum plate surface to construct a visual backbone network. Channel expansion is performed through the visual backbone network to obtain the learning result. Multi-scale feature extraction is performed based on the learning result to generate multi-scale surface texture features. Geometric features of the aluminum plate surface are extracted based on the three-dimensional digital model of the aluminum plate surface. The geometric features are fused with the multi-scale surface texture features across modes. Defect detection is performed based on the fusion result to generate a defect-sensitive feature map. The defect-sensitive feature map is traversed to locate defects and lock multiple defect regions. Structured feature processing is performed on the multiple defect regions to generate the defect feature vector set.

[0065] Furthermore, the defect detection module 13 is used to perform the following method:

[0066] Geometric attention is calculated based on the geometric features to generate multiple geometric position curvature values; feature enhancement is performed on the geometric features according to the multiple geometric position curvature values ​​to construct a geometric feature map; the multi-scale surface texture features are analyzed to extract high curvature texture features and flat texture features; weights are assigned to the high curvature texture features and the flat texture features to generate multiple weight coefficients; the high curvature texture features and the flat texture features are calibrated according to the multiple weight coefficients to generate a texture feature map; the geometric feature map and the texture feature map are fused across modes to generate the fusion result.

[0067] Furthermore, the optimization control module 14 is used to perform the following method:

[0068] Based on the defect feature vector set, coordinate transformation is performed back to the aluminum plate surface to determine the three-dimensional coordinates of the defect center point; the aluminum plate surface is divided according to the three-dimensional coordinates of the defect center point to generate the area to be repaired; defect influence analysis is performed based on the defect feature vector to determine the defect influence coefficient; the process is compensated according to the defect influence coefficient to generate process parameter compensation values; the process parameter compensation values ​​are matched with the process compensation to generate process control instructions; the process control instructions are fed back to the remote terminal to perform adaptive defect compensation on the aluminum plate surface.

[0069] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0070] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0071] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for detecting surface defects in aluminum plates based on image recognition, characterized in that, The method includes: The high-speed industrial camera is activated according to the parameters of the multi-band ring light source to collect data on the surface of the aluminum plate online, generating an original image dataset. The original image dataset includes original structural point data and original surface pixel data. Based on the original structural point data, a three-dimensional reconstruction is performed to construct a three-dimensional digital model of the aluminum plate surface. Based on the original surface pixel data, the three-dimensional digital model of the aluminum plate surface is transferred to learn, and multi-scale defect detection is performed according to the learning results to generate a defect feature vector set. The step of performing multi-scale defect detection based on the learning results and generating a defect feature vector set includes: Based on the learning results, multi-scale feature extraction is performed to generate multi-scale surface texture features; Based on the three-dimensional digital model of the aluminum plate surface, the geometric features of the aluminum plate surface are extracted, and the geometric features are fused with the multi-scale surface texture features across modes. Based on the fusion result, defect detection is performed to generate a defect-sensitive feature map. The defect-sensitive feature map is traversed to locate defects, multiple defect regions are locked, and structured feature processing is performed on the multiple defect regions to generate the defect feature vector set. The defect feature vector set is traced back to the aluminum plate surface for process optimization, and process control instructions for the aluminum plate surface are generated and fed back to the remote terminal.

2. The method for detecting surface defects of aluminum plates based on image recognition as described in claim 1, characterized in that, The process and methods for constructing parameters for a multi-band ring light source include: Optical reflectance analysis was performed based on the material data of the aluminum plate to obtain the target spectral reflectance curve; The infrared band absorption analysis is performed by iterating through the target's spectral reflectance curve, and the infrared light source intensity data is set. The reflectance of the target spectrum is analyzed in the visible light band by traversing the target spectrum reflectance curve, and the color temperature threshold of the visible light source is set. The parameters of the multi-band ring light source are constructed by adjusting the light surround according to the infrared light source intensity data and the light source color temperature threshold to activate the light source controller.

3. The method for detecting surface defects of aluminum plates based on image recognition as described in claim 1, characterized in that, The high-speed industrial camera is activated according to the multi-band ring light source parameters to perform online acquisition of the aluminum plate surface, generating a raw image dataset. The raw image dataset includes raw structural point data and raw surface pixel data. The method includes: The surface of the aluminum plate is gridded, and the sampling interval threshold is set according to the gridding result; Based on the parameters of the multi-band ring light source, the high-speed industrial camera is triggered to perform equal-interval acquisition according to the sampling interval threshold, thereby obtaining multiple equally spaced image sets. Based on the multiple equally spaced image sets, dual-path processing is performed to generate original structural point data and original surface pixel data.

4. The method for detecting surface defects of aluminum plates based on image recognition as described in claim 3, characterized in that, Based on the multiple equally spaced image sets, dual-path processing is performed to generate original structural point data. The method includes: Feature analysis is performed on the multiple equally spaced image sets to generate surface feature parameters of the aluminum plate. The multiple equally spaced image sets are identified according to the surface feature parameters of the aluminum plate, and multiple feature points are determined. A three-dimensional coordinate system is constructed on the surface of the aluminum plate. The multiple feature points are synchronized to the three-dimensional coordinate system on the surface of the aluminum plate and converted into multiple three-dimensional spatial coordinate points to construct the original structural point data.

5. The method for detecting surface defects of aluminum plates based on image recognition as described in claim 3, characterized in that, The method involves performing dual-path processing on the multiple equally spaced image sets to generate raw surface pixel data, including: The multiple equally spaced image sets are transmitted to the RGB channel for image analysis to generate multiple RGB pixel blocks; Using the multiple RGB pixel blocks as target separation points, image separation is performed on the multiple equally spaced image sets to determine multiple separation pixel values; Based on the multiple separated pixel values, image blur fusion is performed according to multi-band information to construct multimodal image data; The original surface pixel data is obtained by filling the multimodal image data according to the pixel distribution.

6. The method for detecting surface defects of aluminum plates based on image recognition as described in claim 4, characterized in that, Based on the original structural point data, a three-dimensional reconstruction is performed to construct a three-dimensional digital model of the aluminum plate surface. The method includes: Traverse the original structural point data, filter outout points from the multiple three-dimensional spatial coordinate points, and determine multiple data points; Based on the multiple data points, a topology verification is performed, and feature analysis is conducted on the original structural point data according to the verification results to generate point cloud distribution feature data. Based on the point cloud distribution feature data, a dynamic reference surface is fitted to generate a reference plane. A three-dimensional digital model of the aluminum plate surface is constructed by performing multi-resolution reconstruction based on the reference plane.

7. The method for detecting surface defects of aluminum plates based on image recognition as described in claim 1, characterized in that, The method for performing transfer learning on the three-dimensional digital model of the aluminum plate surface based on the original surface pixel data includes: The original surface pixel data is merged and aligned with the three-dimensional digital model of the aluminum plate surface through multiple channels. Based on the alignment result, the three-dimensional digital model of the aluminum plate surface is transferred to construct a visual backbone network. The learning results are obtained by expanding the channels through the visual backbone network.

8. The method for detecting surface defects of aluminum plates based on image recognition as described in claim 7, characterized in that, The method for cross-modal fusion of the geometric features and the multi-scale surface texture features includes: Geometric attention is calculated based on the geometric features to generate multiple geometric position curvature values; Geometric features are enhanced based on the curvature values ​​of the multiple geometric positions to construct a geometric feature map; Based on the multi-scale surface texture features, high curvature texture features and flat texture features are extracted. The high curvature texture feature and the flat texture feature are weighted and multiple weight coefficients are generated. The high curvature texture feature and the flat texture feature are calibrated according to the multiple weighting coefficients to generate a texture feature map; The geometric feature map and the texture feature map are fused across modes to generate the fusion result.

9. The method for detecting surface defects of aluminum plates based on image recognition as described in claim 1, characterized in that, The defect feature vector set is traced back to the aluminum plate surface for process optimization, and process control instructions for the aluminum plate surface are generated and fed back to the remote terminal. The method includes: Based on the defect feature vector set, coordinate transformation is performed back to the aluminum plate surface to determine the three-dimensional coordinates of the defect center point; The surface of the aluminum plate is divided according to the three-dimensional coordinates of the defect center point to generate the area to be repaired; Based on the defect feature vector, a defect impact analysis is performed to determine the defect impact coefficient. Based on the defect impact coefficient, a compensation calculation is performed on the process to generate process parameter compensation values. The process parameter compensation values ​​are matched with process compensation to generate process control instructions, which are then fed back to the remote terminal to perform adaptive defect compensation on the aluminum plate surface.

10. A surface defect detection system for aluminum plates based on image recognition, characterized in that, The system is used to implement the image recognition-based aluminum plate surface defect detection method according to any one of claims 1-9, the system comprising: Image acquisition module: Activates a high-speed industrial camera according to the parameters of a multi-band ring light source to acquire images of the aluminum plate surface online and generate an original image dataset, which includes original structural point data and original surface pixel data. 3D Reconstruction Module: Based on the original structural point data, perform 3D reconstruction to construct a 3D digital model of the aluminum plate surface; Defect detection module: Based on the original surface pixel data, the three-dimensional digital model of the aluminum plate surface is transferred to learn, and multi-scale defect detection is performed according to the learning results to generate a defect feature vector set; Optimization control module: The defect feature vector set is traced back to the aluminum plate surface for process optimization, and process control instructions for the aluminum plate surface are generated and fed back to the remote terminal.

Images