Aluminum profile surface defect online detecting and sorting system
By combining high-resolution image acquisition with multi-branch neural networks and three-dimensional morphology quantization, the problems of consistency in aluminum profile surface defect detection and single sorting logic were solved, achieving accurate evaluation and flexible sorting, and improving production efficiency and material utilization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGDONG WEIYE ALUMINUM CO LTD
- Filing Date
- 2026-03-06
- Publication Date
- 2026-05-15
AI Technical Summary
Existing aluminum profile surface defect detection mainly relies on manual visual inspection, which is inefficient and inconsistent. Two-dimensional automatic optical inspection systems cannot accurately assess the three-dimensional morphological parameters of defects, resulting in high rates of missed detection and false detection. Furthermore, the simple sorting logic cannot meet the diverse needs of customers.
High-resolution industrial cameras combined with a ring light source system are used to acquire images of aluminum profile surfaces. By fusing two-dimensional visual recognition and three-dimensional morphology quantization through a multi-branch neural network, and combining process limit thresholds and dynamic customer thresholds to make multi-level decisions, accurate assessment of defects and flexible sorting are achieved.
It enables precise assessment and flexible sorting of surface defects in aluminum profiles, improves the yield of high-quality products and resource utilization, provides a basis for process optimization, and ensures production feasibility and the satisfaction of customers' differentiated needs.
Smart Images
Figure CN122048804A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of industrial inspection and sorting technology, specifically relating to an online inspection and sorting system for surface defects in aluminum profiles. Background Technology
[0002] Aluminum profiles, as important structural and decorative materials, are widely used in construction, transportation, electronics, and machinery manufacturing. During their production, due to fluctuations in extrusion process parameters, die wear, uneven cooling, or subsequent handling collisions, various defects such as scratches, pits, bumps, pinholes, cracks, and color differences can easily occur on the surface. These defects not only affect the product's appearance but can also become stress concentration points, reducing the material's fatigue strength and corrosion resistance, ultimately impacting the quality, safety, and service life of the final product.
[0003] Traditional surface defect inspection primarily relies on manual visual inspection. This method is inefficient, labor-intensive, and the results are significantly affected by human experience, fatigue, and subjective judgment, resulting in poor consistency, high rates of missed detections and false detections, and is no longer suitable for the fast-paced, continuous production rhythm of modern manufacturing. To improve automation, two-dimensional automated optical inspection (AOI) technology based on machine vision has been gradually introduced. These systems acquire images using industrial cameras and utilize image processing algorithms or simple machine learning models to identify and locate defects, thus automating the inspection process.
[0004] However, existing AOI technologies and systems still have the following significant limitations: Most systems rely solely on two-dimensional image information (such as color, texture, and contrast) for judgment, failing to capture crucial three-dimensional morphological parameters of defects, such as precise depth, pit volume, or protrusion height. A defect that appears obvious in an image may actually have a negligible depth, not affecting usability; conversely, a pit with inconspicuous image features may have an actual depth exceeding safety standards. Two-dimensional information alone cannot provide a scientific and accurate assessment of the severity of defects, their actual impact on material properties, and the feasibility of repair.
[0005] The sorting logic is simplistic and lacks flexibility: Existing automated sorting systems typically employ a binary "qualified / unqualified" judgment logic, with fixed thresholds. This model fails to consider two crucial dimensions: first, the technical feasibility at the production end, i.e., the limits of defects that the factory's existing repair equipment and technical capabilities can handle; and second, the diverse needs of customers, with different customers and orders having varying requirements for surface quality. The result is either the misclassification of repairable profiles that meet specific customer requirements as scrap, leading to material waste, or the allowing unrepairable or non-compliant defective products to flow into subsequent processes, causing quality disputes. Summary of the Invention
[0006] The purpose of this invention is to provide an online detection and sorting system for surface defects of aluminum profiles. By integrating two-dimensional visual recognition, three-dimensional morphology quantification and multi-level intelligent decision-making, it can achieve accurate assessment and flexible sorting of surface defects, thereby maximizing the satisfaction of customers' differentiated quality requirements and improving the rate of high-quality products and resource utilization while ensuring production feasibility.
[0007] The specific technical solution adopted by this invention is as follows: A method for online detection and sorting of surface defects in aluminum profiles includes the following steps: Acquire at least one framed image during the initial sorting of the target on the main conveyor line; For each of the framed images, the framed images are processed based on an image recognition model to determine whether there are defects on the target surface and to determine the location of the defects; Targets identified as defective are sorted to the secondary sorting line; For the location of the defect area, obtain a set of three-dimensional point quantization parameters of the defect; It also utilizes multi-branch neural networks to perform fine-grained classification and cause tracing of defects, which can be used for process optimization and repair guidance; Then, a second comparison is made using the three-dimensional point quantization parameter set and the process limit threshold set; When the set of three-dimensional point quantization parameters does not exceed the corresponding process limit threshold set, repairable targets are obtained, and unrepairable targets are sorted to the first collection area. Within the allowable range of the process limit threshold set, a dynamic threshold set is set according to the quality level of the customer order; A third comparison is performed using the set of three-dimensional point quantization parameters and the set of dynamic thresholds for the repairable target; When the set of three-dimensional point quantization parameters of the repairable target does not exceed the corresponding dynamic threshold set, the acceptable target is sent back to the main conveyor line, and the unacceptable target is sorted to the second collection area.
[0008] The acquisition of framed images includes: using a high-resolution industrial camera in conjunction with a ring light source system to continuously acquire images of the aluminum profile surface at a frame rate of 500fps-600fps, and preprocessing the acquired images, the preprocessing including at least one of image denoising, brightness equalization, and contrast enhancement.
[0009] The set of three-dimensional point quantization parameters is obtained by processing the three-dimensional point cloud data of the defect area, and specifically includes: the depth or height value, the projected area value, the volume value, and the length and width value of the defect; The depth or height is obtained by calculating the maximum vertical distance from the defect point cloud to the reference plane; the projected area is obtained by calculating the area of the projected polygon of the defect point cloud on the reference plane; the volume is obtained by calculating the cumulative value of the space enclosed by the defect point cloud and the reference plane; and the length and width are obtained by calculating the circumscribed rectangle size of the defect projection in the main direction.
[0010] The method of using a multi-branch neural network to perform fine classification and cause tracing of defects includes: fusing a two-dimensional image and a three-dimensional topographic image of the defect area and inputting the result into a multi-task neural network; after the neural network extracts the backbone network through shared features, it connects the classification branch and the cause tracing branch in parallel, and outputs the fine category label of the defect and the probability distribution of the potential production process cause, respectively.
[0011] The set of process limit thresholds is a set of fixed values that are comprehensively set based on the upper limit of the technical capability of the production line repair equipment and the minimum requirements of the subsequent processing technology of aluminum profiles for the surface condition of the substrate. It is pre-stored in the database and is used as an absolute standard to determine whether the defect has the possibility of physical repair.
[0012] The dynamic threshold set is set according to the customer order quality level within the allowable range of the process limit threshold set; the dynamic threshold set includes allowable defect depth / height thresholds, area thresholds, volume thresholds, and length / width thresholds; The customer order quality level includes at least three levels: A (high-end), B (mid-range), and C (normal). Different levels correspond to different sets of dynamic thresholds. Among them, the thresholds for defect depth, area, volume, and size are the most stringent for level A, while those for level C are the most lenient. The system automatically calls the corresponding set of dynamic thresholds for comparison and decision-making based on the input order information.
[0013] An online detection and sorting system for surface defects in aluminum profiles includes: The image acquisition module, located above the main conveyor line, is used to acquire framed images of the aluminum profile surface. The image processing and defect recognition module is communicatively connected to the image acquisition module and is used to run the image recognition model, process the framed image to determine whether a defect exists and locate the defect; The first sorting execution mechanism, located on the side of the main conveyor line and controlled by the control center, is used to sort aluminum profiles with identified defects to the secondary sorting line. The 3D scanning module is located above the secondary sorting line and is used to acquire 3D point cloud data for the located defect areas. The data processing and analysis module is used to process 3D point cloud data to extract a set of 3D point quantization parameters and store a set of process limit thresholds and a set of dynamic thresholds. The comparison and decision module is used to perform the comparison logic between the three-dimensional point quantization parameter set and the process limit threshold set and dynamic threshold set, and output the sorting decision instruction; The second sorting execution mechanism, located on the secondary sorting line and controlled by the comparison and decision module, is used to sort irreparable targets to the first collection area. The third sorting execution mechanism, located on the secondary sorting line and controlled by the comparison and decision module, is used to sort unacceptable items to the second collection area and send acceptable items back to the main conveyor line. The control center is connected to the image processing and defect identification module, data processing and analysis module, comparison and decision-making module, and each sorting execution mechanism to coordinate and control the entire system process.
[0014] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.
[0015] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the above-described method.
[0016] The technical effects achieved by this invention are as follows: This invention establishes a two-level intelligent decision-making mechanism between the process and the customer, enabling refined management of quality control and optimal resource allocation. The process limit threshold ensures technical feasibility, strictly screens out unrepairable products that exceed the factory's existing repair capabilities or affect the fundamental requirements of subsequent processing, avoids ineffective repair attempts and waste of production resources, and ensures that all profiles entering the repair process are technically processable. The dynamic customer threshold meets the diversity of market demands. The system can dynamically adjust the acceptance standards according to different customer order levels, realizing flexible production of "first-class products for high-end customers and qualified products to meet ordinary needs". This not only ensures the stringent quality of high-end orders, but also avoids downgrading or scrapping a large number of profiles with minor defects but meeting the requirements of specific orders due to the uniform adoption of the strictest standards, which significantly improves the rate of high-quality products and the overall material utilization rate.
[0017] This invention uses a multi-branch neural network to perform detailed classification and causal analysis of defects while detecting them, outputting the defect type and possible causes in the production process. This provides process engineers with intuitive data support, helps to quickly locate and solve root cause problems in the production process, and realizes the transformation from "post-process sorting" to "pre-process prevention". Attached Figure Description
[0018] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system block diagram of the present invention. Detailed Implementation
[0019] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.
[0020] like Figure 1 As shown, an online detection and sorting method for surface defects in aluminum profiles includes the following steps: At least one frame image is acquired during the first sorting of the target on the main conveyor line. A high-resolution industrial camera is used in conjunction with a ring light source system to continuously acquire images of the aluminum profile surface at a frame rate of 500fps-600fps or higher. The acquired images are preprocessed, including image denoising, brightness equalization, and contrast enhancement, to improve image quality and provide clear input data for subsequent defect identification. For each frame image, the image recognition model is used to process the frame images to determine the location of the defects; based on the location of the defects, a bounding box is added to the conveyor belt frame image. The image recognition model is a deep learning-based convolutional neural network model, using the YOLO architecture. It is trained using a large amount of labeled aluminum profile surface defect image data to achieve rapid and accurate identification and pixel-level localization of various defect types. The control center drives the first robotic arm on the main conveyor line to sort defective aluminum profiles from the main conveyor line to the secondary sorting line, while defect-free profiles continue to flow on the main conveyor line and enter the subsequent process of qualified products. For defective targets entering the secondary sorting line, precise three-dimensional morphological data of the defects are obtained using three-dimensional scanning technology for the identified defect locations. At the same time, multi-branch neural networks are used to perform fine classification and cause tracing of defects, which can be used for process optimization and repair guidance. Continuously collect defect image data generated during historical production processes and their corresponding set of quantitative parameters verified by 3D scanning; Process experts double-label historical defect cases with detailed category tags and root cause tags; Construct a multi-branch, multi-task neural network model. Its core is a shared feature extraction backbone network used to extract high-level semantic features from the input defect region image, fusing a normalized 3D depth map as an additional channel. The backbone network is followed by two parallel, task-specific branches: Classification branch: Consists of a fully connected layer classifier, which outputs the probability distribution of defects belonging to each fine category; The source tracing branch consists of another fully connected layer classifier that outputs the probability distribution of defects attributable to various potential manufacturing process causes. The multi-task model is trained end-to-end using a labeled dataset, and the loss function is a weighted multi-task loss. Through training, the model learns to simultaneously decouple information used to distinguish defect types and infer causes from image features. After the aluminum profile target is sorted to the secondary line and the three-dimensional scan is completed, the two-dimensional image block and the three-dimensional topography map of the defect area are simultaneously fed into the multi-task neural network model. One is used to output the specific defect category, and the other is used to output the probability distribution of possible causes of the defect. The diagnostic results are bound to the three-dimensional parameters to form a defect profile. The traceability results can provide additional reference for sorting decisions and can also generate reports showing the frequency, distribution location, and corresponding three-dimensional severity of defects with different causes. This provides process engineers with intuitive and quantitative evidence to accurately locate problems in the production process; A specially coded grating pattern is projected onto the defect area, and the deformed light spot modulated by the profile surface is captured by a camera. The three-dimensional coordinates of each point on the surface are calculated by the principle of triangulation. The surface height information is calculated by scanning the defect area with a laser line and calculating the displacement of the laser line in the camera image. Defects are captured from different perspectives using two or more cameras, and three-dimensional information is calculated by matching corresponding points and utilizing parallax. By processing the acquired 3D point cloud data and extracting the 3D point quantization parameter set, the acquired complete point cloud is first filtered and denoised. Then, the random sampling consistency algorithm is used to fit the reference plane of the aluminum profile surface from the defect surrounding area. By calculating the distance from each point in the defect area point cloud to the reference plane, the accurate segmentation of the defect and the quantization calculation of the 3D morphology parameters are realized. The set of 3D point quantization parameters includes: Depth / Height: The maximum distance from the bottom of the defect to the surface reference plane, or the distance from the highest point to the reference plane; Area: The projected area of the defect on the surface; Volume: The volume of material missing or superfluous that is contained in the defect; Length / Width: The dimension of the defect in the principal direction; The obtained set of 3D defect point quantization parameters is compared item by item with the set of process limit thresholds; the comparison logic is as follows: If all parameters do not exceed their corresponding process limit thresholds, the defect is determined to be technically repairable, and the target enters the repairable queue. If any parameter exceeds its corresponding process limit threshold, the defect is determined to be unrepairable. The control center immediately drives the second robotic arm on the secondary sorting line to sort the target to the first collection area. The set of process limit thresholds is determined by combining the minimum requirements of subsequent aluminum profile processing technology on the surface condition of the substrate and the upper limit of the technical capabilities of the factory's existing repair equipment. Within the allowable range of the process limit threshold set, a dynamic threshold set is set according to the customer order quality level. The customer order quality level is divided into three levels: A-level high-end, B-level mid-range, and C-level ordinary. Each level corresponds to different surface quality requirements. The dynamic threshold set includes allowable defect depth / height thresholds, area thresholds, volume thresholds, and length / width thresholds. These thresholds are automatically adjusted according to the order level to achieve dynamic management of quality standards. The comparison is performed using a set of 3D point quantization parameters of the repairable target and a set of dynamic thresholds; the comparison logic is as follows: If all parameters do not exceed their corresponding dynamic thresholds, the defect is determined to be within the customer's acceptable range and is acceptable after repair. The target is marked as an acceptable target and sent back to the main conveyor line by the secondary sorting line to be combined with qualified products. If any parameter exceeds its corresponding dynamic threshold, the defect is determined to be repairable but does not meet the current order quality requirements and is therefore an unacceptable target. The control center then drives the third robotic arm to sort the defect to the second collection area.
[0021] like Figure 2 As shown, an online detection and sorting system for surface defects in aluminum profiles includes: The image acquisition module, located above the main conveyor line, is used to acquire framed images of the aluminum profile surface. The image processing and defect recognition module communicates with the image acquisition module and is used to run the image recognition model, process framed images to determine whether defects exist and locate defects. The first sorting execution mechanism, located on the side of the main conveyor line and controlled by the control center, is used to sort aluminum profiles with identified defects to the secondary sorting line. The 3D scanning module is located above the secondary sorting line and is used to acquire 3D point cloud data for the located defect area. The 3D scanning module is any one or more of the following combinations: a 3D scanner based on structured light coding, a laser line scanner, or a multi-view stereo vision camera. The data processing and analysis module is used to process 3D point cloud data to extract a set of 3D point quantization parameters and store a set of process limit thresholds and a set of dynamic thresholds. The comparison and decision module is used to perform comparison logic between the three-dimensional point quantization parameter set and the process limit threshold set and dynamic threshold set, and output sorting decision instructions; The second sorting execution mechanism, located on the secondary sorting line and controlled by the comparison and decision module, is used to sort irreparable targets to the first collection area. The third sorting execution mechanism, located on the secondary sorting line and controlled by the comparison and decision module, is used to sort unacceptable items to the second collection area and send acceptable items back to the main conveyor line. The control center communicates with the image processing and defect identification module, the data processing and analysis module, the comparison and decision-making module, and each sorting execution mechanism to coordinate and control the entire system process.
[0022] To ensure the real-time performance of online detection, the control center integrates a high-performance industrial computer and a graphics processor to accelerate image recognition model inference and 3D point cloud data processing. All modules are synchronized through a high-speed industrial bus to ensure that the total processing time from image acquisition to final sorting decision meets the production line cycle time requirements.
[0023] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.
[0024] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the above-described method.
[0025] Example 1: Processing procedure for minor scratches This embodiment demonstrates the system's processing of an aluminum profile with minor surface scratches, which are repairable and meet the requirements of high-end orders. Image Acquisition and Preliminary Recognition: The aluminum profile moves at a constant speed on the main conveyor line. The image acquisition module, equipped with a high-resolution industrial camera with a ring light source, continuously acquires surface images at a frame rate of 550fps. After preprocessing, including noise reduction and brightness equalization, the acquired images are sent to the image processing and defect recognition module. The pre-trained YOLO model in the module identifies a suspected "scratch" defect in the middle of the profile and precisely locates its area with a bounding box (pixel coordinates [200:250, 500:550]). The control center records the target ID and defect location.
[0026] Initial sorting and 3D scanning: The control center instructs the first sorting actuator to sort the aluminum profile from the main conveyor line to the secondary sorting line. The secondary sorting line transports the profile to the 3D scanning module below, where a laser line scanner is used. Based on the recorded defect location, the system controls the scanner to perform a high-density 3D scan of the area to obtain accurate point cloud data. Parameter extraction and intelligent analysis: The data processing and analysis module processes the defect point cloud by fitting a reference plane, segmenting the defect point cloud, and calculating a set of three-dimensional point quantization parameters. The results show that the maximum scratch depth is 0.05 mm and the projected area is 1.2 mm. 2 Its volume is 0.04mm. 3 The image is 8mm long and 0.15mm wide. At the same time, the two-dimensional image block and the three-dimensional topography of this area are fed into a multi-branch neural network. The network output is finely classified as "slight friction scratches". The source tracing branch gives "slight wear in the mold guide area" and "foreign objects in the conveyor roller" as possible causes. This information is recorded in the defect file and can be used for process optimization reference.
[0027] Secondary decision-making based on process limits: The comparison and decision module calls a pre-stored set of process limit thresholds, with a maximum allowable repair depth of 0.8 mm and a maximum allowable repair area of 15 mm. 2 The obtained parameters are compared with the corresponding thresholds. Since all parameters do not exceed the process limits, the defect is determined to be a "repairable target". The system does not trigger the second sorting execution mechanism to sort the unrepairable items to the first collection area. Final decision based on dynamic thresholds: The current production order is a high-end order from customer A. The system automatically calls the most stringent set of dynamic thresholds corresponding to this level; the maximum allowable depth for level A is 0.1mm, and the maximum area is 2mm. 2 The comparison and decision module performs a third comparison, examining the scratch parameters: depth 0.05mm < 0.1mm, area 1.2mm. 2 <2mm 2 …; All A-level standards were met again; Final sorting action: The system determines the target to be "acceptable". The control center instructs the third sorting actuator to return the aluminum profile from the secondary sorting line to the main conveyor line, allowing it to re-enter the qualified product flow channel. After subsequent repair treatments such as rapid polishing, the profile can be delivered as a Grade A product. Example 2: Processing flow for deep pit defects This embodiment demonstrates the system's processing of an aluminum profile with a deep dent, a defect that is beyond the factory's repair capabilities. Image acquisition and preliminary identification: Similar to step 1 of Example 1, the system identifies and locates a "pit" defect on the surface of another aluminum profile; Initial sorting and 3D scanning: Same as step 2 in Example 1, the target is sorted to the secondary line and 3D scanning is performed; Parameter extraction and intelligent analysis: The data processing and analysis module calculates the 3D point quantization parameters: the maximum depth of the pit is 1.5mm, and the projected area is 8mm². 2 Its volume is 7.5mm. 3The multi-branch neural network classifies the issue as "extrusion bubble pits," with the source tracing indicating either "high gas content in the cast rod" or "excessive extrusion speed." Secondary decision-making based on process limits: The comparison and decision module compares the parameters with a set of process limit thresholds. Among these, the depth parameter of 1.5mm exceeds the process limit threshold of "maximum allowable repair depth of 0.8mm". Final sorting process: Based on the logic that "any parameter exceeding the limit determines it as irreparable," the system immediately identifies the target as "irreparable." The control center instructs the second sorting mechanism to directly sort it to the first collection area. The subsequent third comparison based on dynamic thresholds is not executed. This avoids sending profiles that cannot be fundamentally repaired into the repair process or attempting to deliver them to customers, thus saving production resources.
[0028] In summary, the system described in this invention not only achieves automatic defect identification and sorting, but more importantly, it achieves a leap from "simple binary judgment" to "multi-dimensional intelligent decision-making" by introducing three-dimensional quantitative parameters, two-level decision-making based on process limits and customer dynamic thresholds, and intelligent cause analysis. This ensures technical feasibility, adapts to diverse market demands, and provides a basis for process optimization, thereby significantly improving production flexibility and material utilization while ensuring quality.
[0029] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention are implemented according to conventional methods in the art unless otherwise specified or limited.
Claims
1. A method for online detection and sorting of surface defects in aluminum profiles, characterized in that, Includes the following steps: Acquire at least one framed image during the initial sorting of the target on the main conveyor line; For each of the framed images, the framed images are processed based on an image recognition model to determine whether there are defects on the target surface and to determine the location of the defects; Targets identified as defective are sorted to the secondary sorting line; For the location of the defect area, obtain a set of three-dimensional point quantization parameters of the defect; It also utilizes multi-branch neural networks to perform fine-grained classification and cause tracing of defects, which can be used for process optimization and repair guidance; Then, a second comparison is made using the three-dimensional point quantization parameter set and the process limit threshold set; When the set of three-dimensional point quantization parameters does not exceed the corresponding process limit threshold set, repairable targets are obtained, and unrepairable targets are sorted to the first collection area. Within the allowable range of the process limit threshold set, a dynamic threshold set is set according to the quality level of the customer order; A third comparison is performed using the set of three-dimensional point quantization parameters and the set of dynamic thresholds for the repairable target; When the set of three-dimensional point quantization parameters of the repairable target does not exceed the corresponding dynamic threshold set, the acceptable target is sent back to the main conveyor line, and the unacceptable target is sorted to the second collection area.
2. The method according to claim 1, characterized in that: The acquisition of framed images includes: using a high-resolution industrial camera in conjunction with a ring light source system to continuously acquire images of the aluminum profile surface at a frame rate of 500fps-600fps, and preprocessing the acquired images, the preprocessing including at least one of image denoising, brightness equalization, and contrast enhancement.
3. The method according to claim 1, characterized in that: The set of three-dimensional point quantization parameters is obtained by processing the three-dimensional point cloud data of the defect area, and specifically includes: the depth or height value, the projected area value, the volume value, and the length and width value of the defect; The depth or height is obtained by calculating the maximum vertical distance from the defect point cloud to the reference plane; the projected area is obtained by calculating the area of the projected polygon of the defect point cloud on the reference plane; the volume is obtained by calculating the cumulative value of the space enclosed by the defect point cloud and the reference plane; and the length and width are obtained by calculating the circumscribed rectangle size of the defect projection in the main direction.
4. The method according to claim 1, characterized in that: The method of using a multi-branch neural network to perform fine classification and cause tracing of defects includes: fusing a two-dimensional image and a three-dimensional topographic image of the defect area and inputting the result into a multi-task neural network; after the neural network extracts the backbone network through shared features, it connects the classification branch and the cause tracing branch in parallel, and outputs the fine category label of the defect and the probability distribution of the potential production process cause, respectively.
5. The method according to claim 1, characterized in that: The set of process limit thresholds is a set of fixed values that are comprehensively set based on the upper limit of the technical capability of the production line repair equipment and the minimum requirements of the subsequent processing technology of aluminum profiles for the surface condition of the substrate. It is pre-stored in the database and is used as an absolute standard to determine whether the defect has the possibility of physical repair.
6. The method according to claim 1, characterized in that: The dynamic threshold set is set according to the customer order quality level within the range allowed by the process limit threshold set; the dynamic threshold set includes allowed defect depth / height thresholds, area thresholds, volume thresholds, and length / width thresholds.
7. The method according to claim 1, characterized in that: The customer order quality level includes at least three levels: A (high-end), B (mid-range), and C (normal). Different levels correspond to different sets of dynamic thresholds. Among them, the thresholds for defect depth, area, volume, and size are the most stringent for level A, while those for level C are the most lenient. The system automatically calls the corresponding set of dynamic thresholds for comparison and decision-making based on the input order information.
8. An online detection and sorting system for surface defects of aluminum profiles, used to implement the method as described in any one of claims 1 to 6, characterized in that, include: The image acquisition module, located above the main conveyor line, is used to acquire framed images of the aluminum profile surface. The image processing and defect recognition module is communicatively connected to the image acquisition module and is used to run the image recognition model, process the framed image to determine whether a defect exists and locate the defect; The first sorting execution mechanism, located on the side of the main conveyor line and controlled by the control center, is used to sort aluminum profiles with identified defects to the secondary sorting line. The 3D scanning module is located above the secondary sorting line and is used to acquire 3D point cloud data for the located defect areas. The data processing and analysis module is used to process 3D point cloud data to extract a set of 3D point quantization parameters and store a set of process limit thresholds and a set of dynamic thresholds. The comparison and decision module is used to perform the comparison logic between the three-dimensional point quantization parameter set and the process limit threshold set and dynamic threshold set, and output the sorting decision instruction; The second sorting execution mechanism, located on the secondary sorting line and controlled by the comparison and decision module, is used to sort irreparable targets to the first collection area. The third sorting execution mechanism, located on the secondary sorting line and controlled by the comparison and decision module, is used to sort unacceptable items to the second collection area and send acceptable items back to the main conveyor line. The control center is connected to the image processing and defect identification module, data processing and analysis module, comparison and decision-making module, and each sorting execution mechanism to coordinate and control the entire system process.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 6.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 6.