Surface defect detection system

Abstract

A surface defect detection system provides input data for a modified YOLOv10 algorithm. The surface defect detection system is provided with a surface defect detection platform forming a machining plane, and a to-be-detected object is placed on the machining plane. A light source in a light source module emits illumination light of a corresponding wave band to the machining plane according to detection requirements. An image of the to-be-detected object is collected by a camera module, in particular, a 8K line scanning camera and a 3D line scanner are synchronously started to collect high-resolution 2D image and 3D point cloud data of the to-be-detected object, and after preprocessing and labeling the image of the to-be-detected object, a surface defect image dataset is formed. Thus, the surface defect image dataset can be used to train a surface defect detection model for performing small target defect detection on the to-be-detected object to perform small target defect detection on the to-be-detected object.

Description

Technical Field

[0001] The specification discloses a defect detection technique, specifically a surface defect detection system for establishing a surface defect detection model. Background Technology

[0002] In the modern electronics manufacturing industry, high-density interconnect printed circuit boards (HDI-PCBs) are widely used in various electronic devices due to their high performance and miniaturization.

[0003] As electronic products become increasingly miniaturized and complex, the types and numbers of defects on HDI-PCBs are also increasing, especially small-target defects such as tiny water stains and minor scratches. These small-target defects are often difficult to detect using traditional inspection methods, leading to decreased product performance, reduced reliability, and even safety hazards. Therefore, developing efficient small-target defect detection algorithms has become an urgent industry need to improve inspection accuracy and reduce the generation of defective products. Furthermore, real-time inspection is crucial in HDI-PCB production, especially on high-speed production lines, where any delay can impact overall production efficiency. Traditional manual inspection methods are inefficient and have a high false positive rate, making it difficult to meet the high-precision, high-speed requirements of modern industrial production lines. Utility Model Content

[0004] To address the problems existing in the current technology and meet the need for real-time detection of small target defects, this specification proposes a surface defect detection system. The system utilizes a defect detection technology based on a modified YOLO model to effectively improve the accuracy of small target defect detection in HDI-PCBs, significantly reduce data processing time, and provide a reliable solution for automated defect detection in the HDI-PCB manufacturing process.

[0005] According to an embodiment, the surface defect detection system includes a surface defect detection platform forming a processing plane for placing the object to be inspected. It includes a light source module and a photographic module. The light source module includes a light source and a light source controller. The light source controller controls the light source to emit illumination light onto the object to be inspected according to the inspection requirements. The photographic module includes an 8K line scan camera and a 3D line scanner for acquiring surface defect images of the object to be inspected. This includes synchronously activating the 8K line scan camera and the 3D line scanner with the light source to acquire high-resolution 2D images and 3D point cloud data of the object to be inspected, respectively. After preprocessing and annotation, multiple sets of surface defect images of the acquired objects to be inspected are combined to form a surface defect image dataset.

[0006] Furthermore, this surface defect image dataset is used to train a surface defect detection model that performs small target defect detection on the object to be detected.

[0007] Furthermore, the surface defect detection model runs a modified YOLOv10 algorithm, with the surface defect image dataset providing the input data for the modified YOLOv10 algorithm.

[0008] Furthermore, the modified YOLOv10 algorithm improves the efficiency of extracting key features of small target defects during small target defect detection by using depthwise separable convolution and point convolution design.

[0009] For surface defect image datasets, data augmentation can be performed to expand the surface defect image dataset to establish training, validation, and test sets.

[0010] Furthermore, the light source is a multi-band light source, and the light source controller controls the light source to flicker so as to synchronously drive the imaging module to acquire surface defect images of the object to be inspected.

[0011] This multi-band light source integrates ultraviolet, visible, and infrared light, and can detect various types of defects on the object under test by dynamically switching between different bands.

[0012] Furthermore, the light source module incorporates polarizers and polarizing filters, which can adjust the polarization angle to reduce light crosstalk in highly reflective areas of the object being tested.

[0013] Furthermore, the surface defect detection platform may be equipped with a color mark sensor to detect the incoming material condition of the object to be inspected.

[0014] Furthermore, the camera module and the light source module can be activated based on the incoming material condition of the object to be inspected detected by the color mark sensor, and the light source can be activated synchronously through the light source controller to acquire surface defect images of the object to be inspected.

[0015] Furthermore, the 3D line scanner employs a blue laser triangulation system to obtain a three-dimensional shape detection image of the object to be inspected using the laser triangulation principle and high-speed image processing technology.

[0016] Furthermore, the object to be inspected can be a high-density interconnect printed circuit board, and the surface defect image includes at least small target defects such as circuit board warpage, solder joint height, and microstructural defects.

[0017] To further understand the features and technical content of this utility model, please refer to the following detailed description and drawings of this utility model. However, the drawings provided are for reference and illustration only and are not intended to limit this utility model. Attached Figure Description

[0018] Figure 1 The diagram shows an example of a surface defect detection system utilizing various intelligent tools. Figure 2 A schematic diagram of an embodiment of the surface defect detection system equipment is shown; Figure 3 A schematic diagram of a hardware architecture embodiment of a surface defect detection system is shown; Figure 4 An embodiment of the process diagram showing the method for establishing a defect detection model applied to a surface defect detection system; and Figure 5 Figure shows an example of a small target defect on a high-density interconnect printed circuit board. Detailed Implementation

[0019] This specification describes a surface defect detection system for inspecting objects such as printed circuit boards (PCBs), including high-density interconnect printed circuit boards (HDI-PCBs). A modified YOLO model running within the surface defect detection system improves the neural network structure and optimizes feature extraction, effectively enhancing the detection accuracy of small target defects and enabling real-time processing, thus meeting the industry's urgent need for efficient and accurate detection.

[0020] Figure 1 The diagram shows an example of a surface defect detection system utilizing various intelligent tools.

[0021] The surface defect detection system shown in the figure constructs a surface defect detection platform 100 to form a processing plane for placing the object to be inspected 10. An image acquisition module 101 (including a camera module 121 and a light source module 123) is used to acquire images of the object to be inspected 10. According to an embodiment, the image acquisition module 101 includes a camera module 121 and a light source module 123. The light source module 123 includes a light source and a light source controller. The light source controller can control the light source to emit illumination light onto the object to be inspected 10 on the processing plane according to the inspection requirements. The camera module 121 can use an 8K line scan camera and a 3D line scanner to acquire surface defect images of the object to be inspected. The computer system's control program synchronously activates the 8K line scan camera and the 3D line scanner with the light source to acquire high-resolution 2D images and 3D point cloud data of the object to be inspected 10.

[0022] The image acquisition module 101 acquires surface defect images of the object to be inspected 10. After the images are preprocessed (such as adjusting the brightness, contrast, and color of the images, and excluding defective images) and labeled (such as labeling regions of interest in the images that may have defects), a high-quality defect image dataset is generated and stored in the storage module 103 of the defect detection system.

[0023] The defect detection system incorporates a modified YOLOv10 model 110, which combines a Simple Attention Module (SimAM) 111 and a multi-scale feature extraction module 112 to enhance the model's ability to extract features from small target images. Furthermore, the modified YOLOv10 model 110 can also be combined with a Channel Efficient Small Target Enhancement Module 113 (CESOEM) to further improve the model's ability to detect defects in small targets.

[0024] It's worth mentioning that this simple attention module is an attention mechanism used in convolutional neural networks (CNNs), enabling intelligent models to focus on key regions of an image. This allows them to capture local and global features with minimal parameters, thus improving model performance. The multi-scale feature extraction module can capture small, medium, and large features using convolutional kernels of different sizes (e.g., 7×7, 11×11, 23×23) and horizontal / vertical convolutions (H-Conv / V-Conv). This design can simultaneously handle targets of different sizes, significantly enhancing the model's ability to detect small targets. Furthermore, the channel-efficient small target enhancement module significantly enhances the detection capability of small targets by introducing higher-resolution feature layers and employing a bidirectional feature fusion strategy, while drastically reducing model parameters and computational complexity.

[0025] According to the embodiment, the modified YOLOv10 model 110 introduces the MobileNet General Information Bottleneck C2f module (MU-C2f) 114, employing depthwise separable convolution and pointwise convolution to reduce redundant features and improve feature extraction efficiency. A dynamic sampling fusion block 115 replaces the traditional upsampling module, improving upsampling accuracy through dynamic sampling. Finally, an optimization strategy based on a sliding loss function 116 is adopted to improve the model's convergence speed and detection accuracy. Based on the implementation of the defect detection system, by modifying the YOLOv10 model using the various modules and mechanisms proposed above, the detection accuracy of small target defects can be effectively improved, while reducing the number of model parameters and computational burden.

[0026] According to the embodiment, the channel high-efficiency small target enhancement module 113 introduces a higher resolution small target detection layer and adopts an innovative dynamic upsampling method to improve the upsampling accuracy, realizing more comprehensive multi-scale feature interaction of 1 / 32, 1 / 16, 1 / 8 and 1 / 4. This design can capture target features of different sizes, especially significantly improving the detection capability of small target defects.

[0027] The efficient small target enhancement module 113 achieves more comprehensive multi-scale feature interaction by introducing a higher-resolution feature layer. The innovation of the small target detection layer 21, as shown in the figure, lies in firstly constructing a more refined feature pyramid network with a small target detection layer 21 within the feature extraction layer 23, based on the backbone network feature extraction. This feature pyramid network can progressively upsample features at a depth of 1 / 32 scale to high-resolution features at a shallower 1 / 4 scale through dynamic upsampling 25. Secondly, target detection is performed on the feature map at 1 / 4, 1 / 8, and 1 / 16 scales. Compared to the YOLOv10n model, which only performs detection on low-resolution feature maps, the method for establishing a surface defect detection model proposed in the specification adds a detection head to the high-resolution feature map, enabling the model to more accurately detect small objects. This multi-scale detection strategy ensures that the model can effectively detect large, medium, and small target defects, enhancing the overall robustness of the detection.

[0028] Furthermore, by removing the 1 / 32 scale feature layer of 1024 channels from the YOLOv10n model, the complexity of the surface defect detection model is effectively reduced. This not only reduces the number of parameters and computational load but also avoids information loss that may occur when detecting on extremely low-resolution feature maps. Therefore, the channel-efficient small target enhancement module 113 improves the small target detection performance of the modified YOLOv10 model 110 while increasing computational efficiency. Finally, after improving the feature fusion block, the model's detection capability is also significantly improved, especially for the identification of small-sized defects, while reducing computational complexity and the number of parameters. Thus, the channel-efficient small target enhancement module 113 effectively captures local and global features by combining a parameter-free attention mechanism and multi-scale feature extraction, and improves the model's defect detection accuracy by optimizing the energy function and feature refinement. Moreover, the introduction of multi-scale feature interaction and high-resolution feature layers significantly enhances the detection capability of small target defects while effectively reducing the complexity and computational load of the defect monitoring method.

[0029] Figure 2 Next, a schematic diagram of an embodiment of the surface defect detection system is shown.

[0030] The figure shows that the surface defect detection system includes a surface defect detection platform 20, which is a high-precision and high-efficiency hardware platform for surface defect detection built by means of a mechanism. It provides high-quality input data for the improved YOLOv10 algorithm for objects such as high-density interconnect printed circuit boards (HDI-PCB) to be inspected.

[0031] The surface defect detection platform 20 integrates a variety of advanced imaging technologies, including a light source module (light source and light source controller 213), a color mark sensor 207 that enables automatic activation and acquisition, and a photographic module, such as a 3D line scanner 201 and a 2D imaging module (implemented as an 8K line scan camera 203). It is also equipped with a computer 215 that enables automation, synchronous startup and image processing, and can meet the needs of complex industrial inspection scenarios.

[0032] According to the embodiment, the imaging module is also equipped with a polarizing lens 205, the light source is a multi-band light source 209, and may be equipped with a polarizer 211. The light source controller 213 is a controller for controlling the flashing synchronization of the light source. A color mark sensor 207 is provided on the surface defect detection platform 20 for detecting the incoming material condition of the object to be inspected. When paired with a computer 215, it can automatically perform image acquisition and data processing.

[0033] According to the embodiment, an 8K line scan camera 203 is equipped with a polarizing lens 205. The 8K line scan camera 203 is electrically connected to a computer 215 via a data cable for image acquisition and processing. Next, a light source controller 213 is connected to a multi-band light source 209 and can control the synchronous flashing of the light source. After the multi-band light source 209 is turned on and the light source controller 213 performs flashing synchronization control, illumination light of the corresponding band will be emitted onto the processing plane according to the detection requirements. Simultaneously, the polarization angle of the polarizer 211 is adjusted to adapt to different surface characteristics.

[0034] In addition, the color mark sensor 207 is fixed at an appropriate position on the production line to detect the incoming material status of the object to be inspected (such as HDI-PCB). Once the object to be inspected is detected to enter the production line, the 3D line scanner 201, the 8K line scanner 203 and the multi-band light source 209 are activated to run synchronously. When the object to be inspected is transported on the processing plane of the surface defect inspection platform 20, the 8K line scanner 203 and the 3D line scanner 201 are controlled to start synchronously to collect high-resolution 2D images and 3D point cloud data of the object to be inspected, ensuring the comprehensiveness and accuracy of defect detection.

[0035] Furthermore, a high-resolution 8K line scan camera 203 (e.g., 8192×4320@65kHz resolution) is used as the core imaging device. When the 8K line scan camera 203 receives an automatic activation signal, it immediately starts the high-speed line scan mode to acquire images of the object to be inspected on the processing plane line by line. The 8K line scan camera 203 accurately acquires every detail of the object to be inspected through an activation signal synchronized with the production line, and transmits the acquired high-precision image data to the computer 215 in real time for subsequent analysis.

[0036] According to the light source embodiment, the multi-band light source 209 integrates ultraviolet light (such as 365nm UV-LED), visible light (such as 520nm green light), and infrared light (such as 850nm IR). By dynamically switching different bands of the multi-band light source 209, images of various types of defects on the object to be inspected can be captured. This design is optimized for the online defect characteristics of brown chemical products. Firstly, the light source module integrates ultraviolet, visible, and infrared light. By dynamically switching different bands of light, it can target different types of defects (such as...) Figure 5 The system displays defects such as roller marks, chemical residues, scratches, and water stains, providing optimal imaging conditions.

[0037] For example, infrared light can be used for penetration testing to identify potential defects or foreign objects inside the material of the object being inspected; ultraviolet light can be used to detect defects that are difficult to detect with the naked eye, such as surface chemical residues and minor scratches. Furthermore, to address the issue of surface reflection on the object being inspected (such as HDI-PCB), the light source module incorporates polarizers and polarizing filters. Polarizers, by adjusting the polarization angle, effectively reduce light crosstalk in highly reflective areas. A polarizing filter in front of the light source lens enhances the contrast of surface details, significantly improving the visibility of defects, especially in highly reflective areas (such as scratches and water stains). The combination of polarizers and polarizing filters allows for accurate detection of minute defects (scratches, chemical residues) even in highly reflective environments. Finally, the light source module integrates a light source controller 213 that controls the synchronous operation of the light source, the 3D line scanner 201, and the 8K line scanner camera 203. According to an embodiment, a high-frequency LED driver can enable the light source to flash at an extremely high frequency, avoiding image blurring caused by the movement of the object being inspected on the processing plane. Meanwhile, the precise synchronization between the light source flicker and the camera shutter ensures that each frame of the image has the best exposure effect, thereby significantly improving the image quality in high-speed detection scenarios.

[0038] Furthermore, through a computer 215 and related automatic control mechanisms, such as in conjunction with a color mark sensor 207, the feeding status of the object to be detected is obtained to achieve an automatic activation acquisition function. This is specifically designed for high-speed browning production lines to achieve precise activation of incoming material signals. According to an embodiment, the highly sensitive color mark sensor 207 monitors the position of the object to be detected or other target objects in real time. When the object to be detected enters the designated detection area, an activation signal is immediately generated and sent to subsequent equipment (such as a camera module and a light source module). The output of the activation signal can be delayed according to the production line speed and target position to ensure a high degree of synchronization between the data acquisition equipment and the production line operation, thereby achieving precise image acquisition or program control of the object to be detected.

[0039] According to an embodiment, the 3D line scanner 201 can employ a blue laser triangulation system (e.g., wavelength 405nm, line width 0.05mm), based on the principle of laser triangulation and high-speed image processing technology, specifically designed to obtain high-precision three-dimensional shape detection images of objects to be inspected in industrial production lines. The laser emitting unit can be equipped with a temperature-controlled LD driver (stability ±0.1℃), output power 30mW, and line uniformity >95%; the optical receiving unit can consist of a 12MP global shutter CMOS (pixel size 3.45μm) and a Scheimpflug lens group, with a depth of field of 50mm, implementing a three-dimensional reconstruction algorithm, namely, based on an improved RANSAC point cloud processing framework, achieving sub-pixel-level edge extraction (accuracy 0.1 pixel), and a point cloud density ≥1000 points / mm². This technology projects a continuous laser line onto the surface of the target object using a line laser generator, and uses a high-resolution industrial camera to capture the contour changes formed by the laser line on the surface of the object to be inspected (e.g., surface defects) from a specific angle, calculating the height information of the object's surface through a triangulation algorithm. The system combines a high-performance image processing unit and a data acquisition module, enabling real-time acquisition and reconstruction of high-density point cloud data even when objects are moving rapidly. Furthermore, the 3D line scanner 201 utilizes 3D line scanning technology, which features non-contact operation, high precision, and high speed, making it suitable for detecting complex surface morphologies, such as HDI-PCB board warpage, solder joint height, and minute structural defects.

[0040] use Figure 1 The defect detection system shown is referenced further. Figure 3 The schematic diagram of the hardware architecture of the surface defect detection system shown illustrates a method for establishing a surface defect detection model based on a modified YOLO model through collaborative operation between hardware and software components. The method for establishing this surface defect detection model and using the model to perform defect detection can be found in [reference needed]. Figure 4 The process diagram of the embodiment shown is illustrated.

[0041] according to Figure 3 The illustrated hardware architecture diagram demonstrates how a high-precision, high-efficiency surface defect detection platform 300 provides high-quality input data for small-target defect images to the modified YOLOv10 algorithm. The surface defect detection platform 300 integrates various advanced imaging technologies, including a light source module (light source 302 and light source controller 303 shown in the embodiment), a computer 30 with a signal processing module 306, an automatic activation acquisition module 308 (which can detect the feeding status of the object 301 to be inspected via a color mark sensor 307), image processing functions, and an intelligent model running therein, such as... Figure 1The defect detection system shown is a modified YOLOv10 model 110, and an image acquisition module 101. The image acquisition module 101 is as follows: Figure 1 The display includes a camera module 121 and a light source module 123.

[0042] The embodiment includes a 3D line scanner 304 and an 8K line scanner camera 305 of a 2D imaging module. The connections primarily include an electric connection between a light source 302 and its light source controller 303, which is then electrically connected to a computer 30; the 3D line scanner 304 and the 8K line scanner camera 305 are both electrically connected to the computer 30; the image acquisition module 101 operates synchronously with the light source module to acquire image data of the object to be inspected 301 placed on the processing plane 310; and the modified YOLOv10 algorithm is run through the signal processing module 306 in the computer 30 to achieve the goal of industrial defect detection.

[0043] According to an embodiment, Figure 1 The image acquisition module 101 shown is a photographic module for acquiring images of the surface of the object to be inspected 301. Specifically, it includes an 8K line scan camera 305, a 3D line scanner 304, a polarizing lens, a multi-band light source, a polarizer, a light source 302, a light source controller 303 (implemented as, for example, a light source flicker synchronization control controller), and a color mark sensor. The image data of the object to be inspected acquired by the image acquisition module 101 is processed by the signal processing module 306 of the computer 30 to perform small target defect detection. Related process embodiments can be found in the following examples. Figure 4 .

[0044] exist Figure 4 In the process shown, the surface defect detection platform 300 is set up in the defect detection system, which includes a photographic module, such as the 3D line scanner 304 and the 2D imaging module. The 2D imaging module includes a polarizing lens mounted in front of the 8K line scanner 305. The 3D line scanner 304 and the 8K line scanner 305 are connected to the computer 30 and the signal processing module 306 via data cables for image acquisition and processing.

[0045] Initially, the light source controller 303 (which can synchronously control the light source flashing) is connected to the light source 302 (implemented as, for example, a multi-band light source). After the light source controller 303 is started, the light source 302 is driven to emit multi-band light of the corresponding band according to the detection requirements to illuminate the processing plane 310, in which there is the object to be detected 301, such as HDI-PCB. At the same time, the polarization angle of the polarizer can be adjusted to adapt to different surface characteristics (step S401).

[0046] According to the embodiment, the light source module is optimized for the online defect characteristics of brown chemical products. First, the light source 302 is a multi-band light source integrating ultraviolet light (such as 365nm VU LED), visible light (such as 520nm green light), and infrared light (such as 850nm IR). By dynamically switching between different bands, the multi-band light source system provides optimal imaging conditions, enabling the detection of various types of defects on the object under test (such as chemical residues, scratches, and water stains). For example, infrared light can be used for penetration detection to identify potential defects or foreign objects inside the material of the object under test 301; ultraviolet light can be used to detect defects that are difficult to detect with the naked eye, such as chemical residues and micro-scratches on the surface of the object under test 301.

[0047] Furthermore, to address the surface reflection issue of the object under inspection 301 (such as an HDI-PCB), the light source module incorporates a polarizer and a polarizing filter. The polarizer, by adjusting the polarization angle, effectively reduces light crosstalk in highly reflective areas. A polarizing filter can be fitted in front of the lens mounted on the light source 302 to further enhance the contrast of surface details, especially significantly improving the visibility of defects in highly reflective areas (such as scratches and water stains). The combination of the polarizer and filter allows for accurate detection of minute defects (scratches, chemical residues) even in highly reflective environments. Moreover, the light source controller 303 in the light source module can be a light source flashing synchronization control module, used to synchronize the operation of the light source 302 with the photographic module (such as a 3D line scanner 304 and an 8K line scanner camera 305) in the image acquisition module 101. For example, the light source controller 303 uses a high-frequency LED driver, enabling the light source 302 to flash at an extremely high frequency, avoiding image blurring caused by the movement of the object under inspection 301. Meanwhile, the flickering of the light source 302 can be precisely synchronized with the shutter of the camera module, ensuring that each frame of the image has the best exposure effect, thereby significantly improving the image quality in high-speed detection scenarios.

[0048] In subsequent steps, while synchronously controlling the light source 302, the automatic activation acquisition module 308 fixes the color mark sensor 307 at an appropriate position on the production line to detect the incoming material status of the object to be detected 301, activates the image acquisition module 101 to run, and synchronously activates the light source 302 through the light source controller 303 to perform image acquisition (step S403).

[0049] According to an embodiment, the automatic activation acquisition module 308 is specifically designed for high-speed browning production lines to achieve precise activation of incoming material signals. The automatic activation acquisition module 308 monitors the position of the object to be inspected 301 or other target objects in real time using a high-sensitivity color mark sensor 307. When the object to be inspected 301 is detected entering the designated inspection area, an activation signal is immediately generated and sent to other devices, such as the imaging module and light source module of the surface defect inspection platform 300. According to an embodiment, the output of the activation signal generated by the automatic activation acquisition module 308 can be delayed according to the production line speed and the position of the target (object to be inspected 301) to ensure a high degree of synchronization between the data acquisition equipment and the production line operation, thereby achieving precise image acquisition or program control of the object to be inspected 301.

[0050] Next, the production line equipment of the defect detection system is started, so that when the object to be inspected 301 moves on the conveyor, the 8K line scan camera 305 and the 3D line scanner 304 are started synchronously to collect high-resolution planar (2D) images and stereo (3D) point cloud data of the object to be inspected 301, so as to ensure the comprehensiveness and accuracy of defect detection (step S405).

[0051] According to the embodiment, when forming a 2D image, a high-resolution 8K line scan camera 305 is used as the core imaging device. For example, the resolution of the 8K line scan camera 305 can be 8192×4320@65kHz. When the 8K line scan camera 305 receives the incoming material signal sent by the automatic activation acquisition module 308, it immediately starts the high-speed line scan mode to acquire images of the object 301 to be inspected passing through the inspection area line by line. The 8K line scan camera 305 can accurately capture every detail of the object 301 to be inspected through the activation signal synchronized with the production line, and transmit the acquired high-precision image data to the computer 30 and signal processing module 306 in real time for subsequent analysis.

[0052] According to another embodiment, the line scanning technology of the 3D line scanner 304 is based on the laser triangulation principle and high-speed image processing technology, specifically designed for high-precision three-dimensional shape detection in industrial production lines. The line scanning technology projects a continuous laser line onto the surface of the object to be inspected 301 using a line laser generator. A high-resolution industrial camera captures the contour changes formed by the laser line on the surface of the object 301 from a specific angle, and the height information of the object's surface is calculated using a triangulation algorithm. Furthermore, the defect detection system can be combined with a high-performance imaging module and a signal processing module 306 to complete real-time acquisition and reconstruction of high-density point cloud data even when the object 301 is moving rapidly. The proposed 3D line scanner 304 features non-contact, high precision, and high speed, making it suitable for detecting complex surface morphologies, such as HDI-PCB board warpage, solder joint height, and microstructural defects.

[0053] Furthermore, in one embodiment, the 3D line scanner 304 may employ a blue laser triangulation system (wavelength 405nm, line width 0.05mm), based on the principle of laser triangulation and high-speed image processing technology, specifically designed for high-precision 3D topography inspection in industrial production lines. The 3D line scanner 304 may be equipped with a laser emitting unit, featuring a temperature-controlled LD driver (stability ±0.1℃), output power of 30mW, and line uniformity >95%. The 3D line scanner 304 may also be equipped with an optical receiving unit, which may consist of a 12MP global shutter light sensor (e.g., pixel size 3.45μm) and a Scheimpflug lens group, achieving a depth of field of up to 50mm. The 3D line scanner 304 runs a 3D reconstruction algorithm, based on an improved RANSAC point cloud processing framework, achieving sub-pixel-level edge extraction (accuracy 0.1pixel) and a point cloud density ≥1000 points / mm². When the 3D line scanner 304 is running, a continuous laser line is projected onto the surface of the target object (i.e., the object to be inspected 301) by a line laser generator. A high-resolution industrial camera captures the contour changes formed by the laser line on the surface of the object to be inspected 301 from a specific angle, and the height information of the surface of the object to be inspected 301 is calculated by a triangulation algorithm. This 3D line scanning technology features non-contact, high precision, and high speed, and is suitable for the inspection of complex surface morphologies, such as circuit board warpage, solder joint height, and micro-structural defects.

[0054] After acquiring image data of the object to be inspected 301, the original photographs are exported, preprocessed, and annotated to generate a surface defect image dataset of the object to be inspected 301 (step S407). It is worth noting that, unlike traditional methods that use computer vision technology to train datasets, the method proposed in this specification uses an image dataset of a specific object to be inspected 301 (such as an HDI-PCB), making it more suitable for performing online visual inspection tasks on this object to be inspected 301.

[0055] After obtaining the surface defect image dataset of the object to be inspected 301, data augmentation (such as cropping, rotation, noise addition, etc.) is performed to expand the surface defect image dataset (step S409). Taking HDI-PCB as an example of an object to be inspected, the surface defect image dataset can be divided into a training set, a validation set, and a test set according to a set ratio.

[0056] According to the embodiment, data acquisition, preprocessing, and data augmentation of the surface defect image dataset are implemented in the computer 30 and the signal processing module 306. Specifically, during the preprocessing of the surface defect image dataset, the surface defect image of the object to be detected 301 acquired by the 2D imaging module (such as an 8K line scan camera 305) is segmented using pixels as a threshold. According to the embodiment, the image is divided into nine blocks on average. Data augmentation is achieved by adding noise, randomly changing brightness and sharpness, and flipping the image. The processed surface defect image of the object to be detected 301 is then annotated using an annotation tool, for example, in .txt format, resulting in the surface defect image dataset of the object to be detected 301. Furthermore, the ratio of the training set, validation set, and test set is, in an example, 7:2:1.

[0057] It is worth mentioning that, in order to solve the problem of low accuracy of traditional surface defect detection models in detecting small target defects, the method proposed in the specification for establishing a surface defect detection model based on a modified YOLO model has been innovatively improved on the basis of the YOLOv10n network. In particular, in response to the difficulty of detecting small target defects in industrial production, a series of efficient and lightweight module designs are proposed, which significantly improves detection performance and industrial applicability.

[0058] According to the embodiments, the method for establishing a surface defect detection model based on a modified YOLO model utilizes artificial intelligence technologies such as... Figure 1 As shown, the Channel Efficient Small Target Enhancement Module (CESOEM) 113 introduces a higher resolution feature layer and employs a bidirectional feature fusion strategy, significantly enhancing the detection capability of small target defects in industrial applications while substantially reducing model parameters and computational complexity. Furthermore, the defect detection system proposed in the specification combines the attention mechanism of the simple attention module 111 and the multi-scale feature extraction module 112 to implement a multi-scale simple attention module (MsimAM). This multi-scale simple attention module can accurately extract key features of small target defects while reducing computational overhead. Further, Figure 1 The MobileNet General Information Bottleneck C2f module (MU-C2f)114, as shown, further improves the efficiency of extracting key features of small target defects through depthwise separable convolution and point convolution design, ensuring the model's lightweight nature. These artificial intelligence technologies enable the defect detection system to achieve high-precision, real-time detection of minute defects in industrial production environments, meeting the dual requirements of efficiency and accuracy for production lines, and possessing significant application value.

[0059] It is worth mentioning that this multi-scale simple attention module integrates a simple attention module 111 and a multi-scale feature extraction module 112 to capture local and global features with minimal parameters. The multi-scale simple attention module integrates a simple attention module 111 with a parameter-free attention mechanism and a multi-scale feature extraction module 112. The simple attention module 111 can capture local and global features with minimal parameters, while the multi-scale feature extraction module 112 proposes an optimization scheme based on multi-scale feature extraction and directional convolution. It extracts small, medium, and large image features through convolutional kernels of different sizes and horizontal and vertical convolutions. This design can simultaneously process image features of different sizes, further improving the model's ability to detect missing lines in small targets.

[0060] According to an embodiment, the multi-scale simple attention module identifies the most active neurons by optimizing an energy function. Equation 1 describes the activation value of a single neuron. This optimization process is based on well-established neuroscience theory and generates a 3D attention weight tensor. Furthermore, the purpose of identifying these neurons is to assess the linear separability of the target neuron from other neurons, thereby enhancing the ability of the modified YOLO model to effectively distinguish features.

[0061] Equation 1: .

[0062] in, and It is a linear transformation, where It is the activation value of a single neuron. and It refers to the target neuron and other neurons in the input feature single channel. It is an index in the spatial dimension. It is the number of neurons in the channel. It is weight and sum Bias transformation. By minimizing Equation 1, the linear separability between the target neuron and all other neurons in the same channel can be found. This derives an energy function and the importance of each neuron. Next, feature refinement is performed using scaling operands; the entire refinement stage can be referenced in Equation 2.

[0063] Equation 2: .

[0064] in, These are the activation values ​​of all neurons on the feature map. Input feature map, Output feature map. This multi-scale simple attention module effectively captures local and global features by combining a parameter-free attention mechanism (simple attention module 111) and multi-scale feature extraction (multi-scale feature extraction module 112), and improves the defect detection accuracy of the defect monitoring method by optimizing the energy function and refining the features.

[0065] Furthermore, Figure 1 The MobileNet general information bottleneck C2f module 114 shown is an optimized design of the C2f module in YOLOv10n. Its purpose is to solve the problem of feature redundancy and computational overhead caused by multi-layer convolution operations in the traditional C2f module, and to improve computational efficiency and parameter optimization capabilities at the same time. It is worth mentioning that the MobileNet general information bottleneck C2f module 114 introduces lightweight depthwise separable convolution and pointwise convolution, decomposing (split) the MU-bottlenect part of the traditional full convolution operation into a more efficient four-layer structure: the initial depthwise convolution (Start_dw Conv), the expanded convolution (ExpandConv), the middle depthwise convolution (Middle_dw Conv), and the projective convolution (Project Conv).

[0066] According to the embodiment, the MobileNet general information bottleneck C2f module 114 introduces lightweight depthwise separable convolution and pointwise convolution to decompose the traditional full convolution operation into a more efficient four-layer structure: initial depthwise convolution, dilated convolution, intermediate depthwise convolution, and projective convolution. The depthwise convolution kernel size is 3×3×3, the stride is 1, and the padding is 1. The MobileNet general information bottleneck C2f module 114 performs convolution independently on each channel to extract spatial features. Lightweight filtering is performed by applying a single convolutional filter to each input channel. The dilated convolution expands the number of channels through a 1×1×1 convolution, thus... Feature expansion to This enhances feature representation capabilities. The intermediate depthwise convolution kernel is 3×3×3 in size, used for further feature extraction. The projection convolution uses a 1×1×1 convolution to increase the number of feature channels from... Compress back To reduce computational overhead, the MobileNet general information bottleneck C2f module 114 proposed in the method for establishing a surface defect detection model described in the specification optimizes the C2f module in YOLOv10n by introducing depthwise separable convolution and pointwise convolution. This significantly improves computational efficiency, reduces feature redundancy and computational overhead, thereby enhancing the overall performance of the model and improving the real-time performance of the defect detection method.

[0067] After forming a surface defect image dataset, including a training set, a validation set, and a test set, using the training set, a surface defect detection model can be trained using a machine learning algorithm to perform small target defect detection on the object to be detected 301. After completion, the validation set is used to verify the effectiveness of the surface defect detection model (step S411). Then, the test set is input into the validated surface defect detection model to test its capabilities (step S413). Once completed, the surface defect detection model can be used to perform small target defect detection, outputting surface defect detection results for a specific object to be detected 301 (such as an HDI-PCB) (step S415).

[0068] For small target defects of a specific object to be detected, please refer to Figure 5 The diagram shows an example of a small target defect in an HDI-PCB, with defect types including (but not limited to) various types. Figure 5 (a) shows roller mark defect 501 on the HDI-PCB surface, which was collected and marked. Figure 5 (b) shows scratch defect 502 on the HDI-PCB surface, such as Figure 5 (c) shows water stain defect 503 on the HDI-PCB surface, and as shown Figure 5 (d) shows chemical residue defect 504 on the HDI-PCB surface.

[0069] Based on the results of actual operation of the surface defect detection model, the average accuracy (mAP) for detecting small target defects (such as scratches, chemical residues, and other minute defects) can reach 88.3% by modifying the network structure of the YOLOv10 model 110 and the feature extraction module (such as the multi-scale feature extraction module 112), demonstrating excellent high-precision performance. Furthermore, the surface defect detection model has a parameter size of only 2.0MB and a weight file size of 4.3MB, which is relatively lightweight and suitable for resource-constrained industrial environments. The surface defect detection model has a detection speed of up to 226 frames per second (FPS), which fully meets the needs of industrial production lines for real-time defect detection. In addition, through multi-scale feature fusion and dynamic sampling strategies, the surface defect detection model exhibits strong robustness in complex backgrounds and various defect conditions, and can adapt to diverse industrial application scenarios.

[0070] In summary, the surface defect detection system described in the above embodiments acquires surface defect images of the object to be inspected using a hardware device. The acquired images form a surface defect image dataset, which serves as input data for a modified YOLOv10 algorithm to train a defect detection model for detecting small target defects. Experimental results show that using the surface defect detection model based on the modified YOLO model, particularly for high-density interconnect printed circuit boards, achieves an average accuracy of 88.3%, a 12% improvement over the baseline model. The weight file size is reduced by 19%, the number of parameters is reduced by 22%, and the detection speed reaches 226 frames per second (FPS), thus meeting the real-time inspection requirements of industrial production lines. Therefore, the surface defect detection system proposed in this publication, by running this surface defect detection model, not only significantly improves defect detection efficiency, eliminates defective materials, reduces product scrap and rework rates caused by material quality issues, lowers production costs, and ensures product quality, but is also particularly suitable for the automatic detection of surface defects during the manufacturing process of high-density interconnect printed circuit boards.

[0071] The above-disclosed content is only a preferred and feasible embodiment of the present utility model, and is not intended to limit the scope of the patent application of the present utility model. Therefore, all equivalent technical changes made using the contents of the present utility model specification and drawings are included in the scope of the patent application of the present utility model.

Claims

1. A surface defect detection system characterized by The system includes: A surface defect detection platform forms a processing plane for placing an object to be inspected; A light source module includes a light source and a light source controller, wherein the light source controller controls the light source to emit illumination light onto the object to be detected according to a detection requirement; and A photographic module includes an 8K line scan camera and a 3D line scanner, used to acquire surface defect images of the object to be inspected, including activating the 8K line scan camera and the 3D line scanner synchronously with the light source to acquire high-resolution 2D images and 3D point cloud data of the object to be inspected respectively. Among them, multiple sets of surface defect images of multiple objects to be detected, acquired by the surface defect detection system and labeled through a preprocessing procedure, form a surface defect image dataset.

2. The surface defect detection system of claim 1, wherein The surface defect image dataset is used to train a surface defect detection model that performs small target defect detection on the object to be detected.

3. The surface defect detection system of claim 2, wherein The surface defect detection model runs a modified YOLOv10 algorithm, and the surface defect image dataset provides the input data for the modified YOLOv10 algorithm.

4. The surface defect detection system of claim 3, wherein The modified YOLOv10 algorithm improves the efficiency of extracting key features of small target defects during small target defect detection by using depthwise separable convolution and point convolution design.

5. The surface defect detection system of claim 3, wherein Having obtained the surface defect image dataset, a data augmentation is performed to expand the surface defect image dataset, and a training set, a validation set, and a test set are established.

6. The surface defect detection system of claim 1, wherein The light source is a multi-band light source, and the light source controller controls the light source to flicker, synchronously driving the imaging module to acquire surface defect images of the object to be inspected.

7. The surface defect detection system of claim 6, wherein The multi-band light source integrates ultraviolet light, visible light, and infrared light, and detects various types of defects on the object to be detected by dynamically switching between different bands of the multi-band light source.

8. The surface defect detection system of claim 7, wherein The light source module incorporates a polarizer and a polarizing filter. The polarizer is used to adjust the polarization angle, thereby reducing light crosstalk to the highly reflective areas of the object under test.

9. The surface defect detection system of claim 1, wherein The surface defect detection platform is equipped with a color mark sensor for detecting the incoming material condition of the object to be detected.

10. The surface defect detection system of claim 9, wherein, The camera module and the light source module are activated based on the incoming material condition of the object to be inspected detected by the color mark sensor. The light source is then activated synchronously by the light source controller to acquire the surface defect image of the object to be inspected.

11. The surface defect detection system of claim 1, wherein The 3D line scanner uses a blue laser triangulation system to obtain a three-dimensional shape detection image of the object to be detected using the principle of laser triangulation and high-speed image processing technology.

12. The surface defect detection system of any one of claims 1 to 11, wherein The object to be inspected is a high-density interconnect printed circuit board, and the surface defect image includes at least the circuit board warpage, solder joint height, and microstructural defects.

Images