Method and system for detecting surface defects of sheet metal part before coating

Abstract

The invention discloses a method and system for detecting surface defects of a sheet metal part before coating, and the method comprises the following steps: S1, multi-modal data collection: synchronously collecting three-dimensional point cloud data and a hyperspectral image of the surface of the sheet metal part through a laser three-dimensional contourgraph and a hyperspectral camera, and recording environment illumination and temperature parameters at the same time; s2, data preprocessing: performing denoising and smoothing processing on the three-dimensional point cloud, extracting surface contour features, and performing defective pixel repair, waveband selection and reflectivity correction on the hyperspectral image; according to the surface defect detection method and system before coating of the sheet metal part, through multi-modal data acquisition and fusion of the three-dimensional morphology and the hyperspectral features, compared with traditional single visual detection, the surface defects of the sheet metal part can be more comprehensively and accurately recognized, particularly, tiny defects and oxide layer defects which are difficult to distinguish by naked eyes, the recognition rate is remarkably increased, and the detection accuracy is improved. And by adopting the lightweight deep learning model, the detection precision is ensured, and the reasoning speed is improved at the same time.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of surface defect detection, in particular to a surface defect detection method and system for sheet metal parts before coating. BACKGROUND

[0002] Surface defect detection of sheet metal parts before coating is a key link to ensure coating quality. During processing, transportation or storage, surface defects such as scratches, pits, bumps and cracks may occur on the sheet metal parts. If these defects are not detected and treated in time, the coating may appear blistering, peeling and orange peel phenomenon due to stress concentration or uneven surface at the defect site after coating, which seriously affects the appearance and corrosion resistance of the product. In addition, surface defects may also lead to uneven coating thickness, reducing the protective effect, and even causing the sheet metal part to fail prematurely due to local corrosion during use. Through surface defect detection before coating, these problems can be identified and repaired in advance to ensure good bonding between the coating and the substrate, improve the overall quality and reliability of the product, and meet the needs of customers for high-quality coating effect.

[0003] However, the existing detection technology generally adopts manual visual detection, which is low in efficiency, time-consuming for single detection, and has an identification rate of less than 60% for micro-defects below 0.1mm. It is greatly affected by the experience and fatigue of the detector. Traditional machine vision systems mostly use a single visible light camera, which is difficult to distinguish between the oxidation layer and the normal area of the metal surface. The quantization precision of three-dimensional defects such as pit depth is less than 0.1mm. At the same time, the existing detection scheme based on deep learning is mostly optimized for a single defect type, and the model size generally exceeds 200MB, which is difficult to deploy on edge devices in the production line. Moreover, it does not consider the problems of strong reflection and uneven texture specific to sheet metal part detection, and the false detection rate often exceeds 5% in actual application. Although multi-modal detection technology has been applied in some fields, it lacks a feature fusion strategy specific to the characteristics of sheet metal parts, and the spatio-temporal alignment error of three-dimensional data and spectral data can reach more than 5 pixels, affecting the detection accuracy. In addition, the existing system is mostly in offline detection mode and cannot be linked with the production line in real time, resulting in the continuous production of defective products due to the lag of detection results. Therefore, we propose a surface defect detection system for sheet metal parts before coating. SUMMARY

[0004] The purpose of the present application is to provide a surface defect detection method and system for sheet metal parts before coating to solve the problems of low precision, slow speed, poor environmental adaptability and insufficient linkage with the production line in the existing sheet metal surface defect detection.

[0005] To achieve the above purpose, the present application provides the following technical scheme: a surface defect detection method for sheet metal parts before coating, comprising the following steps: S1, multi-modal data acquisition: synchronously acquire the 3D point cloud data and hyperspectral image of the surface of the sheet metal part by laser 3D profiler and hyperspectral camera, and record the environmental light and temperature parameters.

[0006] S2, data preprocessing: denoise and smooth the 3D point cloud, extract the surface contour feature, repair bad points, select wavebands and correct reflectivity of the hyperspectral image.

[0007] S3, feature fusion: decompose the 3D feature and spectral feature by multi-scale multi-direction transform, fuse the low-frequency features by principal component analysis, extract the firing number of high-frequency features by pulse coupled neural network and perform weighted fusion, the specific steps are as follows: S31, perform non-subsampled contourlet transform on the 3D contour feature and the hyperspectral feature respectively to obtain the high-frequency directional subband and the low-frequency subband of each.

[0008] S32, calculate the eigenvalue weight of the low-frequency subband by principal component analysis, and obtain the low-frequency fusion subband by weighted summation.

[0009] S33, calculate the firing number of each pixel point in the high-frequency subband by pulse coupled neural network, assign the fusion weight based on the difference in firing number, and generate the high-frequency fusion subband.

[0010] S34, inverse transform and reconstruct the low-frequency fusion subband and the high-frequency fusion subband to obtain the fusion feature map.

[0011] S4, defect recognition: input the fusion feature into a lightweight deep learning model to output the defect type, position coordinates and size parameters.

[0012] S5, result feedback: send sorting signal or defect marking instruction to the production line control system according to the defect level determination result.

[0013] A sheet metal part surface defect detection system before painting, comprising a multi-modal data acquisition unit, an environment perception module, a data processing unit and a production line linkage module, the multi-modal data acquisition unit comprises a laser 3D profiler, a hyperspectral camera and a ring-shaped adaptive light source array, which is used to synchronously acquire 3D topographic data and hyperspectral image data of the surface of the sheet metal part, the environment perception module is provided with a light sensor and a temperature sensor, which acquires the detection environment parameters in real time, the data processing unit comprises a preprocessing module, a feature fusion module and a defect recognition module.

[0014] The preprocessing module performs point cloud denoising and contour extraction on the 3D topographic data, and performs waveband selection and radiation correction on the hyperspectral image.

[0015] The feature fusion module adopts an improved dual-channel attention mechanism to perform weighted fusion on the 3D geometric feature and the spectral feature.

[0016] The defect identification module deploys a lightweight deep learning model for classifying and identifying defects such as scratches, dents, oxide scale, and microcracks.

[0017] The production line linkage module is connected to the production line control system and can trigger marking or sorting actions based on defect detection results.

[0018] Preferably, the scanning accuracy of the laser three-dimensional profilometer is not less than 0.01 mm, and the spectral range of the hyperspectral camera covers the 400-1000 nm band, with a sampling interval ≤ 5 nm.

[0019] Preferably, the annular adaptive light source array includes eight independently controllable LED light sources. The brightness and color temperature of each light source can be adjusted in real time according to the feedback from the environmental sensing module, with an adjustment response time of ≤10ms.

[0020] Preferably, the lightweight deep learning model is based on the improved YOLO-DC SAM architecture, uses depthwise separable convolution to compress the network volume, adds a three-dimensional feature input branch, and has an inference speed of ≥30fps and a defect detection mAP of ≥96%.

[0021] Preferably, the system also includes a defect classification module, which classifies detected defects into three levels: minor, moderate, and severe, based on the size, depth, and distribution density of the defects. Minor defects are defined as isolated defects with a length of <2mm and a depth of <0.05mm.

[0022] Compared with the prior art, the beneficial effects of the present invention are: This method and system for detecting surface defects in sheet metal parts before painting utilizes multimodal data acquisition, integrating three-dimensional morphology and hyperspectral features. Compared to traditional single-vision inspection, it can more comprehensively and accurately identify surface defects in sheet metal parts, especially for minute defects and oxide layer defects that are difficult to distinguish with the naked eye, significantly improving the recognition rate. By employing a lightweight deep learning model, it improves inference speed while ensuring detection accuracy, meeting the real-time detection needs of the production line. The model's small size facilitates deployment on edge equipment of the production line. Through the cooperation of an environmental perception module and a ring adaptive light source array, the system can adapt to different lighting and temperature environments, ensuring stable imaging quality and reducing the impact of environmental factors on the detection results. The production line linkage module enables real-time interaction between the detection results and the production line, allowing for timely sorting or marking of defective workpieces, reducing the continuous generation of defective products and lowering production costs. The system adopts an edge computing + cloud optimization architecture, ensuring real-time data processing while continuously optimizing the model through the cloud to improve detection performance, exhibiting good scalability and adaptability. Attached Figure Description

[0023] Figure 1 Flow chart of the method of the present application; Figure 2 Schematic diagram of multi-modal feature fusion of the present application; Figure 3 System architecture diagram of the multi-modal data acquisition unit of the present application; Figure 4 System architecture diagram of the environment perception module of the present application; Figure 5 System architecture diagram of the data processing unit of the present application; Figure 6 System architecture diagram of the production line linkage module of the present application. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0025] Please refer to Figures 1-6 The present application provides a technical solution: a sheet metal part surface defect detection method before coating, comprising the following steps: S1, multi-modal data acquisition: three-dimensional point cloud data and hyperspectral images of the surface of the sheet metal part are synchronously acquired by a laser three-dimensional profiler and a hyperspectral camera, and environmental light and temperature parameters are recorded simultaneously, the laser three-dimensional profiler scans the surface of the sheet metal part at a frequency of 200 Hz, and 1024x768 point cloud data are obtained every frame; the hyperspectral camera synchronously acquires spectral images in the waveband of 400-1000 nm, the number of spectral channels is 204, the spatial offset between the laser scanning line and the spectral imaging line is kept ≤0.5 mm during the acquisition process, and the time synchronization error is <1 ms.

[0026] S2, data preprocessing: the three-dimensional point cloud is denoised and smoothed, and the surface contour feature is extracted, the hyperspectral image is repaired for bad points, selected for wavebands, and corrected for reflectivity, the point cloud preprocessing includes: 1, statistical filtering to remove outliers; 2, moving least squares method to smooth the surface; 3, calculation of the normal vector curvature to extract the surface mutation area. The hyperspectral preprocessing includes: 1, dark current correction to remove noise; 2, standard whiteboard correction for reflectivity; 3, SPA waveband selection to retain 15 key wavebands, and the data amount is reduced by 92.7%.

[0027] S3, feature fusion: the three-dimensional features and the spectral features are decomposed by multi-scale multi-directional transformation, the low-frequency features are fused by principal component analysis method, and the firing frequency of the high-frequency features is extracted by pulse coupled neural network and weighted fusion, and the specific steps are as follows.

[0028] S31. Non-subsampled profile wave transform is performed on the three-dimensional contour features and hyperspectral features respectively to obtain their respective high-frequency directional sub-bands and low-frequency sub-bands. Non-subsampled profile wave transform (NSCT) is used to decompose the three-dimensional contour feature map and hyperspectral feature map into three levels to obtain 3 low-frequency sub-bands and 15 high-frequency directional sub-bands.

[0029] S32, the eigenvalue weights of the low-frequency sub-bands are calculated using principal component analysis, and the low-frequency fusion sub-bands are obtained by weighted summation. The low-frequency fusion is performed using principal component analysis to calculate the eigenvalue contribution rate of the two-mode low-frequency sub-bands. The expression for weighted fusion is as follows: ; in, This refers to the low-frequency fusion subband. This refers to the eigenvalue weights of the low-frequency subband of the first mode. This refers to the low-frequency subband of the first mode. This refers to the eigenvalue weights of the low-frequency subband of the second mode. It refers to the low-frequency subband of the second mode, in which + =1, the weight is dynamically adjusted according to the feature value.

[0030] S33: The number of ignitions for each pixel in the high-frequency sub-band is calculated using a pulse-coupled neural network (PCNN). Fusion weights are assigned based on the differences in the number of ignitions to generate a high-frequency fused sub-band. High-frequency fusion is achieved using a PCNN, which calculates the number of ignitions for each pixel in the high-frequency sub-band. and Assign fusion weights according to the following expression: ; in, Refers to high-frequency subband pixels The fusion weights, Refers to the high-frequency sub-band pixels of the first mode. Number of ignitions Refers to the high-frequency sub-band pixels of the second mode. The number of ignitions; The final high-frequency fusion subband is generated according to the following expression: in, This refers to high-frequency fusion subband. This refers to the fusion weights of high-frequency sub-band pixels (i.e., the aforementioned). ), This refers to the first mode high-frequency subband. It refers to the second-mode high-frequency subband.

[0031] S34, inverse transform reconstruction is performed on the low-frequency fusion subband and the high-frequency fusion subband to obtain a fusion feature map, and a 256x256 fusion feature map is obtained through inverse NSCT transform.

[0032] S4, defect recognition: input the fusion feature into a lightweight deep learning model to output a defect type, a position coordinate and a size parameter; input the fusion feature map into the lightweight deep learning model, extract features through a backbone network (a lightweight version of CSPDarkNet53), enhance a multi-scale receptive field through an SPPM-PANet module, and finally output a detection head: 1. a defect category (scratch, depression, scale and micro-crack); 2. a position coordinate (x, y); and 3. a size parameter (length, width and depth).

[0033] S5, result feedback: according to a defect grade determination result, send a sorting signal or a defect marking instruction to a production line control system, and grade according to a preset standard: only record a slight defect (length < 2 mm and depth < 0.05 mm) without intervention; mark a moderate defect (2 mm ≤ length < 5 mm or 0.05 mm ≤ depth < 0.1 mm) for prompting; and trigger sorting for a serious defect (length ≥ 5 mm or depth ≥ 0.1 mm). All results are uploaded to an MES system in real time to form a quality traceability file.

[0034] A sheet metal part surface defect detection system before coating, comprising a multi-modal data acquisition unit, an environment perception module, a data processing unit and a production line linkage module, the multi-modal data acquisition unit includes a laser three-dimensional profiler, a hyperspectral camera and a ring-shaped adaptive light source array, which is used to synchronously acquire three-dimensional topographic data and hyperspectral image data of the surface of the sheet metal part, the environment perception module is configured with an illumination sensor and a temperature sensor, which acquires detection environment parameters in real time, the data processing unit includes a preprocessing module, a feature fusion module and a defect recognition module, the multi-modal data acquisition unit adopts a laser and spectrum dual-sensor architecture, the laser three-dimensional profiler adopts a line laser scanning mode, the scanning frequency is 200 Hz, and the lateral resolution is 0.02 mm, so that the depth information of three-dimensional defects such as depressions and protrusions can be accurately acquired, the hyperspectral camera adopts a push-broom imaging mode, the spectral resolution is 3 nm, and an 8 million pixel optical lens is matched, so that surface chemical or microstructure abnormalities such as oxide skin and microcracks can be captured, the ring-shaped adaptive light source array adopts eight groups of high-brightness LEDs, four groups of which are white polarized light sources (5600K), and four groups of which are near-infrared light sources (850nm), the light source control module realizes 0-100% brightness adjustment through PWM dimming technology, the environment perception module monitors the illumination intensity (0-10000 lux) and the environmental temperature (0-50℃) in real time, when the illumination fluctuation exceeds ±15%, the light source control module completes brightness compensation within 10 ms, and the imaging stability is ensured, the data processing unit adopts an 'edge computing + cloud optimization' architecture, the edge terminal is equipped with an NVIDIA Jetson AGXXavier processor, which is responsible for real-time data processing, and the cloud server is used for model training and updating.

[0035] The preprocessing module performs point cloud denoising and contour extraction on the three-dimensional topographic data, and performs band selection and radiation correction on the hyperspectral image, the three-dimensional point cloud is removed from the noise point by statistical filtering, the filtering radius is 0.5 mm, the threshold coefficient is 1.5, and the surface contour is extracted by the RANSAC algorithm; the hyperspectral image preprocessing includes: repairing bad points by using the interpolation method, selecting 15 characteristic bands by using the SPA algorithm, reducing the data amount while retaining the defect sensitive spectral information.

[0036] The feature fusion module adopts an improved double-channel attention mechanism to perform weighted fusion on three-dimensional geometric features and spectral features, the spatial attention branch focuses on the geometric mutation area of the three-dimensional contour, and the channel attention branch strengthens the defect sensitive bands in the spectral features. The fusion process adopts a multi-scale strategy, and a 256x256 pixel feature map is subjected to four-level pyramid decomposition, and different scale defects from 0.1 mm to 2 mm are processed.

[0037] The defect identification module deploys a lightweight deep learning model for classifying and identifying scratches, dents, scale and micro-cracks, the defect identification module is based on a lightweight network architecture, three-dimensional feature input channels are added on the basis of YOLO-DCSAM, deep separable convolution is used instead of traditional convolution, the model volume is compressed to 35MB, the inference speed reaches 45fps, the model training adopts a transfer learning strategy, is pre-trained on a public metal defect dataset first, and then is fine-tuned by 5000 labeled sheet metal defect images, finally, 96.8% mAP and 0.3% false detection rate are realized.

[0038] The production line linkage module is in communication connection with the production line control system, can trigger marking or sorting actions according to the defect detection result, the production line linkage module communicates with the production line PLC through an EtherCAT bus, when serious defects are detected, the sorting device is triggered within 100ms, and slight defects are marked by a laser marking machine, so that subsequent tracing is facilitated.

[0039] In the application, the scanning accuracy of the laser three-dimensional profilometer is not less than 0.01mm, the spectral range of the hyperspectral camera covers the 400-1000nm band, and the sampling interval is less than or equal to 5nm.

[0040] In the application, the annular adaptive light source array comprises eight groups of independently controllable LED light sources, the brightness and color temperature of each group of light sources can be adjusted in real time according to the feedback of the environment perception module, and the adjustment response time is less than or equal to 10ms.

[0041] In the application, the lightweight deep learning model is improved based on the YOLO-DCSAM architecture, deep separable convolution is used to compress the network volume, and a three-dimensional feature input branch is added, the model inference speed is greater than or equal to 30fps, and the defect detection mAP is greater than or equal to 96%.

[0042] In the application, the defect grading module is also included, the module divides the detected defects into three levels of slight, moderate and serious according to the size, depth and distribution density of the defects, wherein the slight defect is defined as an isolated defect with a length less than 2mm and a depth less than 0.05mm, the defect identification module is based on a lightweight network architecture, three-dimensional feature input channels are added on the basis of YOLO-DCSAM, deep separable convolution is used instead of traditional convolution, the model volume is compressed to 35MB, the inference speed reaches 45fps, the model training adopts a transfer learning strategy, is pre-trained on a public metal defect dataset NEU-DET first, and then is fine-tuned by 5000 labeled sheet metal defect images, finally, 96.8% mAP and 0.3% false detection rate are realized.

[0043] Working principle: The system is installed on the conveying line before sheet metal parts are coated. The detection area adopts a closed dustproof design to avoid external dust interference. The multi-modal data acquisition unit is installed 1.2 m above the conveying line. The center-to-center distance between the laser three-dimensional profiler and the hyperspectral camera is 50 mm. Synchronous acquisition is realized through a trigger signal. The ring light source is 300 mm away from the workpiece surface to ensure that the light uniformity is > 90%. When the sheet metal part enters the detection area along with the conveying line (speed 1.5 m / s), the photoelectric sensor triggers the acquisition process: the laser profiler acquires 200 frames of point cloud data per second, the hyperspectral camera synchronously acquires spectral images, and the environmental perception module updates the light and temperature data 10 times per second. The data is transmitted to the edge computing unit through a gigabit Ethernet. The pre-processing module completes single-frame data processing within 50 ms. In the feature fusion process, a dynamic weight adjustment mechanism based on the material of the sheet metal part is introduced for the weight distribution of the low-frequency subband. The two modal low-frequency subbands are weighted and fused according to the preset weight distribution plan to obtain a low-frequency fusion subband. wherein the expression of weight distribution is as follows: ; wherein, is the low-frequency fusion subband, is the feature value weight of the first modal low-frequency subband, is the first modal low-frequency subband, is the feature value weight of the second modal low-frequency subband, is the second modal low-frequency subband; and is a dynamic adjustment factor based on the material of the sheet metal part, satisfying , and the specific values are as follows: for aluminum alloy parts, , ; for cold-rolled steel sheets, , .

[0044] The lightweight model is implemented on the edge end to achieve an inference speed of 45fps, and the detection result is sent to the production line PLC in real time through the EtherCAT bus. When the system is running, calibration is automatically performed once every 8 hours: a standard defect sample containing scratches and depressions of known size is used, and the three-dimensional scale factor and spectral reflectivity benchmark are corrected through an iterative optimization algorithm to ensure long-term detection accuracy. Every month, 100,000+ defect samples accumulated on the edge end are uploaded to the cloud to update the model parameters through federated learning and continuously optimize the detection performance. Through the combination of multi-modal data fusion and lightweight intelligent algorithm, the present application realizes high-precision and real-time detection of defects on sheet metal parts before painting, with a detection accuracy of 96.8%, a false detection rate of <0.5%, and a single-piece detection time of <200ms, meeting the detection needs of high-speed production lines, significantly reducing the rate of defective products after painting, and having important engineering application value.

[0045] In summary: the sheet metal part surface defect detection method and system before painting, through multi-modal data acquisition, fusion of three-dimensional morphology and hyperspectral characteristics, compared with traditional single visual detection, can more comprehensively and accurately identify the surface defects of sheet metal parts, especially the micro defects and the defects of the oxide layer that are difficult to distinguish with the naked eye, the recognition rate is significantly improved. By using a lightweight deep learning model, the inference speed is improved while ensuring detection accuracy, which can meet the real-time detection needs of the production line. The model is small in size and easy to deploy on the edge equipment of the production line. Through the cooperation of the environmental perception module and the annular adaptive light source array, the system can adapt to different lighting and temperature environments, ensure stable imaging quality, and reduce the influence of environmental factors on the detection result. The production line linkage module is used to realize real-time interaction between the detection result and the production line, so that defective workpieces can be sorted or marked in time, reducing the continuous production of defective products and reducing production costs. The system uses edge computing + cloud optimization architecture to ensure real-time data processing and continuously optimize the model through the cloud to improve detection performance, with good scalability and adaptability.

[0046] It should be noted that, in this text, relational terms such as first and second are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article or device.

[0047] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A method for detecting surface defects of sheet metal parts before painting, characterized in that: Includes the following steps: S1, Multimodal data acquisition: Simultaneously acquire three-dimensional point cloud data and hyperspectral images of the sheet metal surface using a laser three-dimensional profilometer and a hyperspectral camera, while recording ambient light and temperature parameters. S2, Data preprocessing: Denoising and smoothing the 3D point cloud and extracting surface contour features; performing bad pixel repair, band selection and reflectance correction on the hyperspectral image. S3, Feature Fusion: Multi-scale, multi-directional transformation decomposes 3D and spectral features; principal component analysis is used to fuse low-frequency features; and pulse-coupled neural networks are used to extract the ignition count of high-frequency features and perform weighted fusion. The specific steps are as follows: S31, perform non-subsampled contour wave transformation on the three-dimensional contour features and hyperspectral features respectively to obtain their respective high-frequency directional sub-bands and low-frequency sub-bands; S32, the principal component analysis method is used to calculate the eigenvalue weights of the low-frequency sub-bands, and the low-frequency fusion sub-bands are obtained by weighted summation. S33, calculates the number of ignitions of each pixel in the high-frequency sub-band through a pulse-coupled neural network, and assigns fusion weights based on the difference in the number of ignitions to generate a high-frequency fused sub-band; S34, perform inverse transform reconstruction on the low-frequency fusion sub-band and the high-frequency fusion sub-band to obtain the fusion feature map; S4, Defect Identification: Input the fused features into a lightweight deep learning model and output the defect type, location coordinates, and size parameters; S5, Result Feedback: Based on the defect level determination result, send a sorting signal or defect marking instruction to the production line control system.

2. A surface defect detection system for sheet metal parts before painting, used to implement the detection method described in claim 1, characterized in that, It includes a multimodal data acquisition unit, an environmental perception module, a data processing unit, and a production line linkage module. The multimodal data acquisition unit includes a laser 3D profilometer, a hyperspectral camera, and a ring adaptive light source array, used to simultaneously acquire 3D morphological data and hyperspectral image data of the sheet metal surface. The environmental perception module is equipped with a light sensor and a temperature sensor to collect and detect environmental parameters in real time. The data processing unit includes a preprocessing module, a feature fusion module, and a defect identification module. The preprocessing module performs point cloud noise reduction and contour extraction on the three-dimensional topography data, and performs band selection and radiometric correction on the hyperspectral image. The feature fusion module employs an improved dual-channel attention mechanism to perform weighted fusion of three-dimensional geometric features and spectral features; The defect identification module deploys a lightweight deep learning model for classifying and identifying defects such as scratches, dents, oxide scale, and microcracks. The production line linkage module is connected to the production line control system and can trigger marking or sorting actions based on defect detection results.

3. The sheet metal part surface defect detection system before painting according to claim 2, characterized in that: The scanning accuracy of the laser 3D profilometer is not less than 0.01 mm, and the spectral range of the hyperspectral camera covers the 400-1000 nm band with a sampling interval of ≤5 nm.

4. The sheet metal part surface defect detection system before painting according to claim 2, characterized in that: The ring-shaped adaptive light source array contains eight independently controllable LED light sources. The brightness and color temperature of each light source can be adjusted in real time according to the feedback from the environmental sensing module, with an adjustment response time of ≤10ms.

5. The sheet metal part surface defect detection system before painting according to claim 2, characterized in that: The lightweight deep learning model is based on the improved YOLO-DC SAM architecture, uses depthwise separable convolution to compress the network volume, adds a three-dimensional feature input branch, and has an inference speed of ≥30fps and a defect detection mAP of ≥96%.

6. The sheet metal part surface defect detection system before painting according to claim 2, characterized in that: It also includes a defect classification module, which classifies detected defects into three levels: minor, moderate, and severe based on the size, depth, and distribution density of the defects. Minor defects are defined as isolated defects with a length of <2mm and a depth of <0.05mm.

Images