Induction cooker production line part defect detection method based on image recognition

Abstract

The invention discloses an induction cooker production line part defect detection method based on image recognition, and relates to the technical field of industrial automatic detection, and the method comprises the following steps: S1, triggering different illumination units through a multi-light source illumination system in a time-sharing manner to collect a plurality of images of a part under different illumination conditions; and S2, inputting the image and the corresponding illumination parameters into a deep learning model to obtain a part surface intrinsic attribute graph. According to the induction cooker production line part defect detection method based on image recognition, through the technical means of multi-light-source time-sharing imaging and deep learning extraction of surface intrinsic attributes, the problem that imaging of high-reflection curved surface parts is inconsistent in a complex illumination environment is effectively solved, and the accuracy and environmental adaptability of defect detection are improved. The texture and geometry two-dimensional independent detection is combined with an arbitration judgment mechanism, so that false detection and missing detection are reduced, the reliability of a detection result is ensured, and the strict requirement of industrial production on quality control is met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial automation detection, in particular to a method for detecting defects of electromagnetic oven production line parts based on image recognition. BACKGROUND

[0002] In the field of industrial defect detection based on image recognition, technology has been widely applied to quality monitoring on production lines. The surface defect detection of electromagnetic oven production line parts, such as microcrystalline panels and metal coil discs, relies on clear image acquisition. However, these parts generally have high reflectivity and complex curved surface structures, combined with uncontrollable fluctuations in the lighting conditions of the production site environment, such as changes in natural light, interference from other equipment light sources, and decay of the lighting device itself, which poses a serious challenge to stable imaging. The uneven distribution of light, the formation of glints caused by local overexposure, and the shadow blocking result in poor consistency of the features of the collected part images, and the same defect presents significant differences under different light. The existing technology attempts to alleviate this problem by constructing a closed lighting environment or using post-algorithm enhancement, but the former has high implementation costs and insufficient flexibility, and the latter is difficult to recover the details lost in the imaging stage, ultimately causing an increase in the false positive rate of the detection system and a decrease in robustness. Therefore, the technical problem to be solved at present is how to overcome the defect recognition reliability problem caused by inconsistent imaging of high-reflectivity curved surface parts under complex and variable lighting environments. SUMMARY

[0003] The present application aims to provide a method for detecting defects of electromagnetic oven production line parts based on image recognition to solve the problems raised in the background. To solve the above technical problems, the present application provides the following technical scheme: a method for detecting defects of electromagnetic oven production line parts based on image recognition, comprising the following steps:

[0004] S1: Collect multiple images of the part under different lighting conditions by triggering different lighting units of a multi-light source lighting system at different times;

[0005] S2: Input the images and corresponding lighting parameters into a deep learning model to obtain a part surface intrinsic attribute map;

[0006] S3: Based on the intrinsic attribute map, make defect judgments from the texture attribute and geometric attribute dimensions respectively;

[0007] S4: Fuse and arbitrate the judgment results of the two dimensions and output the final defect determination result.

[0008] Preferably, the multi-light source lighting system includes a group of ring-distributed lighting units, each group of lighting units having different incident angles and spectral wavelength configurations.

[0009] Preferably, the time-sharing trigger adopts a sequential trigger mode, and the trigger interval time of adjacent lighting units is determined based on the part conveying speed.

[0010] Preferably, the deep learning model is trained by a training set containing synthetic training data, and the synthetic training data is generated by computer graphics rendering technology and contains surface intrinsic attribute true value labels.

[0011] Preferably, the surface intrinsic attribute map includes a reflectivity attribute map and a normal direction map, wherein the reflectivity attribute map is used to characterize the surface material properties, and the normal direction map is used to characterize the surface geometry.

[0012] Preferably, the texture attribute dimension defect judgment adopts a convolutional neural network to analyze the reflectivity attribute map and identify surface scratches and material defects.

[0013] Preferably, the geometric attribute dimension defect judgment is performed by gradient calculation and morphological processing on the normal direction map to identify surface concave-convex and deformation defects.

[0014] Preferably, the fusion arbitration adopts a double-threshold decision mechanism, and when the confidence scores in the same region in the two dimensions both exceed the corresponding threshold, it is determined as a defect.

[0015] An electronic device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of a kind of electromagnetic furnace production line part defect detection method based on image recognition when executing the program.

[0016] A computer readable storage medium stores a computer program, and the computer program implements the steps of a kind of electromagnetic furnace production line part defect detection method based on image recognition when executed by a processor.

[0017] The present application provides a kind of electromagnetic furnace production line part defect detection method based on image recognition. With the following beneficial effects:

[0018] The electromagnetic furnace production line part defect detection method based on image recognition effectively overcomes the problem of inconsistent imaging of high-reflective curved surface parts under complex lighting environment by using the technical means of multi-light source time-sharing imaging and deep learning to extract surface intrinsic attributes, improves the accuracy and environmental adaptability of defect detection. The mechanism of texture and geometry dual-dimension independent detection and arbitration determination is adopted, which reduces the occurrence of false detection and missed detection, ensures the reliability of the detection result, and meets the strict requirements of quality control in industrial production.

[0019] The image recognition-based induction cooker production line part defect detection method realizes automatic online detection of surface defects of induction cooker parts, reduces dependence on manual inspection, and improves the detection efficiency and consistency of the production line. Through the cooperative optimization of hardware control and algorithm processing, the method has good engineering applicability and scalability, and provides an effective technical solution for quality detection of similar industrial surfaces. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 A system module interaction diagram of the image recognition-based induction cooker production line part defect detection method of the present application is shown in the figure.

[0021] Figure 2 A flowchart of the image recognition-based induction cooker production line part defect detection method of the present application is shown in the figure. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0023] Please refer to Figure 1 and Figure 2 , the present application provides a technical solution: an image recognition-based induction cooker production line part defect detection method, comprising the following steps:

[0024] S1: acquire multiple images of the part under different illumination conditions by time-triggering different illumination units through a multi-light source illumination system;

[0025] S2: input the images and corresponding illumination parameters into a deep learning model to obtain a part surface intrinsic attribute map;

[0026] S3: judge defects based on the intrinsic attribute map from the texture attribute and geometric attribute dimensions respectively;

[0027] S4: fuse and arbitrate the judgment results of the two dimensions and output the final defect judgment result.

[0028] It needs to be further explained that in the specific implementation process, when implementing the image recognition-based induction cooker production line part defect detection method, first, the multi-light source lighting system deployed above the detection station is used to collect images of the parts conveyed to the detection position. The system includes multiple independently controlled lighting units arranged in a ring shape and having different lighting angles and spectral characteristics. During the collection process, the system activates each lighting unit in a predetermined order and synchronously triggers an industrial camera to take a part image when each lighting unit is turned on, thereby obtaining a series of images of the same part under different groups of lighting conditions. These images and the corresponding lighting unit parameters, including the light source angle, wavelength, and intensity information of the unit, are recorded and transmitted to the processing system.

[0029] Subsequently, the collected multiple original images and their corresponding lighting parameters are combined and input into a pre-trained deep learning model for processing. The model, trained on a large number of samples, can analyze the intrinsic properties of the part surface from the input multi-light source image sequence. By analyzing the differences in the same surface under different lighting conditions, the model eliminates the interference caused by environmental lighting changes and outputs intrinsic attribute maps that represent the physical characteristics of the part itself, including reflectivity attribute maps that reflect surface material properties and normal direction maps that reflect surface geometry.

[0030] Based on the obtained surface intrinsic attribute maps, the system analyzes and judges defect features from two independent dimensions respectively. In the texture attribute dimension, the reflectivity attribute map is used to analyze the continuity of the surface material through image segmentation algorithms to identify whether there are scratches, stains, or material unevenness, etc. texture class defects, and generate a first judgment result containing defect location and confidence score. In the geometric attribute dimension, the normal direction map is used to calculate the change gradient of the surface normal direction and perform morphological analysis to detect whether there are concave-convex, warping, or deformation, etc. geometric abnormalities, and generate the corresponding second judgment result.

[0031] Finally, the system fuses and arbitrates the judgment results generated from the two dimensions. This process uses a double-threshold decision mechanism to match the defect regions identified in the two dimensions in terms of location. Only when the same region is identified as a defect in both judgment results and the confidence scores of each are above the corresponding preset threshold, is it finally determined that the region has a real defect. The final output result will mark the location information confirmed as a defect, and can generate a detection report or trigger a sorting mechanism as needed. The entire processing process is completed within the production line beat, realizing online automatic detection of part defects.

[0032] The multi-source lighting system comprises a ring-distributed set of lighting units, each set having different incident angle and spectral wavelength configurations. It is further noted that in the specific implementation, the multi-source lighting system adopts a ring-distributed structure, which is composed of multiple sets of lighting units, each set having specific spatial configuration and optical characteristics. The lighting units are evenly distributed in a ring shape above the part detection area, and the installation angle of each lighting unit is precisely calculated to ensure that the part surface can be illuminated from different directions. Each set of lighting units is configured with different light source incident angles, including low-angle lighting units, high-angle lighting units, and side lighting units. These different angle lighting configurations can highlight surface concave and convex defects, planar flaws, and edge features, respectively. At the same time, the lighting units are configured with multi-spectral light sources, including visible light bands and specific wavelength near-infrared light sources. Different wavelengths of light sources can be optimized for the reflection characteristics of different materials on the part surface, enhancing the contrast of specific types of defects.

[0033] The control unit of the lighting system can independently adjust the brightness and flash duration of each lighting unit, and pre-store multiple lighting modes according to the surface characteristics of different types of parts, and automatically select the most suitable lighting combination scheme during detection. This multi-angle multi-spectral lighting configuration can effectively overcome environmental light interference, eliminate specular reflection caused by high-reflective surfaces, and obtain consistent image quality under different lighting conditions, providing reliable input data for subsequent image processing.

[0034] The time-sharing triggering adopts a sequential triggering mode, and the time interval between triggering of adjacent lighting units is determined based on the part conveying speed. It is further noted that in the specific implementation, the time-sharing triggering mode adopts a sequential triggering mechanism that dynamically adjusts the triggering timing of the lighting units according to the actual conveying speed of the parts on the production line. When the part enters the detection station, the speed data of the conveying belt is obtained in real time through the encoder, and the control system calculates the optimal time interval for triggering of adjacent lighting units based on the speed value. The triggering sequence is performed in order according to the physical position of the ring-shaped lighting units, ensuring that each lighting unit is activated when the part is in the optimal detection position. In high-speed production conditions, the system automatically shortens the triggering interval and adopts overlapping triggering, i.e., the next lighting unit is started before the current lighting unit is completely turned off, to ensure that sufficient images under different lighting conditions are obtained within a limited time.

[0035] In low speed or pause working condition, the system prolongs the irradiation time of each lighting unit and increases the frame number of image acquisition to improve the image signal-to-noise ratio. For parts with different surface characteristics, the system pre-stores multiple trigger sequence modes. For high-reflective parts, the trigger sequence is first side light and then top light. For parts with complex surface texture, the mode of alternating trigger of different spectral light sources is adopted. The flash duration of each lighting unit is independently adjusted according to the part surface reflection characteristics and camera exposure requirements, ensuring that consistent exposure images can be obtained under different trigger intervals. This adaptive trigger mechanism based on conveying speed can effectively avoid image blur caused by part movement, while ensuring the synchronization and consistency of multi-light source image acquisition.

[0036] The deep learning model is trained by a training set containing synthetic training data, which is generated by computer graphics rendering technology and contains surface intrinsic attribute true value labels. It needs to be further explained that in the specific implementation process, the deep learning model is trained by a training set containing a large amount of synthetic training data in a supervised learning manner, wherein the synthetic training data is generated by computer graphics rendering technology. The generation process is based on the physical reflection model and geometric characteristics of the surface of the induction cooker part. By adjusting the lighting conditions, material parameters and surface defect morphology in the virtual scene, high-realistic synthetic images with different lighting angles, spectral characteristics and defect types are generated.

[0037] Each synthetic image is accompanied by accurate surface intrinsic attribute true value labels, including reflectivity attribute maps and normal direction maps, which are directly generated by the rendering engine during the rendering process. In the training process, the model learns the mapping relationship from the multi-light source input image sequence to the surface intrinsic attribute map, and optimizes the reconstruction loss function to make the predicted intrinsic attribute map as close as possible to the true value label. In order to enhance the generalization ability of the model, the training data also contains labeled real production line acquisition images, which are mixed with synthetic images in proportion.

[0038] In the training phase, a multi-task learning framework is used to simultaneously optimize reflectivity attribute prediction and normal direction prediction. The two tasks share the feature extraction layer but have independent decoder branches. After the model is trained, it can effectively handle the image differences caused by changes in lighting, reflection characteristics and other factors in the actual production environment, and accurately estimate the intrinsic attributes of the part surface. This training method based on the combination of synthetic data and real data not only solves the problem of difficulty in obtaining real defect samples, but also ensures the reliability of the model in actual application.

[0039] The surface intrinsic attribute map includes a reflectance attribute map and a normal direction map, wherein the reflectance attribute map is used to characterize the surface material characteristics, and the normal direction map is used to characterize the surface geometry. It needs to be further explained that in the specific implementation process, the surface intrinsic attribute map contains two core components of the reflectance attribute map and the normal direction map. The reflectance attribute map is generated by analyzing the change rule of pixel intensity in the multi-light source image sequence under different lighting conditions, and each pixel value of the reflectance attribute map represents the inherent reflection characteristics of the corresponding surface point under various lighting conditions, which can effectively distinguish the differences between materials and is not affected by environmental lighting.

[0040] The normal direction map is constructed by calculating the three-dimensional orientation information of each point on the surface, using the light and dark change pattern generated under different direction light sources to calculate the direction vector of the surface normal in three-dimensional space, thereby accurately describing the micro geometry of the surface. The generation processes of the two attribute maps are mutually coordinated but independent. The reflectance attribute map is mainly used to detect defects related to material characteristics, such as surface contamination, coating peeling or material unevenness; and the normal direction map is focused on identifying defects related to surface geometry, such as scratches, depressions, protrusions or deformations.

[0041] In actual processing, the system will analyze both attribute maps, and find abnormal conditions by comparing the attribute consistency of adjacent regions. For electromagnetic stove parts with complex curved surface structure, this method can effectively overcome the feature distortion problem caused by curved surface reflection in traditional two-dimensional image analysis, and realize comprehensive evaluation of the surface quality of the parts. This analysis method based on physical intrinsic attributes not only improves the accuracy of defect detection, but also enhances the adaptability to different lighting conditions, ensuring the consistency and reliability of the detection results.

[0042] The texture attribute dimension defect judgment uses a convolutional neural network to analyze the reflectance attribute map to identify surface scratches and material defects. It needs to be further explained that in the specific implementation process, the texture attribute dimension defect judgment uses a convolutional neural network to analyze and process the reflectance attribute map, which is specially trained to identify abnormal changes in surface materials. The network input is the reflectance attribute map, which extracts features through multiple convolutional layers, with the bottom layer capturing local texture patterns and the high layer integrating global context information.

[0043] For scratch defects, the network learns to identify linear continuous low reflectance abnormal areas, and judges by analyzing the differences in texture directionality and continuity features with the surrounding normal areas. For stains or material inconsistency defects, the network detects patch-like areas with differences in reflectance values from surrounding areas, and evaluates the boundary definition and internal uniformity of the area. The network output includes a defect probability map for each pixel point, and then a threshold processing generates a binary initial defect mask.

[0044] Post-processing operations are performed on the mask, including removing noise points with too small areas, connecting disconnected defect segments, and classifying according to defect morphology characteristics. For each detected defect region, the system calculates its area, aspect ratio, contour complexity, and other characteristic parameters, and combines the confidence score output by the network for comprehensive evaluation. The entire process can effectively distinguish between real surface defects and pseudo-defects caused by noise or normal texture changes, improving the reliability of the detection results. The method uses the illumination invariant property of the reflectance attribute map to ensure stable detection performance under different illumination conditions.

[0045] Geometric attribute dimension defect judgment identifies surface concave-convex and deformation defects by gradient calculation and morphological processing of the normal direction map. It needs to be further explained that in the specific implementation process, the geometric attribute dimension defect judgment realizes the detection of surface geometric defects by gradient calculation and morphological processing of the normal direction map. The processing process first calculates the gradient amplitude and direction of each pixel point in the normal direction map, and identifies abnormal areas of surface geometry by analyzing the gradient change characteristics. For concave-convex defects, the system detects areas where the normal direction changes abruptly, and distinguishes between convex and concave defects by evaluating the distribution characteristics of the gradient amplitude. For warping or deformation defects, the continuity and consistency change pattern of the normal direction in a larger area is analyzed to identify the overall geometric deformation of the surface.

[0046] In the processing process, a multi-scale analysis method is used. First, a larger scale structural element is used for morphological opening and closing operation to eliminate noise interference and extract the overall contour features of the defects. Then, a smaller scale structural element is used for fine processing to retain the detail information of the defects.

[0047] For the detected potential defect regions, further calculate their geometric feature parameters, including region curvature distribution, boundary regularity, and depth variation trend. According to these characteristic parameters, the defects are classified and severity evaluated, and combined with the area and spatial distribution characteristics of the defect region for comprehensive judgment. This method makes full use of the three-dimensional geometric information provided by the normal direction map, and can effectively detect various types of surface geometric defects, and accurately describe and classify the morphological characteristics of the defects.

[0048] The fusion arbitration adopts a double-threshold decision mechanism. When the confidence scores of the same region in both dimensions exceed the corresponding threshold, it is determined to be a defect. It needs to be further explained that in the specific implementation process, the fusion arbitration process uses a double-threshold decision mechanism to comprehensively analyze the defect judgment results of the two dimensions. This mechanism first performs position registration on the first defect judgment result generated by the texture attribute dimension and the second defect judgment result generated by the geometric attribute dimension to ensure that the correspondence between the same physical region in different dimension detection results is accurately established.

[0049] For each detected region, the system separately obtains its confidence score in two-dimensional detection, and compares these scores with preset independent thresholds. The texture dimension defect judgment adopts a confidence evaluation method based on regional features, considering features such as reflectivity consistency, boundary definition, and texture continuity of the defect region; the geometric dimension defect judgment calculates the confidence degree according to geometric features such as surface curvature change, gradient distribution characteristics, and shape regularity. Two dimensions are set with independent threshold standards, and these thresholds are obtained through a large amount of experimental data statistics, and are dynamically adjusted according to the detection requirements of different types of defects.

[0050] Only when a certain region is identified as a defect in the detection results of both dimensions, and the respective confidence scores exceed the threshold values of the corresponding dimensions, is the region finally determined to have a real defect. For defects detected in only one dimension, the system will mark it as a region to be reviewed, and start the artificial review process. This double verification mechanism effectively reduces the false positives that may be generated by a single detection method, while ensuring that real defects are not missed, improving the accuracy and reliability of the detection system.

[0051] An electronic device includes a memory and a processor, the memory storing a computer program, and the processor implementing the steps of an electromagnetic stove production line part defect detection method based on image recognition when executing the program. It needs to be further explained that in the specific implementation process, the electronic device contains a memory and a processor, and the memory stores a computer program. When the processor executes the program, it realizes the electromagnetic stove production line part defect detection method based on image recognition. The device is equipped with an image acquisition interface for connecting a multi-light source lighting system and an industrial camera to receive a sequence of part images collected under different lighting conditions. The processor is configured with a special image processing module responsible for preprocessing the input images, including noise suppression and image registration, to ensure the accuracy of subsequent analysis. The built-in deep learning inference engine of the device loads a pre-trained neural network model that can parse the intrinsic attribute map of the part surface from the multi-light source image sequence.

[0052] The processing process includes two parallel defect detection branches: one branch analyzes the surface texture properties, and the other branch evaluates the surface geometric properties. The device is also provided with a decision fusion unit responsible for comprehensive arbitration of the output results of the two detection branches, and adopts a double-threshold decision mechanism to confirm the final defect detection result. The device provides a data output interface that can transmit the detection results to the production line control system, and has a human-computer interaction interface for parameter setting and result display. The entire processing process is completed inside the device, realizing the automatic detection of electromagnetic stove part defects without human intervention, ensuring the consistency and reliability of detection efficiency.

[0053] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of an image recognition-based induction cooker production line part defect detection method. It needs to be further explained that in the specific implementation process, the computer readable storage medium stores a computer program, and the program is executed by the processor to implement the image recognition-based induction cooker production line part defect detection method. The program instructions stored in the storage medium include an image acquisition control module for coordinating the time-sharing triggering of a multi-light source lighting system and the image acquisition of an industrial camera.

[0054] The program also includes an image preprocessing routine that calibrates and registers the acquired multi-light source image sequence to ensure consistency of the image data. The core processing module includes a trained deep learning model that can parse the surface intrinsic attribute map from the multi-light source input image. The program has parallel processing channels for texture defect analysis of the reflectance attribute map and geometric defect detection of the normal direction map. The decision arbitration module implements a double-threshold decision mechanism to fuse the results of the two detection channels.

[0055] The program also includes a result output interface that generates a detection report containing defect location and type and can trigger subsequent sorting operations. The program code in the storage medium is optimized to enable real-time processing under the computational resource constraints of an industrial controller, meeting the beat requirements of the production line detection. The entire program architecture uses modular design, and the functional modules exchange data through clearly defined interfaces, ensuring the stability and maintainability of the system.

[0056] It needs to be further explained that in the specific implementation process, the parts on the induction cooker production line are transported to the detection station by the conveyor belt, at which time the multi-light source lighting system installed above the station starts to work. The lighting system adopts a ring distribution structure and is composed of multiple independently controlled lighting units, each with specific incident angles and spectral characteristics. The lighting units are activated in a predetermined order, and the industrial camera synchronously acquires a part image when each lighting unit is turned on, thereby obtaining a series of images of the same part under different lighting conditions. During the image acquisition process, the system records the light source parameters of each lighting unit, including lighting angle, wavelength, and intensity information. This multi-light source time-sharing imaging method can effectively capture the performance characteristics of the part surface under different lighting conditions, providing rich input data for subsequent processing.

[0057] The collected multiple original images and their corresponding illumination parameters are input into a specially trained deep learning model. The model uses an encoder-decoder structure and can parse the intrinsic properties of the part surface from the multi-light source image sequence. A large amount of synthetic data is used in the model training process, which is generated by computer graphics rendering technology, simulating the performance of different surface characteristics under various lighting conditions. The synthetic data contains accurate surface intrinsic property true value labels, enabling the model to learn the mapping relationship from multi-light source images to surface intrinsic properties. The model output includes two important intrinsic property maps: the reflectance attribute map and the normal direction map. The reflectance attribute map represents the inherent reflection characteristics of the material surface and is not affected by the ambient light; the normal direction map represents the three-dimensional orientation information of each point on the surface, reflecting the geometric feature of the surface.

[0058] Based on the obtained surface intrinsic property map, the system detects defects from two independent dimensions respectively. In the texture attribute dimension, a convolutional neural network is used to analyze the reflectance attribute map to identify defects such as surface scratches, stains, or material inconsistencies. The network learns the normal surface texture features and detects abnormal areas with different reflectance from the surrounding area. In the geometric attribute dimension, the gradient calculation and morphological processing of the normal direction map are used to detect geometric defects such as surface concave-convex, warping, or deformation. This processing process analyzes the change pattern of the surface normal direction and identifies areas that do not conform to the normal geometric features.

[0059] The detection results of the two dimensions then enter the fusion arbitration stage. This process uses a double-threshold decision mechanism. First, the location matching of the defect areas detected in the two dimensions is performed to ensure the accurate establishment of the correspondence of the same physical area in the detection results of different dimensions. For each detected area, the system obtains its confidence score in the texture dimension and the geometric dimension respectively. The texture dimension confidence score is calculated based on features such as reflectance consistency, boundary clarity, and texture continuity; the geometric dimension confidence score is calculated based on geometric features such as surface curvature change, gradient distribution characteristics, and morphological regularity. Independent threshold standards are set for the two dimensions, and these thresholds are obtained through a large number of experimental data statistics. Only when a certain area is identified as a defect in the detection results of both dimensions and the confidence scores of each dimension exceed the corresponding dimension threshold, is the area finally determined to have a real defect.

[0060] The entire detection process is completed within the production line beat, realizing online automatic detection of part defects. The system can adapt to different lighting environment changes, effectively overcome the detection difficulties caused by high-reflective surfaces, and improve the accuracy and reliability of defect detection. The detection results can be output to the production line control system in real time to trigger the sorting mechanism or generate a quality detection report.

[0061] The method obtains rich surface information through multi-light source active imaging technology, extracts illumination-invariant intrinsic features using a deep learning model, and uses a two-dimensional verification mechanism to ensure the accuracy of the detection results, effectively solving the problem of poor consistency of traditional detection methods in detecting high-reflective part surface defects under complex lighting conditions.

[0062] It should be further explained that in the specific implementation process, the electromagnetic oven production line parts are transported to the detection station by the conveying device, and the multi-light source lighting system installed above the detection station starts to work. The lighting system adopts a ring distribution structure and is composed of multiple independently controlled lighting units. Each lighting unit has a specific spatial position, incident angle and spectral characteristics. The lighting units are activated in a predetermined order, and an industrial camera synchronously captures a part image when each lighting unit is turned on, thereby obtaining a series of images of the same part under different lighting conditions. During image acquisition, the system records the light source parameters of each lighting unit, including lighting angle, wavelength and intensity information. This multi-light source time-sharing imaging method can effectively capture the performance characteristics of the part surface under different lighting conditions.

[0063] The collected multiple original images and their corresponding lighting parameters are input into a specially trained deep learning model. The model uses an encoder-decoder structure and can analyze the intrinsic properties of the part surface from the multi-light source image sequence. During model training, a combination of synthetic data and real data is used, where synthetic data is generated through computer graphics rendering technology to simulate the performance of different surface characteristics under various lighting conditions. The model output includes two important intrinsic attribute maps: reflectivity attribute map and normal direction map.

[0064] The generation of the reflectivity attribute map is based on the Lambertian reflectance model, the basic principle of which can be represented as: ; wherein, represents the observed image intensity, represents the surface reflectivity, represents the surface normal vector, represents the light source direction vector, represents the vector dot product, i.e. , is the angle between the normal and the light source direction; to ensure that the dot product result is non-negative. Through observations under multiple light source conditions, the reflectivity attribute of the surface can be solved.

[0065] The generation of the normal direction map uses photometric stereo vision technology and solves the following overdetermined equation set by the least squares method: ; wherein, represents the light source direction matrix, represents the surface normal vector, represents the observed image intensity, represents the surface reflectance; the actual solution is obtained by least squares: .

[0066] Based on the obtained surface intrinsic attribute maps, the system conducts defect detection from two independent dimensions respectively. In the texture attribute dimension, a convolutional neural network is used to analyze the reflectance attribute map to identify defects such as surface scratches, stains or material inconsistencies. The network structure includes multiple convolutional layers and pooling layers, and finally outputs a defect probability map through a fully connected layer.

[0067] In the geometric attribute dimension, the gradient calculation and morphological processing of the normal direction map are used to detect geometric defects such as surface concave-convex, warping or deformation. The gradient calculation uses Sobel operators to calculate the partial derivatives of the normal direction in the and directions respectively: , ; wherein, , are the gradients in the and directions respectively, is the partial derivative symbol, , are the components of the normal vector in the and directions respectively. Through the analysis of the gradient amplitude and direction, the geometric abnormal area of the surface can be identified.

[0068] The detection results of the two dimensions enter the fusion arbitration stage. This process uses a double-threshold decision mechanism. First, the positions of the defect regions detected in the two dimensions are matched. For each detected region, the system calculates its confidence score in the texture dimension and the geometric dimension respectively. The texture dimension confidence score is calculated based on features such as reflectance consistency, boundary sharpness and texture continuity; the geometric dimension confidence score is calculated based on geometric features such as surface curvature change, gradient distribution characteristics and morphological regularity. The confidence score calculation uses the following formula: ; wherein, represents the texture dimension confidence score; is a weight coefficient, which is learned through training data; represents the th feature value, including reflectance consistency, boundary sharpness, etc. represents the feature index, from 1 to .

[0069] Two dimensions set independent threshold standards, which are obtained through a large number of experimental data statistics. Only when a certain area is identified as a defect in the detection results of the two dimensions, and the respective confidence scores exceed the threshold of the corresponding dimension, is the area finally determined to exist a real defect.

[0070] The entire detection process is completed within the production line beat, realizing online automatic detection of part defects. The system can adapt to different changes in lighting environment and effectively overcome the detection difficulties brought by high-reflective surfaces. The detection results are output to the production line control system in real time, which is used to trigger the sorting mechanism or generate a quality detection report.

[0071] The method obtains rich surface information through multi-light source active imaging technology, extracts illumination-invariant intrinsic features using a deep learning model, and uses a two-dimensional verification mechanism to ensure the accuracy of the detection results, effectively solving the problem of poor consistency of traditional detection methods in detecting high-reflective part surface defects under complex lighting conditions.

[0072] The following loss function is used in the model training process:

[0073] ; wherein, total loss value, predicted reflectance map, real reflectance map, predicted normal map, real normal map, gradient of normal map, weight coefficient of each loss term, square of L2 norm.

[0074] The Adam optimizer is used in the optimization process, and the learning rate is dynamically adjusted according to the training progress. The initial learning rate is set to 0.001, and the learning rate is reduced to half of the original value every 50 epochs. The early stopping strategy is used in the training process, and the training is terminated when the validation set loss does not decrease for 10 consecutive epochs.

[0075] The method also includes the following quality control measures in the implementation process: periodically calibrate the multi-light source lighting system to ensure the stability of the lighting intensity; establish a standard sample library and periodically verify the performance of the detection system; set up a detection parameter self-adaptive adjustment mechanism to automatically optimize the detection parameters according to different part types.

[0076] Through the implementation of the above technical solutions, accurate and stable detection of electromagnetic stove production line part surface defects can be realized, meeting the strict requirements of industrial production for quality control. The system has good engineering applicability and scalability, providing an effective technical solution for quality detection of similar industrial surfaces.

[0077] The technical means of extracting surface intrinsic properties through multi-source time-sharing imaging and deep learning effectively overcome the problem of inconsistent imaging of high-reflective curved parts under complex lighting environments, improving the accuracy and environmental adaptability of defect detection. The mechanism of independent detection of texture and geometry in two dimensions and combined arbitration determination reduces the occurrence of false positives and false negatives, ensuring the reliability of the detection results and meeting the strict requirements of quality control in industrial production.

[0078] The realization of automatic online detection of electromagnetic oven part surface defects reduces the dependence on manual inspection and improves the detection efficiency and consistency of the production line. Through the cooperative optimization of hardware control and algorithm processing, this method has good engineering applicability and scalability, providing an effective technical solution for quality detection of similar industrial surfaces.

[0079] 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 actual such relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment. Without further limitation, the element defined by the statement "including a" does not exclude the presence of additional identical elements in the process, method, article or equipment including the element.

[0080] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for detecting defects in parts of an induction cooker production line based on image recognition, characterized in that, Includes the following steps: S1: Multiple images of the part under different lighting conditions are acquired by triggering different lighting units in a time-sharing manner through a multi-source lighting system; S2: Input the image and corresponding lighting parameters into a deep learning model to obtain the intrinsic property map of the part surface; S3: Defects are judged based on the intrinsic attribute map from the dimensions of texture attributes and geometric attributes, respectively; S4: Merge and arbitrate the judgment results from the two dimensions and output the final defect judgment result.

2. The method for detecting defects in parts of an induction cooker production line based on image recognition according to claim 1, characterized in that: The multi-source lighting system comprises a ring-shaped group of lighting units, each group having a different incident angle and spectral wavelength configuration.

3. The method for detecting defects in parts of an induction cooker production line based on image recognition according to claim 2, characterized in that: The time-division triggering adopts a sequential triggering mode, and the triggering interval between adjacent lighting units is determined based on the component conveying speed.

4. The method for detecting defects in parts of an induction cooker production line based on image recognition according to claim 1, characterized in that: The deep learning model is trained using a training set containing synthetic training data, which is generated using computer graphics rendering techniques and contains ground truth labels for intrinsic surface attributes.

5. The method for detecting defects in parts of an induction cooker production line based on image recognition according to claim 4, characterized in that: The intrinsic surface property map includes a reflectance property map and a normal direction map, wherein the reflectance property map is used to characterize the surface material properties, and the normal direction map is used to characterize the surface geometry.

6. The method for detecting defects in parts of an induction cooker production line based on image recognition according to claim 5, characterized in that: The texture attribute dimension defect judgment uses a convolutional neural network to analyze the reflectivity attribute map and identify surface scratches and material defects.

7. The method for detecting defects in parts of an induction cooker production line based on image recognition according to claim 5, characterized in that: The geometric attribute dimension defect judgment is performed by performing gradient calculation and morphological processing on the normal direction map to identify surface unevenness and deformation defects.

8. The method for detecting defects in parts of an induction cooker production line based on image recognition according to claim 1, characterized in that: The fusion arbitration adopts a dual-threshold judgment mechanism. When the confidence scores of the same area in both dimensions exceed the corresponding thresholds, it is judged as a defect.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the program, it implements the steps of the method as described in any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-8.

Images