Motor rear cover injection molding defect identification method and system combined with visual inspection

Abstract

The invention relates to the technical field of visual inspection, and provides a motor rear cover injection molding defect identification method and system combined with visual inspection. The method comprises the steps that a collection network is connected, multi-level image information of an injection molding area of the motor rear cover is obtained, and the image information has a light source auxiliary identifier; performing feature recognition on the image information based on features of the light source auxiliary identifier to obtain recognition features of each level; and performing hierarchical fusion interaction on the identification features of each level according to the multi-level image acquisition features, determining a defect identification result, and performing defect positioning feedback, so as to solve the technical problems of poor reliability of defect identification and difficulty in surface and deep defect identification under different illumination conditions, realize multi-light-source cooperation and multi-level feature fusion, and improve the defect identification efficiency. And the appearance and deep defects of the motor rear cover injection molding part can be comprehensively and accurately identified and positioned.

Description

Technical Field

[0001] This application relates to the field of visual inspection technology, specifically to a method and system for identifying injection molding defects in motor rear covers using visual inspection. Background Technology

[0002] With the continuous development of industrial automation, precision manufacturing and intelligent inspection technologies have gradually become important means to improve production efficiency and product quality. In the field of motor manufacturing, especially in the injection molding process of motor back covers, the occurrence of injection molding defects can seriously affect product quality and may even lead to motor failure, resulting in huge costs for production and maintenance. Therefore, accurate and efficient defect detection and identification technologies have become an indispensable part of the production process. Traditional defect detection methods, such as manual inspection or automated inspection based on simple image processing, can detect obvious defects in some cases, but they often have high rates of missed detection and false detection when faced with complex production environments and minute defects under various lighting conditions. In particular, the identification of surface and deep defects is more difficult due to changes in lighting and the diversity of defect types. Summary of the Invention

[0003] This application provides a method and system for identifying injection molding defects in motor back covers by combining visual inspection, aiming to solve the technical problems of poor reliability in defect identification under different lighting conditions and difficulty in identifying surface and deep defects.

[0004] The first aspect disclosed in this application provides a method for identifying injection molding defects in a motor rear cover using visual inspection. The method includes: connecting to a data acquisition network to obtain multi-level image information of the injection molding area of ​​the motor rear cover, wherein the image information has light source auxiliary markers; performing feature recognition on the image information based on the features of the light source auxiliary markers to obtain recognition features at each level; performing hierarchical fusion interaction on the recognition features at each level according to the multi-level image acquisition features to determine the defect identification result and provide defect location feedback.

[0005] Another aspect of this application discloses a defect identification system for motor rear cover injection molding combined with visual inspection. The system includes: an image acquisition module connected to an acquisition network to obtain multi-level image information of the injection molding area of ​​the motor rear cover, wherein the image information has light source auxiliary markers; a feature recognition module that performs feature recognition on the image information based on the features of the light source auxiliary markers to obtain recognition features at each level; and a defect localization module that performs hierarchical fusion interaction on the recognition features at each level according to the multi-level image acquisition features to determine the defect identification result and provide defect localization feedback.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: The aforementioned method for identifying injection molding defects in motor rear covers, combined with visual inspection, first acquires multi-layered image information of the injection molding area of ​​the motor rear cover by connecting to a data acquisition network, and then combines this information with auxiliary marking features from different light sources. This image information includes different features of the surface, depth, and material. Next, based on the features of the light source-assisted markings, the images are analyzed to extract recognition features at each level. Finally, by fusing and interacting the features from different levels, the final defect identification result is generated, and defect location feedback is provided, effectively improving the accuracy and reliability of defect identification.

[0007] The above description is merely an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a flowchart illustrating a method for identifying injection molding defects in a motor rear cover that incorporates visual inspection in one embodiment.

[0010] Figure 2 This is an architecture diagram of a motor rear cover injection molding defect recognition system that incorporates visual inspection in one embodiment.

[0011] Figure labeling: Image acquisition module 11, feature recognition module 12, defect location module 13. Detailed Implementation

[0012] This application provides a method and system for identifying injection molding defects in motor back covers by combining visual inspection, thereby solving the technical problems of poor reliability in defect identification under different lighting conditions and difficulty in identifying surface and deep defects.

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0014] It should be noted that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such process, method, product, or device.

[0015] Example 1, as Figure 1 As shown, this application provides a method for identifying injection molding defects in motor rear covers using visual inspection, the method comprising: The system connects to the acquisition network to obtain multi-level image information of the injection molding area of ​​the motor rear cover, where the image information includes light source auxiliary markers.

[0016] In this embodiment, a connection is first established with a data acquisition network, which includes various sensors such as high-resolution 2D cameras and thermal imaging cameras, capable of covering multiple levels of image features. Then, this acquisition network is used to acquire images of the injection-molded area of ​​the motor's rear cover, obtaining multi-level image information. This multi-level image information is labeled with light source auxiliary identifiers, which contain specific light source information, such as light source intensity, light source type, and light source acquisition parameters. This labeled multi-level image information provides sufficient visual data for subsequent feature extraction and defect identification, ensuring that potential defect types can be comprehensively and accurately identified during the inspection process.

[0017] Furthermore, this application provides, prior to connecting to the data acquisition network, the following: A multi-level image acquisition device and a light source device are connected to form a multi-level combined sub-unit. The multi-level combined sub-unit is connected via an RS485 digital potentiometer network. The RS485 communication protocol is used to establish the connection between the multi-level combined sub-units, and the digital potentiometers are used to control the configuration parameters of each light source.

[0018] Preferably, before connecting to the acquisition network, the multi-level image acquisition devices and light source devices are first connected to construct multi-level combined sub-units. Each combined sub-unit includes a set of image acquisition devices and corresponding light source devices, which together constitute the required acquisition network. To ensure effective collaborative work between multiple sub-units, these sub-units are connected via an RS485 digital potentiometer network. The RS485 communication protocol, as a reliable data transmission protocol, supports the serial connection of multiple devices and has good anti-interference capabilities and high data transmission reliability. Through the RS485 bus, multiple combined sub-units can share data in the same network, achieving synchronous data transmission. In each combined sub-unit, the intensity and other configuration parameters of the light source need to be dynamically adjusted according to actual needs. To achieve this function, a digital potentiometer is used as a controller. The digital potentiometer can change the current of the light source by adjusting the resistance value, thereby controlling parameters such as the brightness and color temperature of the light source. Through the control signals transmitted via the RS485 network, the digital potentiometer can precisely adjust the working state of each light source, ensuring optimal illumination conditions during image acquisition, thereby improving the performance and accuracy of the entire defect identification system.

[0019] Furthermore, this application provides that the multi-level image acquisition device includes at least: a surface feature acquisition device, a deep feature acquisition device, and a material feature acquisition device; wherein, the surface feature acquisition device includes a high-resolution 2D camera and a polarization imaging system for acquiring surface texture, color distribution, and gloss features; the deep feature acquisition device includes a 3D line laser scanner and a structured light projection system for acquiring three-dimensional morphology, internal structure, and stress distribution features; and the material feature acquisition device includes a multispectral imaging system and a thermal imaging camera for acquiring spectral characteristics, thermal conductivity characteristics, and material uniformity features.

[0020] Optionally, the multi-level image acquisition device includes at least three main parts: a surface feature acquisition device, a deep feature acquisition device, and a material feature acquisition device. Each part has a specific function and role to ensure comprehensive and accurate capture of different layers of features in the injection-molded area of ​​the motor back cover. The surface feature acquisition device mainly includes a high-resolution 2D camera and a polarization imaging system to acquire surface information of the motor back cover, including surface texture, color distribution, and gloss. These features help identify surface defects such as cracks, dents, and scratches. During operation, the high-resolution 2D camera can acquire images of surface texture and color distribution, while the polarization imaging system acquires images using polarized light at different angles, highlighting surface gloss and minor imperfections. The deep feature acquisition equipment mainly includes a 3D line laser scanner and a structured light projection system, used to acquire the internal structure, three-dimensional morphology, and stress distribution characteristics of the motor back cover. These features help identify deep structural defects, such as internal cracks, bubbles, and molding defects. During operation, the 3D line laser scanner scans the surface of the motor back cover to acquire three-dimensional morphology and depth data. The structured light projection system projects a beam of light with a known pattern onto the object surface, analyzes the deformation of the light pattern, and acquires more refined three-dimensional morphology and structural information. The material feature acquisition equipment mainly includes a multispectral imaging system and a thermal imaging camera, used to analyze the material properties of the motor back cover, including spectral characteristics, thermal conductivity, and material uniformity. These features help identify potential defects caused by material defects, such as material inhomogeneity and thermal stress. During operation, the multispectral imaging system acquires images of the motor back cover at different wavelengths to analyze the spectral characteristics and surface changes of the material. The thermal imaging camera captures thermal radiation images to analyze the temperature distribution and thermal conductivity characteristics of the motor back cover and identify heat-related defects. By comprehensively utilizing these surface feature acquisition devices, deep feature acquisition devices, and material feature acquisition devices, the injection molding area of ​​the motor rear cover can be fully inspected from multiple angles and levels, providing more accurate and efficient defect identification and location results.

[0021] Furthermore, this application provides a method for controlling the configuration parameters of each light source using a digital potentiometer, including: Adjustment of light source intensity and adjustment of light source-assisted image type.

[0022] Optionally, digital potentiometers are used to precisely control the light source configuration parameters of each light source device in the acquisition network, including the light source intensity and the type of light source-assisted image, to optimize the quality and features of the acquired images. Specifically, when adjusting the light source intensity, the digital potentiometer precisely controls the intensity of the light source by adjusting the resistance. The intensity of each light source can be dynamically adjusted according to actual acquisition needs by changing the resistance to control the current passing through the light source. Generally, a lower resistance value allows more current to flow, thus increasing the light source intensity; conversely, a higher resistance value reduces the current and decreases the brightness of the light source. This control method ensures that the brightness of the light source can be precisely adjusted according to different acquisition needs, adapting to the defect identification requirements under different lighting environments. When adjusting the type of light source-assisted image, the digital potentiometer can also adjust the image type under different light source configurations. For example, by adjusting the light source with different wavelengths, the light source can be changed to different illumination methods such as diffuse light, grazing light, and backlight, changing the illumination characteristics of the image so that the image can highlight different defect features. For surface defects, low-angle grazing light can enhance the shadow effect, helping to highlight surface defects such as dents and cracks. For internal material defects, backlighting can enhance the contour contrast of the defects, making deeper defects such as flash and bubbles more visible. By adjusting the digital potentiometer, the most suitable light source configuration can be selected according to different defect types to obtain the best image effect, ensuring that the system can more effectively identify various defects that may occur during the injection molding process of the motor back cover, improving the accuracy and robustness of defect detection.

[0023] Based on the characteristics of the light source-assisted marker, feature recognition is performed on the image information to obtain recognition features at each level.

[0024] In one embodiment, after acquiring multi-level image information, differentiated feature enhancement processing is performed on the acquired image information based on the features in the light source-assisted markers and the corresponding image acquisition levels. This highlights defect areas in image information at different levels. For example, by adjusting the angle and intensity of the light source, surface defects can be more clearly displayed, while deep defects can be better revealed through different lighting effects. Subsequently, feature extraction is performed on the enhanced image information at each level to obtain recognition features at different levels. The recognition features at each level reflect defects of different depths or characteristics, including surface defects, deep defects, and material problems, thereby reducing the occurrence of missed and false detections.

[0025] Furthermore, this application provides a method for feature recognition of the image information based on the features of the light source-assisted identifier, obtaining recognition features at each level, including: Based on the features of the light source-assisted markers and the corresponding image acquisition levels, differential feature enhancement processing is performed on multi-level image information; feature extraction is performed on the image information at each level of feature enhancement processing to obtain the recognition features at each level.

[0026] Preferably, during the acquisition phase, each image includes light source auxiliary identification information, which records data such as light source type, illumination angle, and light intensity parameters. The system first identifies the response of the current optical environment to image acquisition based on the image acquisition level, such as surface, deep, or material level. This response may be enhancement or suppression. Subsequently, based on the determined response relationship and combined with the type of injection molding defect, multi-level image information differential feature enhancement processing is performed, optimizing the image in terms of brightness, contrast, and texture directionality, ensuring that defects are clearly presented at different levels. Next, the enhanced image information at each level is input into a pre-constructed convolutional neural network. This network consists of a surface feature branch network, a deep feature branch network, and a material feature branch network. The surface feature branch network uses a shallow convolutional structure, such as ResNet-18 or MobileNet, to extract two-dimensional features such as surface texture, color distribution, and gloss. The deep feature branch network uses a three-dimensional convolutional structure or a dual-channel CNN structure that fuses depth maps to extract morphological changes and stress distribution. The material feature branch network uses a multi-channel convolutional structure to extract spectral reflectance and thermal gradient distribution features. These branch networks all use corresponding labeled image data and are iteratively trained through forward propagation, loss calculation, backpropagation, and parameter optimization. By extracting features from the received image information at each level through these branch networks, recognition features at each level can be obtained. These features describe the morphological, optical, or physical characteristics of the motor back cover at the corresponding level, providing basic data for subsequent multi-level feature fusion and interaction, thus improving the accuracy and robustness of injection molding defect recognition for the motor back cover.

[0027] Furthermore, this application provides features based on light source-assisted identification and corresponding image acquisition levels, and performs differential feature enhancement processing on multi-level image information, including: Based on the light source type and light source acquisition parameters, determine the enhancement or suppression response relationship of the optical environment to image acquisition; establish the recognition mapping relationship between the enhancement or suppression response relationship and the injection molding defect type; based on the recognition mapping relationship, perform differentiated feature enhancement processing on the multi-level image information according to the injection molding defect type.

[0028] Optionally, feature data such as light source type and light source acquisition parameters are first extracted from the light source auxiliary identifiers. The light source type can be grazing light, diffuse light, backlight, etc., and the light source acquisition parameters can be illumination angle, intensity, wavelength range, etc. This feature data is then input into a support vector machine (SVM), which is trained using light source auxiliary identifier samples with response relationship labels. This SVM can establish enhancement or suppression response relationships between light source parameters and images based on the feature data in the new light source auxiliary identifiers. For example, low-angle grazing light can enhance the shadow response of concave areas; high-intensity diffuse light can reduce overexposure in high-reflectivity areas; backlighting can strengthen boundary features and suppress background noise; and near-infrared light can penetrate some material layers, enhancing internal cavity features. Subsequently, based on the determined response relationships, light source feature signatures for each image feature are established. These light source feature signatures are then correlated with the mapping relationship from defect type to core image features, establishing an enhancement or suppression response relationship and an identification mapping relationship between injection molding defect types. Subsequently, based on this identification mapping relationship, the current defect type is determined, and the enhancement algorithm corresponding to the defect type, such as edge detection algorithm, texture analysis algorithm, region growing algorithm, etc., is used to perform multi-level image information differential feature enhancement processing to suppress noise interference in non-defect areas, so that the defect is more obvious and easier to identify in the image, thereby improving the recognition rate and robustness of injection molding defects in motor back cover.

[0029] Furthermore, this application provides a mapping relationship for establishing the enhanced or suppressed response relationship with the injection molding defect type, including: Based on the enhancement or suppression response patterns of light source type and light source parameters to image acquisition features, light source feature signatures for each image feature are established. By performing image analysis under multiple light source conditions on a known injection molding defect sample library, the core defect features that stably present the most distinctive characteristics of each type of defect under different optical environments are extracted and summarized, and a mapping relationship from defect type to core image features is established. Through the relationship transmission between light source feature signatures and core image features, the enhancement or suppression response relationship and the identification mapping relationship of injection molding defect type are established.

[0030] Optionally, to achieve a precise correspondence between optical response characteristics and injection molding defect types, image acquisition experiments are first conducted under relatively stable conditions, controlling other environmental factors such as temperature, humidity, and background illumination, according to the set light source type and parameter combinations, to obtain multiple sets of image data under different light source conditions. For these image data, texture features are calculated using the gray-level co-occurrence matrix, edge features are calculated using edge detection algorithms such as Canny and Sobel, and color features are calculated by statistically analyzing the mean values ​​of different color channels in the image. Subsequently, the changes in the same image features under different light source parameters are compared to analyze the enhancement or suppression effects of light source type and parameters on image features. For example, it was found that when the light source intensity increases, the edge features of the image become clearer, indicating that high-intensity light sources have an enhancement effect on edge features; while when the incident angle of the light source is small, some texture features may be masked, indicating that small incident angle light sources have a suppression effect on these texture features. Based on these analysis results, corresponding light source feature signatures are established for each image feature. These light source feature signatures are obtained by encoding the light source type, parameters, and corresponding image response results, usually existing in the form of vectors, providing a stable optical response basis for subsequent defect identification. Subsequently, using a sample library of injection molding defects containing various defects such as dents, weld lines, flash, bubbles, shrinkage cavities, and impurities, image data under various light source environments is extracted. Similar feature extraction is then performed on the image data of each type of defect to obtain features such as texture changes, edge shapes, and color differences in the defect area. For each type of defect, principal component analysis is used to extract the most distinctive core defect features. For example, for dent defects, it may be found that under a specific light source angle, the grayscale value of the defect area differs significantly from the surrounding normal area, and this difference remains relatively stable under different light source intensities. By summarizing the defect feature analysis results for each type of defect, a preliminary mapping relationship from defect type to core image features can be formed. For example, the core image feature corresponding to dent defects is "significantly increased shadow intensity and clear edge gradient under grazing light." Then, based on the image features involved in the mapping relationship from defect type to core image features, the corresponding light source feature signature is matched. Using this image feature as a bridge, a transmission relationship between the light source feature signature and the core image features is established. Based on this transmission relationship, the enhancement or suppression response relationship is then linked to the injection molding defect type, establishing a mapping relationship between the enhancement or suppression response relationship and the injection molding defect type. In summary, through the above process, a dynamic mapping between light source parameters, image features, and defect types is achieved, enabling the accurate identification and utilization of image features under different optical conditions. This solves the problem of inconsistent defect performance of injection molded parts under different lighting conditions, improving the adaptability, stability, and accuracy of multi-light source visual inspection.

[0031] Furthermore, this application provides differential feature enhancement processing of the multi-level image information based on the aforementioned recognition mapping relationship and according to the injection molding defect type, including: When the identification mapping relationship is a dent defect type, low-angle grazing light is used to enhance the shadow effect, and the dent contour is extracted by an edge detection algorithm; when the identification mapping relationship is a weld line defect type, high-angle diffuse light is used in conjunction with polarization filtering, and material fusion anomalies are identified by a texture analysis algorithm; when the identification mapping relationship is a flash defect type, backlighting is used to enhance the contour contrast, and abnormal edge protrusions are detected by morphological operations; when the identification mapping relationship is a bubble shrinkage defect type, near-infrared transmission imaging is used, and internal cavities are identified by a region growing algorithm.

[0032] Optionally, when an image feature is detected to match the mapping relationship for dent defect identification, low-angle grazing light is automatically used. This low-angle grazing light allows the light to illuminate in a direction nearly parallel to the surface of the injection molded part, resulting in refraction and reflection at the edge of the dent, thus reducing the amount of light inside the dent and creating a noticeable shadow effect at the dent. Subsequently, the image is Gaussian smoothed using the Canny edge detection algorithm to suppress noise. Then, the gradient magnitude and direction of the image are calculated to find points with drastic gray-level changes. Non-maximum suppression is used to retain points with local maximum gradient magnitudes, refining the edges. Finally, dual threshold detection and edge connection are used to determine the final edge contour, which serves as the dent contour to clearly distinguish the dented area from the surrounding normal area. When an image feature is detected to match the mapping relationship for weld line defect identification, high-angle diffuse light is automatically used to illuminate the surface of the injection molded part in a more uniform manner, reducing interference from reflected light on the injection molded part surface. At the same time, polarization filtering is applied to the collected light to further eliminate the polarization component in the surface reflected light, enhancing the texture information inside the material and making the material fusion anomaly at the weld line more obvious. Next, the image undergoes texture enhancement processing. Texture directionality, energy, and contrast are extracted using the gray-level co-occurrence matrix, and microstructural changes at the material fusion point are analyzed using a local binary mode algorithm. When a discontinuity in texture direction or abnormal local gray-level distribution is detected, it is identified as a weld line defect. When the image features match the mapping relationship for flash defect identification, backlighting is automatically applied to create a strong brightness contrast between the motor back cover edge and the background, highlighting the abnormal protruding parts of the flash in the image. Then, the acquired image is binarized, converting it into a black-and-white binary image to completely separate the flash area from the background area. Morphological opening and closing operations are then used to process the binary image. The opening operation first performs erosion to remove small noise points and tiny protrusions, followed by dilation to restore the original shape of the object, but not to restore small parts that have been completely eroded. The closing operation first performs dilation to fill small holes inside the object and connect adjacent objects, followed by erosion to restore the object's boundaries. By combining morphological opening and closing operations, the edges of the flash area can be smoothed, noise interference removed, and the flash outline made clearer. By analyzing the shape and position information of the obtained edges, abnormal protrusions, i.e., flash defects, can be identified. Combined with area calculations, the defect size can be further quantified, enabling automatic flash detection. When the detected image features match the mapping relationship for bubble shrinkage defects, near-infrared transmission imaging is automatically used. Near-infrared light can penetrate the injection molded part, causing internal cavities such as bubbles to exhibit different grayscale characteristics in the image. Next, the acquired near-infrared image undergoes denoising processing, for example, using wavelet transform denoising methods to remove noise interference and improve image quality. Finally, grayscale processing is performed to convert the image into a grayscale image.Furthermore, based on the grayscale characteristics of bubble shrinkage defects in the image, a suitable seed point is selected. The seed point is the starting point for region growth, and pixels with lower grayscale values ​​are usually chosen as seed points. Then, with the seed point as the center, according to pre-set growth criteria, such as grayscale similarity criteria, adjacent pixels with similar grayscale values ​​to the seed point are merged into the same region. This process is repeated until no new pixels can be merged into the region. Through this region growth process, internal cavity regions such as bubble shrinkage defects can be segmented from the image. Based on the size, shape, and other features of the segmented region, it can be determined whether bubble shrinkage defects exist. When the area or shape parameters of the segmented region exceed a threshold, it is determined that bubble shrinkage defects exist. Through the above-mentioned light source configuration and algorithm matching strategy for different defect types, the optimal imaging and processing method can be automatically selected according to the recognition mapping relationship, so that various defects are maximized in the most suitable optical environment, thereby improving the detection accuracy of injection molding defects and reducing recognition errors under different lighting conditions.

[0033] Based on the multi-level image acquisition features, the recognition features of each level are fused and interacted to determine the defect recognition result and provide defect location feedback.

[0034] In one embodiment, after obtaining the recognition features at each level, these features are spatially aligned and scaled to ensure comparability in spatial location, pixel density, and grayscale range across different levels. Subsequently, a feature interaction model between levels is established in a unified feature space, dynamically allocating the weights of features at each level through an attention mechanism. For surface features, their weight in surface defect detection is increased based on texture clarity and gloss variation; for deep features, their sensitivity to internal defects such as bubbles and shrinkage cavities is enhanced based on structural depth and stress gradient information; for material-level features, their contribution to material inhomogeneity identification is strengthened through spectral consistency and thermal distribution characteristics. After horizontal and vertical fusion interactions, the fused features are cross-validated and confidence scores are calculated to remove noise and false positives, generating the final defect recognition result. Finally, the identified defect areas are spatially located and fed back. Coordinate mapping is used to present the actual location of the defect on the motor back cover on the detection result interface, achieving visual labeling and precise location of the defect, thus providing a basis for subsequent quality control and production process optimization.

[0035] Furthermore, this application provides the method of performing hierarchical fusion interaction on the recognition features of each level according to the multi-level image acquisition features to determine the defect recognition result, including: Spatial registration and scale normalization are performed on the identification features at each level to establish a unified multi-level feature space. In the multi-level feature space, weight allocation based on attention and feature interaction is performed to conduct horizontal fusion, vertical fusion, and spatial fusion. The defect identification result is generated through multi-level feature cross-validation and confidence fusion. The weight allocation based on attention and feature interaction includes hierarchical attention weight allocation, channel attention allocation, and spatial attention allocation. The hierarchical attention weight allocation is to configure fusion weights for vertical fusion based on the recognition relativity of features at each level. The channel attention allocation is to configure fusion weights for horizontal fusion based on the recognition influence relationship of defects on features at different light sources within the same level. The spatial attention allocation is to configure fusion weights for spatial fusion based on the spatial distribution influence of defect regions.

[0036] Optionally, since the image data at each level originates from different sensing devices and exhibits differences in acquisition angle and coordinate deviation, the geometric alignment of the recognition features at each level is first achieved based on feature point matching algorithms, such as SIFT, SURF, or structured light depth information, mapping the images at each level to a unified spatial coordinate system. To ensure consistency of features at different resolutions and imaging scales, interpolation and maximum-minimum normalization are used to standardize feature sizes and grayscale ranges, constructing a unified multi-level feature space to provide a unified foundation for fusion computation. Subsequently, within the multi-level feature space, a multi-dimensional attention mechanism and feature interaction strategy are introduced to perform fusion from three aspects: vertical (hierarchical fusion), horizontal (channel fusion), and spatial fusion. In vertical fusion, weights are dynamically assigned based on the importance of different level features to defect identification. For example, for surface defects such as dents and weld lines, the weight of surface-related features is increased; for structural defects such as bubbles and shrinkage cavities, the weight of deep and material features is increased. The attention module then adaptively updates the fusion ratio based on the recognition confidence and information content of features at each level, achieving vertical multi-layer feature aggregation. In horizontal fusion, where multi-channel images acquired from different light sources reflect different feature responses at the same level, a channel attention mechanism, such as the SE-Net structure, calculates the correlation between each light source feature and defect saliency, and dynamically adjusts the channel weights based on their recognition contribution, achieving horizontal fusion of light source features. In spatial fusion, a spatial attention mechanism analyzes the spatial distribution saliency of defect areas, assigning higher weights to abnormal areas. For example, heatmap analysis is used to locate defect-prone areas, and corresponding weights are increased. Next, the fused feature results need to undergo inter-level consistency verification to improve the stability and reliability of the identification. This process involves surface-to-depth consistency verification: for detected surface anomalies, it checks whether there are corresponding geometric deformations or morphological anomalies in the deeper features to rule out surface lighting artifacts or noise-induced misjudgments. It also involves depth-to-material consistency verification: for structural anomalies detected in the deeper layers, the spectral or thermal properties of the material layer are analyzed simultaneously to verify whether they are accompanied by abnormal heat distribution or changes in spectral reflectance, thus confirming that they are genuine defects rather than material texture fluctuations. Following this, a consistency scoring system is established to quantify the surface-to-depth and depth-to-material verification levels. For surface-to-depth verification, cosine similarity is used to calculate the feature similarity between the surface texture features and the deep morphological features. The intersection-union ratio (IUU) of surface defect regions and deep defect regions is calculated to obtain the spatial overlap. By weighting these two calculation results, a consistency score for the verification process can be obtained.For deep-material verification, principal component analysis is performed on the material spectrum to calculate the spectral entropy of the anomalous region. Local curvature analysis is performed on the deep morphological features to calculate the average absolute value of the curvature of the anomalous region. The overlap between the material anomalous region and the deep morphological anomalous region is calculated, and the Euclidean distance between their center points is statistically analyzed. By weighting these three calculation results, a consistency score for the verification process can be obtained. If the consistency score is high, the confidence of the defect in that region is improved; conversely, it is marked as a low-confidence region requiring further verification. Then, to obtain the final defect identification conclusion, each level of identification feature is treated as an independent source of evidence, and a confidence fusion calculation is performed based on evidence theory. In this process, for each level of feature, its support function for different defect types is calculated, representing the identification strength of that level of feature for a certain defect. Then, using the DS evidence combination rule, the support of multiple evidence sources is weighted and fused to form a comprehensive confidence function. During the fusion process, conflicting evidence is automatically suppressed, and the feature responses with high consistency across multiple layers are strengthened. When the overall confidence level is higher than the threshold, the defect type and location coordinates are output; when the confidence level is lower than the threshold, it is marked as an area to be reviewed or further data collection and processing is required. Through the above-mentioned fusion, verification, and confidence level decision-making mechanism, the system can achieve efficient complementarity and dynamic weight allocation of multi-level features, enabling the defect identification results to simultaneously possess high accuracy, high robustness, and interpretability, providing a reliable basis for subsequent quality control and process optimization.

[0037] In summary, the embodiments of this application have at least the following technical effects: This embodiment first connects to a data acquisition network to obtain multi-level image information of the injection-molded area of ​​the motor rear cover, wherein the image information includes light source auxiliary markers. Then, based on the features of the light source auxiliary markers, feature recognition is performed on the image information to obtain recognition features at each level. Finally, the recognition features at each level are fused and interacted according to the multi-level image acquisition features to determine the defect recognition result and provide defect location feedback. These technical effects collectively solve the technical problems of poor reliability in defect recognition under different lighting conditions and difficulty in identifying surface and deep defects. They achieve the technical effect of comprehensive and accurate identification and location of surface and deep defects in the injection-molded part of the motor rear cover through multi-light source collaboration and multi-level feature fusion.

[0038] Example 2, based on the same inventive concept as the motor rear cover injection molding defect identification method combined with visual inspection in the aforementioned examples, such as... Figure 2As shown, this application provides a motor rear cover injection molding defect identification system combined with visual inspection. The system includes: an image acquisition module 11, which connects to an acquisition network to obtain multi-level image information of the motor rear cover injection molding area, wherein the image information has light source auxiliary markers; a feature recognition module 12, which performs feature recognition on the image information based on the features of the light source auxiliary markers to obtain recognition features at each level; and a defect localization module 13, which performs hierarchical fusion interaction on the recognition features at each level according to the multi-level image acquisition features to determine the defect identification result and provide defect localization feedback.

[0039] Furthermore, the image acquisition module 11 is also used to perform the following method: A multi-level image acquisition device and a light source device are connected to form a multi-level combined sub-unit. The multi-level combined sub-unit is connected via an RS485 digital potentiometer network. The RS485 communication protocol is used to establish the connection between the multi-level combined sub-units, and the digital potentiometers are used to control the configuration parameters of each light source.

[0040] Furthermore, the image acquisition module 11 is also used to perform the following method: The multi-level image acquisition device includes at least: a surface feature acquisition device, a deep feature acquisition device, and a material feature acquisition device; wherein, the surface feature acquisition device includes a high-resolution 2D camera and a polarization imaging system for acquiring surface texture, color distribution, and gloss characteristics; the deep feature acquisition device includes a 3D line laser scanner and a structured light projection system for acquiring three-dimensional morphology, internal structure, and stress distribution characteristics; and the material feature acquisition device includes a multispectral imaging system and a thermal imaging camera for acquiring spectral characteristics, thermal conductivity characteristics, and material uniformity characteristics.

[0041] Furthermore, the image acquisition module 11 is also used to perform the following method: Adjustment of light source intensity and adjustment of light source-assisted image type.

[0042] Furthermore, the feature recognition module 12 is also used to perform the following method: Based on the features of the light source-assisted markers and the corresponding image acquisition levels, differential feature enhancement processing is performed on multi-level image information; feature extraction is performed on the image information at each level of feature enhancement processing to obtain the recognition features at each level.

[0043] Furthermore, the feature recognition module 12 is also used to perform the following method: Based on the light source type and light source acquisition parameters, determine the enhancement or suppression response relationship of the optical environment to image acquisition; establish the recognition mapping relationship between the enhancement or suppression response relationship and the injection molding defect type; based on the recognition mapping relationship, perform differentiated feature enhancement processing on the multi-level image information according to the injection molding defect type.

[0044] Furthermore, the feature recognition module 12 is also used to perform the following method: When the identification mapping relationship is a dent defect type, low-angle grazing light is used to enhance the shadow effect, and the dent contour is extracted by an edge detection algorithm; when the identification mapping relationship is a weld line defect type, high-angle diffuse light is used in conjunction with polarization filtering, and material fusion anomalies are identified by a texture analysis algorithm; when the identification mapping relationship is a flash defect type, backlighting is used to enhance the contour contrast, and abnormal edge protrusions are detected by morphological operations; when the identification mapping relationship is a bubble shrinkage defect type, near-infrared transmission imaging is used, and internal cavities are identified by a region growing algorithm.

[0045] Furthermore, the feature recognition module 12 is also used to perform the following method: Based on the enhancement or suppression response patterns of light source type and light source parameters to image acquisition features, light source feature signatures for each image feature are established. By performing image analysis under multiple light source conditions on a known injection molding defect sample library, the core defect features that stably present the most distinctive characteristics of each type of defect under different optical environments are extracted and summarized, and a mapping relationship from defect type to core image features is established. Through the relationship transmission between light source feature signatures and core image features, the enhancement or suppression response relationship and the identification mapping relationship of injection molding defect type are established.

[0046] Furthermore, the defect location module 13 is also used to perform the following method: Spatial registration and scale normalization are performed on the identification features at each level to establish a unified multi-level feature space. In the multi-level feature space, weight allocation based on attention and feature interaction is performed to conduct horizontal fusion, vertical fusion, and spatial fusion. The defect identification result is generated through multi-level feature cross-validation and confidence fusion. The weight allocation based on attention and feature interaction includes hierarchical attention weight allocation, channel attention allocation, and spatial attention allocation. The hierarchical attention weight allocation is to configure fusion weights for vertical fusion based on the recognition relativity of features at each level. The channel attention allocation is to configure fusion weights for horizontal fusion based on the recognition influence relationship of defects on features at different light sources within the same level. The spatial attention allocation is to configure fusion weights for spatial fusion based on the spatial distribution influence of defect regions.

[0047] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0048] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0049] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for recognizing injection defect of motor rear cover combined with visual detection, characterized in that, The method comprises the following steps: connecting a collection network to obtain multi-level image information of the motor rear cover injection molding area, wherein the image information has light source auxiliary identification; based on the characteristics of the light source auxiliary identification, feature recognition is performed on the image information to obtain each level recognition feature; according to the multi-level image collection characteristics, the level fusion interaction of each level recognition feature is determined to determine the defect recognition result and perform defect positioning feedback.

2. The method for recognizing injection defect of motor back cover combined with visual detection according to claim 1, characterized in that, Before connecting the collection network, the method comprises the following steps: connecting a multi-level image collection device and a light source device to construct a multi-layer combination subunit; connecting the multi-layer combination subunit through an RS485 digital potentiometer network, wherein the RS485 communication protocol establishes the connection of the multi-layer combination subunit, and the digital potentiometer is used to control the configuration parameters of each light source.

3. The method for recognizing injection defect of motor back cover combined with visual detection according to claim 2, characterized in that, The multi-level image collection device at least comprises: a surface feature collection device, a deep feature collection device, and a material feature collection device; wherein the surface feature collection device comprises a high-resolution 2D camera and a polarization imaging system for collecting surface texture, color distribution, and glossiness features; the deep feature collection device comprises a 3D line laser scanner and a structured light projection system for collecting three-dimensional topography, internal structure, and stress distribution features; and the material feature collection device comprises a multispectral imaging system and a thermal imaging camera for collecting spectral characteristics, thermal conductivity characteristics, and material uniformity characteristics.

4. The method for recognizing injection molding defects of a motor back cover combined with visual inspection according to claim 2, characterized in that, Using a digital potentiometer to control the configuration parameters of each light source includes adjusting the intensity of the light source and adjusting the type of light source auxiliary image.

5. The method for recognizing injection molding defects of a motor back cover combined with visual inspection according to claim 1, characterized in that, Based on the characteristics of the light source auxiliary identification, feature recognition is performed on the image information to obtain each level recognition feature, which comprises the following steps: based on the characteristics of the light source auxiliary identification and the corresponding image collection level, differentiating feature enhancement processing is performed on the multi-level image information; feature extraction is performed on each level image information subjected to feature enhancement processing to obtain the each level recognition feature.

6. The method for recognizing injection defect of motor back cover combined with visual detection according to claim 5, characterized in that, Based on the characteristics of the light source auxiliary identification and the corresponding image collection level, differentiating feature enhancement processing is performed on the multi-level image information, which comprises the following steps: determining the enhancement or inhibition response relationship of the optical environment to image collection according to the type of light source and the collection parameters of the light source; establishing the recognition mapping relationship between the enhancement or inhibition response relationship and the injection defect type; based on the recognition mapping relationship, differentiating feature enhancement processing is performed on the multi-level image information according to the injection defect type.

7. The method for recognizing injection defect of motor back cover combined with visual detection according to claim 6, characterized in that, Based on the recognition mapping relationship, differentiating feature enhancement processing is performed on the multi-level image information according to the injection defect type, which comprises the following steps: when the recognition mapping relationship is a dent defect type, low-angle grazing light is used to enhance the shadow effect, and the edge detection algorithm is used to extract the dent profile; when the recognition mapping relationship is a weld line defect type, high-angle diffuse light is used in combination with polarization filtering, and the texture analysis algorithm is used to identify material fusion abnormalities; when the recognition mapping relationship is a flash defect type, backlight illumination is used to enhance the contrast of the profile, and morphological operation is used to detect edge abnormal protrusions; when the recognition mapping relationship is a bubble shrinkage defect type, near-infrared transmission imaging is used, and the region growing algorithm is used to identify internal cavities.

8. The method for recognizing injection defect of motor back cover combined with visual detection according to claim 6, characterized in that, The mapping relationship between the enhancement or inhibition response relationship and the identification of the injection defect type is established, including: According to the enhancement or inhibition response law of the light source type and the light source parameter on the image acquisition feature, the light source feature signature of each image feature is established; By analyzing the images of the known injection defect sample library under the condition of multiple light sources, the core defect features of each type of defect that are most distinguishable and stable under different optical environments are extracted and summarized, and the mapping relationship from the defect type to the core image feature is established; The mapping relationship between the enhancement or inhibition response relationship and the identification of the injection defect type is established by conducting relationship transmission between the light source feature signature and the core image feature.

9. The method for recognizing injection molding defects of a motor back cover combined with visual inspection according to claim 1, characterized in that, The hierarchical fusion interaction of the hierarchical identification features is performed according to the multi-level image acquisition features, and the defect identification result is determined, including: The spatial registration and scale normalization of the hierarchical identification features are performed, and a unified multi-level feature space is established; In the multi-level feature space, the weight distribution based on attention and feature interaction is performed, horizontal fusion, vertical fusion and spatial fusion are performed, and the defect identification result is generated through multi-level feature cross-validation and confidence fusion; The weight distribution based on attention and feature interaction includes hierarchical attention weight distribution, channel attention distribution and spatial attention distribution, the hierarchical attention weight distribution is vertical fusion according to the identification relative of each level feature and the configuration of fusion weight, the channel attention distribution is horizontal fusion according to the identification influence relationship of defects in different light source features and the configuration of fusion weight, and the spatial attention distribution is spatial fusion according to the spatial distribution position influence of the defect area and the configuration of fusion weight.

10. A motor back cover injection defect recognition system combined with visual detection, characterized in that, The system is used to perform the motor rear cover injection defect identification method combined with visual detection according to any one of claims 1-9, including: An image acquisition module: connected to a collection network, obtaining multi-level image information of the motor rear cover injection area, wherein the image information has a light source auxiliary identifier; A feature recognition module: based on the features of the light source auxiliary identifier, the image information is subjected to feature recognition to obtain hierarchical identification features; A defect positioning module: the hierarchical fusion interaction of the hierarchical identification features is performed according to the multi-level image acquisition features, the defect identification result is determined, and the defect positioning feedback is performed.

Images