An image enhancement-based automobile part defect identification method and system

Abstract

This invention relates to the field of automotive part defect recognition technology, and discloses an image enhancement-based method for automotive part defect recognition. The method includes acquiring historical inspection image data of parts, determining the processing cycle and constructing a periodic inspection image dataset, and obtaining a periodic characteristic inspection image dataset through part type filtering; performing image enhancement processing on the dataset, constructing enhanced feature paths and analyzing scores, and determining the optimal periodic characteristic inspection image data; using the enhanced feature path nodes as references for clustering processing, analyzing suitable defect features, and determining the suitable defect recognition score range for the part. This system improves the accuracy and adaptability of automotive part defect recognition through image enhancement, periodic characteristic analysis, and dynamic score range construction, and is suitable for automated defect detection scenarios in automotive parts.

Description

Technical Field

[0001] This invention relates to the field of automotive parts defect recognition technology, specifically to an image enhancement-based automotive parts defect recognition method and system. Background Technology

[0002] In the automotive manufacturing industry, defect detection of automotive parts is a crucial step in ensuring product quality. Traditional methods for identifying automotive parts defects often involve manual visual inspection, but this approach is highly susceptible to subjective factors. Differences in experience, responsibility, and visual sensitivity among inspectors can lead to discrepancies in defect assessments of the same part. Prolonged inspection can also cause fatigue, increasing the probability of missed or false positives. This is especially true for minute defects such as micro-cracks and scratches on the surface of parts, or hidden defects in structurally complex parts, where the accuracy of manual identification is difficult to guarantee, severely impacting inspection efficiency and reliability.

[0003] Besides manual inspection, some automated inspection methods rely on fixed thresholds for judgment, but these methods also have significant limitations. During the production of automotive parts, different production batches exhibit significant differences in image presentation. For example, image resolution, lighting conditions, and surface reflectivity may vary between batches. Fixed thresholds cannot flexibly adapt to these variations. For some batches, they may over-identify, misclassifying normal areas as defects, while for others, they may under-identify, missing actual defects, leading to unstable overall inspection results.

[0004] While machine vision technology has advanced defect recognition techniques, it still faces numerous challenges. Inspection images of automotive parts often suffer from poor quality due to various factors. For example, uneven lighting in the production environment can cause some areas of the image to be too dark or too bright; fluctuations in the parameters of the imaging equipment can lead to a decrease in image resolution; and reflections on the part's surface can obscure potential defect features. Traditional image enhancement methods often employ a uniform processing approach, making it difficult to specifically improve the contrast between defective and normal areas, and also failing to effectively preserve defect details. This results in blurred defect features in the enhanced image, directly impacting the accuracy of subsequent defect recognition.

[0005] Most existing defect identification systems are based on static models and lack attention to the characteristics of the parts' production cycle. The production process of automotive parts involves cyclical changes. For example, production equipment wears down over long-term operation, leading to differences in dimensional accuracy and surface finish between parts produced in different cycles. Batch changes of raw materials can also cause subtle changes in the physical properties of parts, which are directly reflected in the features of the inspection images. If the identification model remains fixed and is not adjusted according to the cyclical characteristics, "misjudgment drift" will gradually occur. That is, as the cycle progresses, the model's judgment error on defects will increase, making it unable to stably adapt to actual production scenarios in the long term.

[0006] Traditional methods often neglect the systematic analysis of feature paths during image enhancement when constructing defect recognition rules. The optimal image enhancement method differs for different types of part defects. For example, for surface scratch defects, edge enhancement algorithms may be needed to highlight the scratch boundaries; for surface stain defects, contrast adjustment may be needed to emphasize the stain area. Existing technologies fail to dynamically select the optimal enhancement path based on part type and image characteristics, resulting in enhanced images that still cannot effectively highlight defect features, thus reducing the efficiency of defect recognition.

[0007] Existing defect identification systems often employ simplistic scoring mechanisms, relying heavily on single features or basic threshold judgments. However, automotive parts exhibit complex and diverse defect characteristics; a single part may contain multiple defect types, and these defect features can vary subtly. A simplistic scoring method struggles to comprehensively cover the complex defect feature space, easily leading to blind spots and an inability to accurately define the boundary between defects and normal conditions, further impacting the reliability of defect identification. Furthermore, most systems lack periodic self-updating capabilities. When production processes are adjusted, equipment status changes, or raw materials are replaced, the fixed identification model gradually becomes ineffective, requiring manual parameter readjustment. This not only increases maintenance costs but can also cause production downtime, failing to meet the real-time and adaptive requirements of intelligent manufacturing. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for identifying defects in automotive parts based on image enhancement, so as to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for defect identification of automotive parts based on image enhancement, the method comprising:

[0010] Acquire historical inspection image data of parts, determine the processing cycle, construct the corresponding cycle inspection image dataset, and filter and analyze based on part type to determine the cycle characteristic inspection image dataset;

[0011] Based on the periodic feature detection image dataset, corresponding image enhancement processing is performed, the enhancement feature path of the corresponding periodic feature detection image data is constructed, and the score of the corresponding periodic feature detection image data is analyzed to determine the optimal periodic feature detection image data.

[0012] Using the enhanced feature path nodes of the periodic best feature detection image data as reference nodes, clustering processing of the enhanced feature path nodes corresponding to the periodic feature detection image data is performed respectively. The adaptation defect features of the enhanced feature path nodes of each best feature detection image data are analyzed, and the adaptation defect identification score range of the corresponding parts is determined comprehensively.

[0013] Based on the zero-fit defect identification scoring range, the detected image data is scored and filtered; if the score of the detected image data falls within the comprehensive fit defect identification scoring range, it is marked as defect; otherwise, the image data is marked as normal.

[0014] Preferably, the step of acquiring historical inspection image data of parts, determining the processing cycle, constructing a corresponding cycle inspection image dataset, and determining the cycle characteristic inspection image dataset based on part type filtering and analysis includes:

[0015] By calling the historical inspection image database of the parts, the processing cycle window is determined, the historical inspection image data of the parts within the corresponding cycle is retrieved, and the corresponding cycle inspection image dataset is constructed; the cycle inspection image dataset includes the image resolution data and illumination parameter data of the corresponding parts and the corresponding part number data;

[0016] Based on the average clarity of each image in the periodic inspection image dataset for the corresponding part, a type screening analysis is performed to determine the periodic characteristic inspection image dataset after the type screening analysis.

[0017] Preferably, the step of performing corresponding image enhancement processing based on the periodic feature detection image dataset, constructing the enhancement feature path of the corresponding periodic feature detection image data, and analyzing the score of the corresponding periodic feature detection image data to determine the optimal periodic feature detection image data includes:

[0018] Based on the periodic characteristic detection image data in the periodic characteristic detection image dataset, image enhancement processing is performed on the corresponding periodic characteristic detection image data, corresponding enhancement feature nodes are constructed, and the nodes are concatenated to construct the enhancement feature path of the corresponding periodic characteristic detection image data.

[0019] Based on the enhanced feature paths of the corresponding periodic characteristic detection image data, a scoring analysis is performed on the periodic characteristic detection image data; and a corresponding enhanced feature node feature matrix is ​​constructed for the enhanced feature nodes of the periodic characteristic detection image data.

[0020] The enhanced feature node feature matrix is ​​a combination of the image data proportion weight corresponding to the enhanced feature and the detail preservation score; based on the enhanced feature node feature matrix of the image data detected for each periodic characteristic, enhanced feature node scoring analysis is performed;

[0021] Based on the enhanced feature node score data of the corresponding periodic characteristic detection image data, the score data of the corresponding periodic characteristic detection image data is determined by summing the node scores on the enhanced feature path of the corresponding periodic characteristic detection image data.

[0022] Preferably, the step of analyzing the adaptation defect features of the enhanced feature path nodes of each optimal characteristic detection image data, and comprehensively determining the adaptation defect identification scoring range for the corresponding part, includes:

[0023] Based on the scoring data of the image data detected by each periodic feature, the maximum value of the output scoring data is the image data detected by the best periodic feature. Then, each enhanced feature node on the enhanced feature path corresponding to the image data detected by the best periodic feature is taken as a reference node, and the enhanced feature nodes corresponding to the remaining image data detected by each periodic feature in the image data are used for node clustering. According to the node clustering results, the clustering adaptation state of the enhanced feature nodes of the image data detected by the best periodic feature corresponding to each reference node and the enhanced feature nodes corresponding to the remaining image data detected by each periodic feature is determined.

[0024] Rectangular regions are divided with each reference node as the center, and the distribution analysis of the enhanced feature nodes corresponding to the remaining periodic feature detection image data contained in the rectangular regions is performed respectively.

[0025] By gradually expanding the side length of the rectangular region and calculating the ratio of the number of nodes to the area in the rectangular region corresponding to each side length, the node distribution density value of the rectangular region corresponding to each side length is determined, and the rectangular region corresponding to the maximum distribution density value is taken as the best fitted node rectangular region for the corresponding reference node.

[0026] Based on the periodic best characteristics, the best-fitting node rectangular region corresponding to each enhanced feature node in the image data is detected. The type and score of each enhanced feature node in the rectangular region are divided in a coordinated manner. If there are two or more enhanced feature nodes of the same type, the enhanced feature nodes of the same type are classified and filtered, and the maximum and minimum scores of the enhanced feature nodes of the same type are retained respectively.

[0027] In the best-fitting node rectangle region of each enhanced feature node in the periodic best feature detection image data, a full path traversal is performed for different types of enhanced feature nodes starting from the center. For enhanced feature nodes of the same type, a branch traversal is performed respectively, and the traversal path of the maximum value node and the minimum value node in the current best-fitting node rectangle region is output.

[0028] The node scores of the maximum value node traversal path and the minimum value node traversal path are calculated and summed respectively. The upper and lower limits of the adaptation defect recognition score of each enhanced feature node corresponding to the best feature detection image data of the current period are determined respectively, and the adaptation defect recognition score interval corresponding to the current enhanced feature node is constructed.

[0029] By integrating the best feature detection image data of the current cycle with the corresponding enhanced feature node path for each node's adaptation defect recognition score interval, the upper and lower limits of each interval are summed and averaged. The average sum of the calculated upper limits is the upper limit of the cycle's comprehensive adaptation defect recognition score interval, and the average sum of the calculated lower limits is the lower limit of the cycle's comprehensive adaptation defect recognition score interval. Thus, the comprehensive adaptation defect recognition score interval for the current part cycle is constructed.

[0030] Preferably, the method further includes:

[0031] It implements a periodic self-updating cycle, allowing users to set the update time and re-analyze the periodic comprehensive adaptation defect identification score range.

[0032] Preferably, the present invention also includes an image enhancement-based automotive parts defect recognition system, comprising an image acquisition module, an enhancement processing module, a defect recognition module, and a feedback update module, wherein:

[0033] The image acquisition module is used to acquire historical inspection image data of parts, determine the processing cycle, construct the corresponding cycle inspection image dataset, and determine the cycle characteristic inspection image dataset based on the part type screening and analysis.

[0034] The enhancement processing module is used to perform corresponding image enhancement processing based on the periodic characteristic detection image dataset, construct the enhancement feature path of the corresponding periodic characteristic detection image data, analyze the score of the corresponding periodic characteristic detection image data, and determine the optimal periodic characteristic detection image data.

[0035] The defect identification module is used to take the enhanced feature path nodes of the periodic best feature detection image data as reference nodes, perform clustering processing on the enhanced feature path nodes corresponding to the periodic feature detection image data, analyze the adaptation defect features of the enhanced feature path nodes of each best feature detection image data, and comprehensively determine the adaptation defect identification score range of the corresponding part.

[0036] The feedback update module is used to judge the defect marking of the detected image based on the adapted defect recognition scoring interval, and to perform periodic adaptation defect recognition scoring interval cyclic update processing.

[0037] Preferably, the image acquisition module includes an image data acquisition unit and an image data filtering unit, wherein:

[0038] The image data acquisition unit calls the historical inspection image database of the parts to determine the processing cycle window, retrieves the historical inspection image data of the parts within the corresponding cycle, and constructs the corresponding cycle inspection image dataset; the cycle inspection image dataset includes the image resolution data and illumination parameter data of the corresponding parts and the corresponding part number data;

[0039] The image data filtering unit performs type filtering analysis based on the average clarity of each image data in the periodic detection image dataset of the corresponding part, in order to determine the periodic characteristic detection image dataset after type filtering analysis.

[0040] Preferably, the enhancement processing module includes an image enhancement path construction unit and an image data scoring and analysis unit, wherein:

[0041] The image enhancement path construction unit uses the periodic characteristic detection image data in the periodic characteristic detection image dataset to perform image enhancement processing on the corresponding periodic characteristic detection image data, construct corresponding enhancement feature nodes, and concatenate the nodes to construct the enhancement feature path of the corresponding periodic characteristic detection image data.

[0042] The image data scoring and analysis unit is used to score and analyze the image data of each periodic characteristic detection according to the enhanced feature path of the image data of each periodic characteristic detection; it constructs the corresponding enhanced feature node feature matrix for the enhanced feature nodes of the image data of each periodic characteristic detection.

[0043] The enhanced feature node feature matrix is ​​a combination of the image data proportion weight and detail preservation score corresponding to the enhanced feature; the enhanced feature node score analysis is performed based on the enhanced feature node feature matrix of the image data of each periodic feature detection; the score data of the image data of each periodic feature detection is determined by summing the node scores on the enhanced feature path of the image data of each periodic feature detection according to the enhanced feature node score data of the image data of each periodic feature detection.

[0044] Preferably, it includes a feature node clustering processing unit and a defect identification and analysis unit, wherein:

[0045] The feature node clustering processing unit, based on the scoring data of each periodic feature detection image data, outputs the maximum score data as the periodic best feature detection image data; then, taking each enhanced feature node on the enhanced feature path corresponding to the periodic best feature detection image data as a reference node, and performing node clustering processing on the enhanced feature nodes corresponding to the remaining periodic feature detection image data in the periodic feature detection image dataset; based on the node clustering processing results, determining the clustering adaptation state between the enhanced feature nodes of the periodic best feature detection image data corresponding to each reference node and the enhanced feature nodes corresponding to the remaining periodic feature detection image data;

[0046] The defect identification and analysis unit divides rectangular regions centered on each reference node to perform distribution analysis on the enhanced feature nodes corresponding to the remaining periodic characteristic detection image data within each rectangular region. By progressively expanding the side length of the rectangular region and calculating the ratio of the number of nodes to the area in the rectangular region corresponding to each side length, the node distribution density value of the rectangular region corresponding to each side length is determined. The rectangular region corresponding to the maximum distribution density value is taken as the optimal fitting node rectangular region for the corresponding reference node. Based on the optimal fitting node rectangular regions corresponding to each enhanced feature node of the periodic optimal characteristic detection image data, the type and score of each enhanced feature node in the optimal fitting node division rectangular region are considered. If there are two or more enhanced feature nodes of the same type, they are classified and filtered, and the maximum and minimum scores of the enhanced feature nodes of the same type are retained. The optimal fitting node rectangular regions of each enhanced feature node of the periodic optimal characteristic detection image data are then analyzed. Starting from the center, a full path traversal is performed on heterogeneous enhanced feature nodes. For nodes of the same type, branch traversal is performed separately, outputting the traversal paths of the maximum and minimum values ​​in the rectangular region of the current best-fit node. The node scores of the maximum and minimum value traversal paths are calculated and summed to determine the upper and lower limits of the adaptation defect recognition score for each enhanced feature node in the current period's best feature detection image data, constructing the adaptation defect recognition score interval for the current enhanced feature node. The adaptation defect recognition score intervals of each node on the path of the enhanced feature node in the current period's best feature detection image data are integrated, and the upper and lower limits of each interval are summed and averaged. The average sum of the upper limits is output as the upper limit of the period's comprehensive adaptation defect recognition score interval, and the average sum of the lower limits is output as the lower limit of the period's comprehensive adaptation defect recognition score interval. Thus, the comprehensive adaptation defect recognition score interval for the current part period is constructed.

[0047] Preferably, the feedback update module includes a defect marking judgment unit and a parameter self-updating unit, wherein:

[0048] The defect marking judgment unit is based on the comprehensive adaptation defect recognition scoring range within the part cycle. It is used to score and calculate the detection image data and filter it. If the score of the detection image data is in the comprehensive adaptation defect recognition scoring range, it is marked as defect; otherwise, the image data is marked normally.

[0049] The parameter self-updating unit implements a periodic self-updating cycle, which is used to update and re-analyze the periodic comprehensive adaptation defect identification scoring range by allowing users to set the update time themselves.

[0050] Compared with the prior art, the beneficial effects of the present invention are:

[0051] By acquiring historical inspection image data of parts and filtering and analyzing it based on part type, high-quality image data reflecting periodic characteristics can be selectively retained, avoiding the omission of defect features due to insufficient image clarity or fluctuations in lighting parameters, thus laying a data foundation for subsequent analysis.

[0052] In the image enhancement process, by constructing enhancement feature paths and analyzing node scores, the optimal feature detection image data for each period can be dynamically determined. This scoring mechanism based on a feature matrix (including image data proportion weights and detail preservation scores) can quantitatively evaluate the highlighting effect of different enhancement methods on defect features, ensuring that the enhanced image retains sufficient details while highlighting defect areas. Compared with traditional fixed enhancement algorithms, this significantly improves the targeting and effectiveness of image preprocessing.

[0053] Clustering is performed using enhanced feature path nodes in the best-featured detection image data as a reference. Through rectangular region division and node distribution density analysis, the appropriate defect identification scoring range for each feature node can be accurately determined. This process fully considers the feature fluctuations of similar parts in different periods. By retaining the maximum and minimum scores of the same type of nodes, a scoring range covering a wide feature space is constructed, effectively solving the problem that traditional single thresholds cannot adapt to feature changes, making the defect identification scoring mechanism more robust.

[0054] The periodic self-updating mechanism enables dynamic adjustment of the comprehensive adaptation defect identification scoring range through user-defined update times. This function can respond in real time to factors such as changes in equipment status and raw material differences during the production process, ensuring that the identification model is always based on the latest data features. This avoids a decrease in identification accuracy due to production cycle fluctuations, significantly improves the system's long-term applicability and adaptability, and reduces manual maintenance costs.

[0055] At the system level, the collaborative work of various modules (image acquisition module, enhancement processing module, defect identification module, and feedback update module) achieves full automation from data collection to defect labeling. The image acquisition module's filtering unit eliminates low-quality data through average sharpness analysis, ensuring the reliability of the input data; the enhancement processing module's path construction and scoring analysis unit quantifies the enhancement effect; the defect identification module's clustering and analysis unit constructs dynamic scoring intervals; and the feedback update module's labeling and self-updating unit enables real-time detection and model optimization. This modular design not only improves the system's maintainability but also enhances the overall efficiency and accuracy of defect identification through refined processing at each stage. Especially in complex defect types and multi-cycle production scenarios, it exhibits stronger adaptability and higher detection accuracy compared to traditional systems. Attached Figure Description

[0056] Figure 1 This is a schematic diagram illustrating the working principle of the image enhancement-based automotive parts defect identification method described in this invention.

[0057] Figure 2 Flowchart for constructing a periodic detection image dataset;

[0058] Figure 3 A flowchart for image enhancement processing and scoring analysis;

[0059] Figure 4 This is a structural diagram of the image acquisition module. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Please see Figures 1-4 The present invention relates to a method for identifying defects in automotive parts based on image enhancement, the specific implementation steps of which are as follows:

[0062] The process involves acquiring historical inspection image data of parts, determining the processing cycle, constructing a corresponding cycle-specific inspection image dataset, and then performing a filtering analysis based on part type to determine the cycle-specific characteristic inspection image dataset. Specifically, the process first accesses the historical inspection image database of parts, determines the processing cycle window, and retrieves historical inspection image data of parts within the corresponding cycle. The constructed cycle-specific inspection image dataset includes image resolution data, illumination parameter data, and part number data for the corresponding parts. Then, a type filtering analysis is performed based on the average sharpness of each image in this dataset to obtain the cycle-specific characteristic inspection image dataset after type filtering analysis.

[0063] Based on a periodic feature detection image dataset, corresponding image enhancement processing is performed. Enhancement feature paths for the corresponding periodic feature detection image data are constructed, and the scores of the corresponding periodic feature detection image data are analyzed to determine the optimal periodic feature detection image data. Specifically, image enhancement processing is performed on each periodic feature detection image data in the dataset, constructing enhancement feature nodes, and then concatenating these nodes into enhancement feature paths. Next, a feature matrix is ​​constructed for the enhancement feature nodes of each periodic feature detection image data. This matrix is ​​composed of the image data proportion weight of the corresponding enhancement feature and the detail preservation score. Based on this, enhancement feature node score analysis is performed. Then, the scores of nodes on the enhancement feature paths of each periodic feature detection image data are summed to obtain the score data for each periodic feature detection image data. The image data corresponding to the maximum score is the optimal periodic feature detection image data.

[0064] Using the enhanced feature path nodes of the periodic optimal feature detection image data as reference nodes, clustering processing is performed on the enhanced feature path nodes corresponding to the periodic feature detection image data. The adaptation defect features of the enhanced feature path nodes of each optimal feature detection image data are analyzed, and the matching defect recognition scoring interval of the corresponding part is comprehensively determined. Specifically, each enhanced feature node on the enhanced feature path corresponding to the periodic optimal feature detection image data is used as a reference node. Clustering processing is performed on the enhanced feature nodes corresponding to the remaining periodic feature detection image data in the periodic feature detection image dataset. Based on the clustering processing results, the clustering adaptation state between the enhanced feature nodes of each reference node and the remaining nodes is determined. Rectangular regions are divided with each reference node as the center. The distribution of the remaining nodes contained in the rectangular regions is analyzed. By gradually expanding the side length of the rectangular regions, the ratio of the number of nodes to the area in the rectangular regions corresponding to each side length is calculated to determine the node distribution density value. The rectangular region corresponding to the maximum distribution density value is taken as the optimal matching node rectangular region of the reference node. Based on the optimal fitting node rectangle region for each enhanced feature node, the type and score of each enhanced feature node are considered. If two or more enhanced feature nodes of the same type exist, they are categorized and filtered, retaining the nodes with the highest and lowest scores. Within each optimal fitting node rectangle region, a full path traversal is performed for enhanced feature nodes of different types, starting from the center. When nodes of the same type exist, a branch traversal is performed, outputting the traversal paths for the nodes with the highest and lowest scores. The node scores of the two paths are summed to determine the upper and lower limits of the fitting defect identification score for each enhanced feature node, constructing a score interval. The fitting defect identification score intervals for each node are integrated, and the upper and lower limits of each interval are summed and averaged to obtain the upper and lower limits of the comprehensive fitting defect identification score interval for the current part cycle, thus constructing the comprehensive fitting defect identification score interval for the current part cycle.

[0065] Example 1:

[0066] The specific implementation method for acquiring historical inspection image data of parts and constructing a periodic inspection image dataset is as follows:

[0067] The system needs to access a historical image database of the parts, which stores a large amount of historical image information about the parts. When accessing the database, a processing cycle window needs to be determined. This window can be set based on actual production needs and the characteristics of the historical data. For example, if the production cycle of the parts is relatively stable, a weekly processing cycle window can be set; if there are many variable factors in the production process, a monthly or quarterly processing cycle window can be selected. After determining the processing cycle window, the historical image data of the parts within that cycle is retrieved from the database.

[0068] The constructed periodic inspection image dataset contains several key data points. Image resolution data is a crucial component, directly affecting image clarity. Different image resolutions significantly impact defect identification; higher resolution reveals detailed features more clearly, aiding in accurate defect recognition, while lower resolution may make subtle defects difficult to detect. Illumination parameter data records the lighting conditions during image capture, including light intensity, angle, and uniformity. Different lighting conditions cause variations in image brightness and contrast, thus affecting defect assessment. For example, under uneven lighting, certain areas of a part may be too bright or too dark, obscuring defect features. Part number data is used to differentiate and manage image data for different parts. The part number allows for quick retrieval of corresponding part information, facilitating subsequent analysis and processing.

[0069] After obtaining the periodic detection image dataset, the images in this dataset need to be screened and analyzed to ensure that the images used for subsequent processing have high quality. The screening analysis is based on the average sharpness of each image. The calculation of average sharpness requires a specific image sharpness evaluation algorithm, which comprehensively considers multiple feature parameters of the image. For example, the algorithm analyzes the edge sharpness of the image; the higher the edge sharpness, the clearer the boundaries of objects in the image, and the better it reflects the shape and structure of the part. At the same time, the algorithm also considers the image contrast; appropriate contrast can make the differences between different areas of the part more obvious, making it easier to identify defects.

[0070] Specifically, for each image data, its average sharpness value is first calculated using an image sharpness evaluation algorithm. Then, a sharpness threshold is set, which needs to be determined based on actual production requirements and the accuracy requirements of image recognition. Image data with an average sharpness below the set threshold is considered to be of poor quality and cannot accurately reflect the true condition of the parts. For example, if the average sharpness of an image is low, it may be due to equipment malfunction, environmental factors, or other reasons causing the image to be blurry. Such images are prone to misjudgment or missed judgment in the subsequent defect identification process, and therefore need to be screened out.

[0071] After the above screening analysis, the resulting image dataset for periodic characteristic detection is the one obtained after type screening analysis. The images in this dataset have high average sharpness, better meeting the needs of subsequent image enhancement processing and defect identification. This processing ensures the quality of the image data input to subsequent processing stages, laying a solid foundation for accurately identifying defects in automotive parts.

[0072] In practical applications, the process of accessing a database needs to consider the integrity and accuracy of the data. Situations where data is missing or incorrect may occur in the database require appropriate measures, such as communicating with relevant production departments to supplement missing or correct erroneous data. Simultaneously, when determining the processing cycle window, the timeliness of the data must also be considered to ensure that the retrieved image data reflects the current production status of the parts.

[0073] Accurate recording of image resolution and illumination parameters is also crucial. During image acquisition, it's essential to ensure the normal operation of the acquisition equipment, performing regular calibration and maintenance to guarantee that the acquired image data meets the required resolution and illumination parameters. For part number data, strict coding rules must be established to prevent duplicate or incorrect numbers, facilitating accurate management and retrieval of image data.

[0074] Different image sharpness evaluation algorithms may yield different results when calculating average sharpness, so it is necessary to select the appropriate algorithm based on the actual situation. At the same time, the sharpness threshold setting also needs to be optimized through multiple experiments and practical applications to ensure that the selected image data is of good quality without excessively filtering out useful data.

[0075] By calling the historical inspection image database of parts and determining the processing cycle window, a cycle inspection image dataset containing image resolution, lighting parameters and part number is constructed. Then, a cycle characteristic inspection image dataset is obtained based on the average sharpness. This can selectively retain high-quality image data that reflects the characteristics of the production cycle, avoid the omission of defect features caused by low-resolution images, and lay a reliable data foundation for subsequent image enhancement and defect recognition. This solves the problem of inconsistent image quality affecting recognition accuracy in traditional methods.

[0076] Example 2:

[0077] The specific implementation method for image enhancement processing and constructing enhanced feature paths on a periodic feature detection image dataset is as follows:

[0078] For each periodic characteristic detection image in the periodic characteristic detection image dataset, corresponding image enhancement processing is required. The purpose of image enhancement processing is to improve the image quality through specific algorithms and techniques, making the defect features of the part more prominent and obvious, so as to facilitate subsequent analysis and identification.

[0079] Image enhancement processing can employ a variety of algorithms and methods. For example, histogram equalization can enhance image contrast by adjusting the image's histogram, making darker or brighter areas clearer; contrast enhancement techniques can highlight differences between different regions in an image, making edges and defects of parts more prominent; and sharpening can enhance edge details, making the outlines of parts clearer and aiding in the identification of subtle defects. In practical applications, it is necessary to select an appropriate image enhancement algorithm or combine multiple algorithms based on the specific characteristics of the image and the needs of defect identification.

[0080] After each image is enhanced, a corresponding enhanced feature node is generated. Each enhanced feature node represents a unique feature of the image after a specific enhancement process. These features can be texture features, edge features, color features, etc., of a specific region of the image. For example, after histogram equalization, the overall contrast of the image changes, and the resulting feature can be used as an enhanced feature node; after sharpening, the edge details of the image are enhanced, and this edge enhancement feature can also be used as an enhanced feature node.

[0081] After generating enhanced feature nodes, these nodes need to be concatenated in a specific order to construct the enhanced feature path for the corresponding periodic feature detection image data. The construction of the enhanced feature path must consider the flow and logical order of image enhancement processing. For example, if contrast enhancement is performed first, followed by sharpening, then the corresponding enhanced feature nodes need to be concatenated in this order to form a complete enhanced feature path. This path reflects the change process of image features during enhancement processing, starting from the initial image features, undergoing a series of enhancement processes, and ultimately forming image features with prominent defect characteristics.

[0082] Scoring analysis is needed to evaluate the image data for each periodic feature detection to determine the optimal image data for each period. The first step in the scoring analysis is to construct a feature matrix for each enhanced feature node. This feature matrix is ​​composed of the image data proportion weights for the corresponding enhanced feature and the detail preservation score.

[0083] Image data proportion weight reflects the importance of the enhancement feature within the overall image. For example, if an enhancement feature involves a large proportion of the image, its proportion weight is relatively high; conversely, if it only involves a small proportion, its proportion weight is low. Detail preservation score evaluates the preservation of image details after enhancement. Good image enhancement should highlight defective features while preserving as much of the original image detail as possible, avoiding the loss of detail due to enhancement. The detail preservation score can be determined by comparing the detail information of the images before and after enhancement, such as analyzing the degree of preservation of fine structures and the increase in noise.

[0084] Based on the constructed feature matrices of each enhanced feature node, a scoring analysis is performed on the enhanced feature nodes. The scoring analysis process needs to comprehensively consider two factors: the image data proportion weight and the detail preservation score. For example, if an enhanced feature node has a high image data proportion weight and a high detail preservation score, then the node's score will be relatively high; conversely, if the image data proportion weight is low or the detail preservation score is low, then the node's score will also be affected.

[0085] After obtaining the score for each enhanced feature node, it is necessary to sum the node scores on the enhanced feature path of each periodic feature detection image data to obtain the score data of the image data. For example, if there are three enhanced feature nodes on the enhanced feature path of an image, and their scores are 80, 85, and 90 respectively, then the score data of the image data is the sum of these three scores, which is 255.

[0086] By comparing the score data of the characteristic detection images for each period, the image data corresponding to the maximum score is output as the optimal characteristic detection image data for that period. This is because the image data corresponding to the maximum score, after image enhancement processing, has the most prominent defect features and the most complete detail preservation, making it the most suitable reference for subsequent defect identification.

[0087] In practical applications, multiple factors need to be considered when selecting an image enhancement algorithm. Different part types may have different defect features and image characteristics, so it is necessary to select an appropriate enhancement algorithm for the specific part type. For example, for parts with complex surface textures, an enhancement algorithm that can highlight texture features may be needed; for parts with fine cracks, a sharpening algorithm that can enhance edge details may be needed.

[0088] When constructing enhanced feature nodes, it is necessary to accurately extract the feature information of the image. This may require the use of advanced image processing techniques and algorithms, such as convolutional neural networks, to automatically extract image features. Simultaneously, it is essential to ensure that the extracted features accurately reflect the changes after image enhancement processing and have a strong correlation with the defect features of the part.

[0089] When constructing the feature matrix, determining the image data proportion weights and detail preservation scores requires a scientific and reasonable approach. Image data proportion weights can be determined by statistically analyzing the number of pixels in the image regions involved in the enhancement features; the detail preservation score can be evaluated by comparing the enhanced image with the original image and analyzing changes in detail information. For example, the structural similarity index (SSIM) between the enhanced image and the original image can be calculated, and this index can be used as a reference for the detail preservation score.

[0090] When performing scoring analysis on enhanced feature nodes, it is necessary to establish a reasonable scoring model and rules. The scoring model should be able to comprehensively consider the influence of image data proportion weight and detail preservation score, and should have a certain degree of flexibility to be adjusted according to actual needs. For example, different weight coefficients can be set for image data proportion weight and detail preservation score to reflect their importance in the scoring.

[0091] When calculating the scoring data for image data, it is essential to ensure the accuracy of the calculations. Various calculation methods can be used, such as simple summation and weighted summation; the specific method chosen should be determined based on the scoring model and actual needs. Simultaneously, the calculation results need to be verified and calibrated to ensure that the scoring data accurately reflects the image quality and the prominence of defect features.

[0092] The process of image enhancement processing and constructing enhanced feature paths for periodic characteristic detection image datasets, as well as conducting scoring analysis to determine the optimal periodic characteristic detection image data, is a complex process that requires comprehensive consideration of multiple factors and the application of various technologies and methods. By rationally selecting image enhancement algorithms, accurately constructing enhanced feature nodes and feature matrices, and scientifically conducting scoring analysis, high-quality optimal periodic characteristic detection image data can be obtained. This provides a reliable reference for subsequent defect identification and analysis, thereby improving the accuracy and efficiency of defect identification in automotive parts.

[0093] By performing image enhancement processing (such as histogram equalization and sharpening) on ​​periodic feature detection image data, an enhancement feature path composed of concatenated enhancement feature nodes is constructed. Based on the feature matrix containing image data proportion weights and detail preservation scores, the node scores are analyzed, and the optimal periodic feature detection image data is finally determined. This method can dynamically select the optimal enhancement method to ensure that defect features are effectively highlighted under various image conditions (such as low light and low resolution). At the same time, the enhancement effect is quantitatively evaluated. Compared with traditional fixed enhancement algorithms, this method significantly improves the targeting and effectiveness of image preprocessing and solves the problem of blurred defect features in low-quality images.

[0094] Example 3:

[0095] When performing clustering processing and determining the optimal fitted node rectangular region by using the enhanced feature path nodes of the periodically optimal characteristic detection image data as reference nodes, the specific implementation method is as follows:

[0096] Based on the scoring data of the periodic characteristic detection image data, the image data corresponding to the maximum score needs to be determined. This image data is the periodic optimal characteristic detection image data. This process requires comparing the scores of all periodic characteristic detection image data one by one to find the one with the highest score. Because the image with the highest score has a better prominence of defect features and better image quality after image enhancement processing, it can be used as the benchmark for subsequent processing.

[0097] After determining the optimal feature detection image data for each period, the enhancement feature nodes along the corresponding enhancement feature path are used as reference nodes. These reference nodes represent the optimal image enhancement features within the current period. Subsequently, the enhancement feature nodes corresponding to the remaining feature detection image data for each period in the periodic feature detection image dataset need to be clustered with these reference nodes. Node clustering can employ common clustering algorithms, such as K-means clustering or DBSCAN clustering. Taking K-means clustering as an example, its basic principle is to first determine the number of clusters, i.e., the number of reference nodes, and then assign the remaining enhancement feature nodes to different clusters based on their distance from the reference nodes, making the nodes within each cluster as similar as possible.

[0098] Clustering can determine the clustering fit between each reference node and the remaining nodes, based on the optimal feature detection image data augmentation node for each period. This fit reflects the similarity between the remaining nodes and the reference nodes. For example, some nodes may be closely clustered around a reference node, indicating that their features are highly similar to those of the reference node and belong to the same class; while other nodes may be far from the reference nodes, indicating that their features differ significantly from those of the existing reference nodes.

[0099] After clustering, the next step is to divide the data into rectangular regions centered on each reference node. For each reference node, an initial small rectangular region is defined, centered on the reference node, with side lengths determined empirically or by an initial setting. Then, a distribution analysis is performed on the enhanced feature nodes corresponding to the remaining periodic feature detection image data within each rectangular region; that is, the number of nodes contained within that rectangular region is counted.

[0100] To find the region that best reflects the node distribution density, the side length of the rectangular region needs to be gradually expanded. For example, the initial side length is d, and then it is expanded to 2d, 3d, and so on. After each expansion, the ratio of the number of nodes to the area of ​​the rectangular region corresponding to that side length is recalculated. This ratio is the node distribution density value. The larger the node distribution density value, the denser the distribution of nodes in that region, and the more representative it is of the feature distribution around the reference node.

[0101] As the rectangular region's side length is gradually expanded, the node distribution density value corresponding to each side length is recorded. When the maximum distribution density value is found, the rectangular region corresponding to that value is the optimal fitted node rectangular region for the corresponding reference node. For example, when the side length is expanded to 5d, the node distribution density value reaches its maximum. Therefore, the rectangular region centered on the reference node with a side length of 5d is the optimal fitted node rectangular region for that reference node, because the node distribution is most dense within this region, maximizing the inclusion of nodes with characteristics similar to the reference node.

[0102] In practical applications, the characteristics of the data and processing requirements need to be considered when choosing a clustering algorithm. If the data distribution of the enhanced feature nodes is relatively regular, the K-means clustering algorithm may be more suitable; if there is noise or uneven density in the data, the DBSCAN clustering algorithm may be more accurate in identifying the cluster structure. At the same time, the determination of the number of clusters also needs to be careful. The optimal number of clusters, i.e., the number of reference nodes, can usually be determined by methods such as the elbow rule, to avoid the impact of too many or too few clusters on the processing effect.

[0103] When dividing a region into rectangular areas, the initial side length needs to be determined by considering both the distribution range of the enhanced feature nodes and the dimensionality of the feature space. If the initial side length is too small, it may not contain enough nodes, leading to inaccurate calculations of the distribution density. If the initial side length is too large, it may contain too many irrelevant nodes, also affecting the accuracy of the results. Therefore, the initial side length can be reasonably estimated based on the scale of the feature space. For example, it can be calculated by taking the average or a certain proportion of the coordinate range of all enhanced feature nodes as the initial side length.

[0104] When calculating node distribution density, it is necessary to accurately count the number of nodes and the area within each rectangular region. For a two-dimensional feature space, the area of ​​a rectangular region is the square of its side length. For higher-dimensional feature spaces, the method of calculating the area will change accordingly, but the basic principle remains the same. At the same time, the accuracy of the node coordinates must be carefully considered, as any deviation in coordinates may lead to errors in determining whether a node is included within a rectangular region, thus affecting the calculation of the distribution density value.

[0105] Furthermore, the step size needs to be set appropriately during the gradual expansion of the rectangular region's side length. A step size that is too small will increase computation and reduce processing efficiency; a step size that is too large may skip the maximum value of the distribution density, making it impossible to find the optimal fitted node rectangular region. Therefore, the step size can be determined based on the initial side length and the accuracy requirements of the feature space. For example, taking 10% of the initial side length as the step size, and then reducing the step size to improve accuracy when approaching the possible maximum value.

[0106] After determining the optimal fitted node rectangle region, further analysis and processing of the nodes within this region are required to ensure the accuracy of subsequent defect identification scoring intervals. For example, it's possible to check whether the features of the nodes within this region match those of the reference nodes, and whether any abnormal nodes exist. If abnormal nodes are found, their causes need to be analyzed; they may be due to noise interference during image acquisition or errors in the enhancement processing. If necessary, these abnormal nodes can be removed to avoid adverse effects on subsequent processing.

[0107] The process of clustering enhanced feature path nodes in periodically optimal feature detection image data as reference nodes and determining the optimal matching node rectangular region is a step-by-step process of screening and determining the optimal feature region through scientific clustering algorithms and region analysis methods. By accurately selecting clustering algorithms, reasonably dividing rectangular regions, and precisely calculating distribution density values, the node regions with the most similar features and the densest distribution to the reference nodes can be found. This lays the foundation for subsequent analysis of the matching defect features of enhanced feature path nodes in each optimal feature detection image data, thereby improving the accuracy and reliability of determining the scoring interval for automotive part defect identification.

[0108] Clustering is performed using enhanced feature path nodes of the image data with the best periodic characteristics as a reference. By gradually expanding the rectangular region and analyzing the node distribution density, the optimal matching node rectangular region is determined. This method can accurately capture defect patterns with similar features to the reference node, fully consider the feature fluctuations of similar parts in different periods, and provide accurate feature basis for the subsequent construction of a scoring interval covering a wide feature space. This solves the problem of "misjudgment drift" caused by feature fluctuations in traditional methods and improves the adaptability to changes in the production cycle.

[0109] Example 4:

[0110] When determining the matching defect identification scoring interval for each enhanced feature node based on the optimal fitting node rectangular region, the specific implementation method is as follows: After determining the optimal fitting node rectangular region corresponding to each enhanced feature node in the periodic best feature detection image data, it is necessary to conduct a comprehensive analysis of the enhanced feature nodes within each rectangular region. Here, we take the detection image of a car engine cylinder block as an example for illustration. Assume that an enhanced feature node corresponds to the enhanced texture feature of the engine cylinder block surface, and its optimal fitting node rectangular region contains 20 enhanced feature nodes from different detection images. These nodes correspond to the texture feature enhancement results at different positions of the cylinder block.

[0111] The types and scores of each enhanced feature node within the rectangular region should be considered. In practical applications, enhanced feature nodes may include edge-enhanced, texture-enhanced, and contrast-enhanced types. Taking engine block detection as an example, if there are 10 edge-enhanced nodes, 8 texture-enhanced nodes, and 2 contrast-enhanced nodes within the rectangular region, the node types need to be categorized first. For nodes with two or more of the same type of enhanced feature node, further categorization and filtering are required. For example, among the 10 edge-enhanced nodes, if their score ranges from 75 to 92 points, the nodes with the highest score of 92 points and the lowest score of 75 points should be retained, while the remaining nodes should be temporarily excluded. This is done to cover the feature range of this type of node by retaining nodes with extreme scores.

[0112] After filtering nodes of the same type, within the optimal fitted node rectangle area of ​​each enhanced feature node, a full path traversal is performed on the nodes of different enhanced feature types, starting from the central reference node. Continuing with the engine block example, assuming the reference node is edge-enhanced, its rectangular area contains nodes of different types such as texture-enhanced and contrast-enhanced. During the traversal, starting from the reference node, each node of a different type is visited according to the logical order of the feature path; for example, texture-enhanced nodes are visited first, followed by contrast-enhanced nodes, ensuring that all possible path connection methods are traversed.

[0113] When encountering nodes with the same type of enhanced features during traversal, branching traversal is required. For example, when traversing to a texture-enhanced node, if another texture-enhanced node is found nearby, different traversal branches need to be generated, starting from the current node and the other node respectively, to ensure that no possible paths are missed. Taking two texture-enhanced nodes as an example, one with a score of 85 and the other with a score of 78, the traversal path will start from these two nodes respectively, generating different branch paths, and finally outputting the traversal paths of the maximum and minimum value nodes in the current best-fit node rectangle area.

[0114] In the example of the engine block, assuming the traversal path of the maximum value node includes a reference node (edge-enhanced, score 90), a high-scoring texture-enhanced node (88), and a high-scoring contrast-enhanced node (85), the sum of the node scores for this path is 90 + 88 + 85 = 263. The traversal path of the minimum value node includes a reference node, a low-scoring texture-enhanced node (75), and a low-scoring contrast-enhanced node (70), the sum of the scores is 90 + 75 + 70 = 235. Therefore, the upper and lower limits of the adaptation defect identification score for this enhanced feature node are determined to be 263 and 235, respectively, constructing the corresponding score interval [235, 263].

[0115] After constructing the scoring interval for a single enhanced feature node, it is necessary to integrate the adaptive defect recognition scoring intervals of each node on the enhanced feature path corresponding to the best feature detection image data of the current period. Assume there are 5 enhanced feature nodes on this path, with scoring intervals of [235, 263], [210, 240], [190, 220], [200, 230], and [220, 250] respectively. The upper and lower limits of each interval are summed and averaged. The upper limit summation is 263 + 240 + 220 + 230 + 250 = 1203, and the average is 1203 ÷ 5 = 240.6. The lower limit summation is 235 + 210 + 190 + 200 + 220 = 1055, and the average is 1055 ÷ 5 = 211. The final output periodic comprehensive adaptive defect recognition scoring interval is [211, 240.6], which is used for subsequent defect recognition scoring of automotive part inspection images.

[0116] In practice, the classification of node types needs to be determined based on the specific image enhancement algorithm and feature extraction method. For example, nodes processed using histogram equalization can be classified as contrast enhancement nodes, nodes processed using sharpening filters can be classified as edge enhancement nodes, and nodes processed using texture analysis algorithms can be classified as texture enhancement nodes. The selection criteria for nodes of the same type must be consistent to ensure that only nodes with the maximum and minimum scores are retained, so as to avoid blurring of feature ranges due to an excessive number of nodes.

[0117] During full path traversal, traversal rules need to be defined, such as depth-first traversal or breadth-first traversal, to ensure consistency and completeness of the traversal order. For complex feature paths, graph theory traversal algorithms may be needed, treating enhanced feature nodes as vertices of the graph and the relationships between nodes as edges, to achieve automatic full path traversal. When there are branches with the same type of nodes, an independent traversal path needs to be generated for each branch to avoid scoring deviations caused by path intersections.

[0118] During the integration of scoring intervals, attention must be paid to data precision processing. For example, the number of decimal places retained for the average value can be determined based on the accuracy requirements of actual defect identification. In automotive parts inspection, retaining one decimal place is usually sufficient to meet engineering needs. Simultaneously, the rationality of the scoring intervals for each node needs to be verified. If the scoring interval of a certain node differs significantly from other nodes, it is necessary to check whether there are any abnormalities in the enhanced feature processing of that node, or whether feature shift has occurred due to changes in image acquisition conditions. If necessary, image enhancement and feature extraction should be redone.

[0119] For different types of automotive parts, such as gears and bearings, the types of enhanced feature nodes and the scoring criteria may differ. For example, defect identification of gears focuses more on the enhanced features of the tooth edges, while defect identification of bearings focuses more on the enhanced features of the surface texture. Therefore, when coordinating node types and scoring, it is necessary to carry out targeted processing based on the structural characteristics and defect types of the parts to ensure that the scoring range can accurately reflect the range of defect features of the parts.

[0120] By performing type filtering, full path traversal, and score calculation on the nodes within the rectangular area of ​​the best-fit node, a reasonable score range can be constructed for each enhanced feature node. Then, by integrating the score ranges of each node, a periodic comprehensive adaptation defect identification score range can be obtained.

[0121] By coordinating the enhanced feature node types and scores within the rectangular area of ​​the best-fitting node, filtering the maximum and minimum values ​​of nodes of the same type, and then determining the upper and lower limits of the fitting defect identification score for each node through full path traversal, the periodic comprehensive fitting defect identification score range is finally integrated to construct a dynamic range covering various defect features within the period. This replaces the traditional single threshold and can comprehensively cover the feature space of complex defects, effectively solving the problem that a single threshold cannot define the boundary between normal and defective features, and significantly improving the robustness of defect identification.

[0122] Example 5:

[0123] When performing image scoring and periodic updates based on the comprehensive fit defect identification scoring range, the specific implementation method is as follows: Taking the inspection of a certain automobile brake pad as an example, assume that the comprehensive fit defect identification scoring range for this part within a period is [180, 240] constructed through preprocessing. When new brake pad inspection image data is input into the system, the system will calculate the score for the inspection image data in the same way as determining the scoring of the inspection image data with the best characteristics of the period.

[0124] Specifically, the input brake pad detection image undergoes image enhancement processing to construct an enhancement feature path. A feature matrix is ​​then built for each enhancement feature node, consisting of the image data proportion weights of the corresponding enhancement feature and a detail preservation score. For example, after edge enhancement processing, an edge enhancement feature node is generated. Its image data proportion weight is determined based on the coverage area of ​​the edge feature within the entire image, and the detail preservation score is evaluated by comparing the detail information of the images before and after enhancement. Then, based on the feature matrix, a scoring analysis is performed on the enhancement feature nodes, and the scores of each node are summed to obtain the score data for the detection image.

[0125] Suppose a brake pad inspection image scores 210 points, which falls within the comprehensive adaptation defect recognition scoring range [180, 240]. The system will mark this image as defective, indicating that the brake pad may have defects such as surface cracks or uneven wear. Another inspection image scores 170 points, below the lower limit of the range (180 points). The system will mark this image normally, considering that no obvious defects were found in the brake pad under the current inspection conditions.

[0126] To enable the system to adapt to changes in processes and equipment wear during brake pad production, a periodic self-updating cycle is required. Users can set the update time according to actual production conditions, such as weekly. When the set update time is reached, such as 2:00 AM every Monday, the system will automatically trigger the periodic comprehensive adaptation defect identification and scoring interval update and reanalysis process.

[0127] During the update and reanalysis, the system first re-acquires the historical detection image data of the brake pads. This historical detection image data includes newly generated detection image data since the last update. For example, if the last update was last Monday, this update will retrieve all brake pad detection image data from last Monday to this Sunday. Then, the processing cycle window is determined, assuming it is still on a weekly basis, and a corresponding cycle detection image dataset is constructed. This dataset contains the brake pad image resolution data, illumination parameter data, and part number data.

[0128] A type-based filtering analysis is performed based on the average sharpness of each image in the dataset. For example, the average sharpness of each image is calculated using an image sharpness evaluation algorithm, and images below a threshold are filtered out to obtain a periodic feature detection image dataset. Image enhancement processing is then applied to each image in this dataset to construct enhancement feature paths, and the scores of each image are analyzed to determine the optimal periodic feature detection image data.

[0129] Using the enhanced feature path nodes of the newly determined optimal feature detection image data for the period as reference nodes, clustering is performed on the enhanced feature nodes of the remaining images to determine the optimal fitting node rectangle region for each reference node. For example, a certain enhanced feature node of a brake pad corresponds to the texture enhancement feature of the surface wear area, and its optimal fitting node rectangle region contains the texture enhancement feature nodes of the surface wear area in different detection images within that period.

[0130] By coordinating the types and scores of each node within the rectangular area, retaining the maximum and minimum scores for nodes of the same type, and performing a full path traversal of nodes of different types, the matching defect identification score interval for each enhanced feature node is determined, and a new periodic comprehensive matching defect identification score interval is obtained. Assuming that after the update and reanalysis, the new comprehensive matching defect identification score interval is [190, 250], this may be due to recent adjustments in the brake pad manufacturing process, leading to changes in the score range of defect features.

[0131] In practical applications, the scoring calculation of detected images needs to ensure the consistency of the algorithm to avoid inconsistencies in scoring standards due to algorithm changes. For example, when calculating the weight of image data proportions, the same statistical method should always be used, such as calculating based on the proportion of pixels; when evaluating detail preservation scores, the same original image should be used as a reference to ensure the accuracy and comparability of the scores.

[0132] The judgment logic for defect marking needs to be clear and explicit. In addition to marking based on whether the score is within the range, other auxiliary indicators can be combined, such as the location and size of the defect features. However, in this implementation, marking is mainly based on the scoring results. For images marked as defects, the system can automatically generate early warning information to prompt quality inspectors to conduct manual review, thereby improving the reliability of defect identification.

[0133] Setting the update cycle time requires a comprehensive consideration of both production efficiency and the timeliness of data updates. If the update interval is too short, it will increase the system's computational burden and affect detection efficiency; if the interval is too long, it may fail to reflect changes in the production process in a timely manner, leading to a decrease in defect identification accuracy. Therefore, users need to set the update time reasonably based on factors such as the production stability of the part and the frequency of defect occurrence. For example, a longer update cycle can be set for parts with stable production processes, while a shorter update cycle should be set for parts with frequent process changes.

[0134] During the update and reanalysis process, attention must be paid to the management and storage of historical data. As time progresses, historical detection image data accumulates, requiring the system to possess efficient data management capabilities. This includes the ability to quickly retrieve data within a specified period and periodically clean up expired data to free up storage resources. Furthermore, when processing large amounts of data, the computational efficiency of the algorithm needs to be optimized to prevent excessively long update and reanalysis times due to large data volumes, which could impact the normal operation of the system.

[0135] Taking the inspection of automobile wheel hubs as an example, if during a periodic update and reanalysis, a new feature node related to the enhancement of the wheel hub bolt hole edges is found in the enhanced feature path of the new periodic best characteristic detection image data, this may be due to recent wheel hub design changes that have increased the detection requirements for bolt hole edge defects. In this case, when constructing a new comprehensive adaptive defect recognition scoring range, it is necessary to incorporate the scoring range of this newly added feature node to ensure that the scoring range can comprehensively reflect all defect features of the wheel hub.

[0136] Based on the comprehensive and adaptive defect recognition scoring range, the detected images are scored, filtered, and marked for defects. The system can be automatically updated periodically by setting the update time, which can directly complete automated defect detection. At the same time, it responds in real time to factors such as changes in equipment status and differences in raw materials during the production process, ensuring that the recognition model is always based on the latest data features. This avoids the decline in recognition accuracy caused by fluctuations in the production cycle, reduces manual maintenance costs, significantly improves the long-term applicability and adaptability of the system, and meets the real-time requirements of intelligent manufacturing.

[0137] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0138] Although embodiments of the invention 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 invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for defect identification of automotive parts based on image enhancement, characterized in that, The method includes: Acquire historical inspection image data of parts, determine the processing cycle, construct the corresponding cycle inspection image dataset, and filter and analyze based on part type to determine the cycle characteristic inspection image dataset; Based on the periodic feature detection image dataset, corresponding image enhancement processing is performed, the enhancement feature path of the corresponding periodic feature detection image data is constructed, and the score of the corresponding periodic feature detection image data is analyzed to determine the optimal periodic feature detection image data. Using the enhanced feature path nodes of the periodic best feature detection image data as reference nodes, clustering processing of the enhanced feature path nodes corresponding to the periodic feature detection image data is performed respectively. The adaptation defect features of the enhanced feature path nodes of each best feature detection image data are analyzed, and the adaptation defect identification score range of the corresponding parts is determined comprehensively. Based on the zero-fit defect identification scoring range, the detected image data is scored and filtered; if the score of the detected image data falls within the comprehensive fit defect identification scoring range, it is marked as defect; otherwise, the image data is marked as normal. The process of performing corresponding image enhancement processing on the periodic feature detection image dataset, constructing the enhancement feature path for the corresponding periodic feature detection image data, and analyzing the scores of the corresponding periodic feature detection image data to determine the optimal periodic feature detection image data includes: Based on the periodic characteristic detection image data in the periodic characteristic detection image dataset, image enhancement processing is performed on the corresponding periodic characteristic detection image data, corresponding enhancement feature nodes are constructed, and the nodes are concatenated to construct the enhancement feature path of the corresponding periodic characteristic detection image data. Based on the enhanced feature paths of the corresponding periodic characteristic detection image data, a scoring analysis is performed on the periodic characteristic detection image data; and a corresponding enhanced feature node feature matrix is ​​constructed for the enhanced feature nodes of the periodic characteristic detection image data. The enhanced feature node feature matrix is ​​a combination of the image data proportion weight corresponding to the enhanced feature and the detail preservation score; based on the enhanced feature node feature matrix of the image data detected for each periodic characteristic, enhanced feature node scoring analysis is performed; Based on the enhanced feature node score data of the corresponding periodic characteristic detection image data, the score data of the corresponding periodic characteristic detection image data is determined by summing the node scores on the enhanced feature path of the corresponding periodic characteristic detection image data.

2. The method for identifying defects in automotive parts based on image enhancement according to claim 1, characterized in that, The process of acquiring historical inspection image data of parts, determining processing cycles, constructing corresponding cycle inspection image datasets, and determining cycle characteristic inspection image datasets based on part type filtering and analysis includes: By calling the historical inspection image database of the parts, the processing cycle window is determined, the historical inspection image data of the parts within the corresponding cycle is retrieved, and the corresponding cycle inspection image dataset is constructed; the cycle inspection image dataset includes the image resolution data and illumination parameter data of the corresponding parts and the corresponding part number data; Based on the average clarity of each image in the periodic inspection image dataset for the corresponding part, a type screening analysis is performed to determine the periodic characteristic inspection image dataset after the type screening analysis.

3. The method for identifying defects in automotive parts based on image enhancement according to claim 2, characterized in that, The analysis of the enhanced feature path nodes of the optimal characteristic detection image data for each part comprehensively determines the matching defect recognition scoring range for the corresponding part, including: Based on the scoring data of the image data detected by each periodic feature, the maximum value of the output scoring data is the image data detected by the best periodic feature. Then, each enhanced feature node on the enhanced feature path corresponding to the image data detected by the best periodic feature is taken as a reference node, and the enhanced feature nodes corresponding to the remaining image data detected by each periodic feature in the image data are used for node clustering. According to the node clustering results, the clustering adaptation state of the enhanced feature nodes of the image data detected by the best periodic feature corresponding to each reference node and the enhanced feature nodes corresponding to the remaining image data detected by each periodic feature is determined. Rectangular regions are divided with each reference node as the center, and the distribution analysis of the enhanced feature nodes corresponding to the remaining periodic feature detection image data contained in the rectangular regions is performed respectively. By gradually expanding the side length of the rectangular region and calculating the ratio of the number of nodes to the area in the rectangular region corresponding to each side length, the node distribution density value of the rectangular region corresponding to each side length is determined, and the rectangular region corresponding to the maximum distribution density value is taken as the best fitted node rectangular region for the corresponding reference node. Based on the periodic best characteristics, the best-fitting node rectangular region corresponding to each enhanced feature node in the image data is detected. The type and score of each enhanced feature node in the rectangular region are divided in a coordinated manner. If there are two or more enhanced feature nodes of the same type, the enhanced feature nodes of the same type are classified and filtered, and the maximum and minimum scores of the enhanced feature nodes of the same type are retained respectively. In the best-fitting node rectangle region of each enhanced feature node in the periodic best feature detection image data, a full path traversal is performed for different types of enhanced feature nodes starting from the center. For enhanced feature nodes of the same type, a branch traversal is performed respectively, and the traversal path of the maximum value node and the minimum value node in the current best-fitting node rectangle region is output. The node scores of the maximum value node traversal path and the minimum value node traversal path are calculated and summed respectively. The upper and lower limits of the adaptation defect recognition score of each enhanced feature node corresponding to the best feature detection image data of the current period are determined respectively, and the adaptation defect recognition score interval corresponding to the current enhanced feature node is constructed. By integrating the best feature detection image data of the current cycle with the corresponding enhanced feature node path for each node's adaptation defect recognition score interval, the upper and lower limits of each interval are summed and averaged. The average sum of the calculated upper limits is the upper limit of the cycle's comprehensive adaptation defect recognition score interval, and the average sum of the calculated lower limits is the lower limit of the cycle's comprehensive adaptation defect recognition score interval. Thus, the comprehensive adaptation defect recognition score interval for the current part cycle is constructed.

4. The method for identifying defects in automotive parts based on image enhancement according to claim 3, characterized in that, The method further includes: It implements a periodic self-updating cycle, allowing users to set the update time and re-analyze the periodic comprehensive adaptation defect identification score range.

5. An image enhancement-based automotive parts defect recognition system, used to implement the image enhancement-based automotive parts defect recognition method as described in any one of claims 1-4, characterized in that, It includes an image acquisition module, an enhancement processing module, a defect identification module, and a feedback update module, among which: The image acquisition module is used to acquire historical inspection image data of parts, determine the processing cycle, construct the corresponding cycle inspection image dataset, and determine the cycle characteristic inspection image dataset based on the part type screening and analysis. The enhancement processing module is used to perform corresponding image enhancement processing based on the periodic characteristic detection image dataset, construct the enhancement feature path of the corresponding periodic characteristic detection image data, analyze the score of the corresponding periodic characteristic detection image data, and determine the optimal periodic characteristic detection image data. The defect identification module is used to take the enhanced feature path nodes of the periodic best feature detection image data as reference nodes, perform clustering processing on the enhanced feature path nodes corresponding to the periodic feature detection image data, analyze the adaptation defect features of the enhanced feature path nodes of each best feature detection image data, and comprehensively determine the adaptation defect identification score range of the corresponding part. The feedback update module is used to judge the defect marking of the detected image based on the adapted defect recognition scoring interval, and to perform periodic adaptation defect recognition scoring interval cyclic update processing.

6. The image enhancement-based automotive parts defect recognition system according to claim 5, characterized in that, The image acquisition module includes an image data acquisition unit and an image data filtering unit, wherein: The image data acquisition unit calls the historical inspection image database of the parts to determine the processing cycle window, retrieves the historical inspection image data of the parts within the corresponding cycle, and constructs the corresponding cycle inspection image dataset; the cycle inspection image dataset includes the image resolution data and illumination parameter data of the corresponding parts and the corresponding part number data; The image data filtering unit performs type filtering analysis based on the average clarity of each image data in the periodic detection image dataset of the corresponding part, in order to determine the periodic characteristic detection image dataset after type filtering analysis.

7. The image enhancement-based automotive parts defect recognition system according to claim 6, characterized in that, The enhancement processing module includes an image enhancement path construction unit and an image data scoring and analysis unit, wherein: The image enhancement path construction unit uses the periodic characteristic detection image data in the periodic characteristic detection image dataset to perform image enhancement processing on the corresponding periodic characteristic detection image data, construct corresponding enhancement feature nodes, and concatenate the nodes to construct the enhancement feature path of the corresponding periodic characteristic detection image data. The image data scoring and analysis unit is used to score and analyze the image data of each periodic characteristic detection according to the enhanced feature path of the image data of each periodic characteristic detection; it constructs the corresponding enhanced feature node feature matrix for the enhanced feature nodes of the image data of each periodic characteristic detection. The enhanced feature node feature matrix is ​​a combination of the image data proportion weight and detail preservation score corresponding to the enhanced feature; the enhanced feature node score analysis is performed based on the enhanced feature node feature matrix of the image data of each periodic feature detection; the score data of the image data of each periodic feature detection is determined by summing the node scores on the enhanced feature path of the image data of each periodic feature detection according to the enhanced feature node score data of the image data of each periodic feature detection.

8. The image enhancement-based automotive parts defect recognition system according to claim 7, characterized in that, It includes a feature node clustering processing unit and a defect identification and analysis unit, wherein: The feature node clustering processing unit, based on the scoring data of each periodic feature detection image data, outputs the maximum score data as the periodic best feature detection image data; then, taking each enhanced feature node on the enhanced feature path corresponding to the periodic best feature detection image data as a reference node, and performing node clustering processing on the enhanced feature nodes corresponding to the remaining periodic feature detection image data in the periodic feature detection image dataset; based on the node clustering processing results, determining the clustering adaptation state between the enhanced feature nodes of the periodic best feature detection image data corresponding to each reference node and the enhanced feature nodes corresponding to the remaining periodic feature detection image data; The defect identification and analysis unit divides rectangular regions centered on each reference node to perform distribution analysis on the enhanced feature nodes corresponding to the remaining periodic characteristic detection image data within each rectangular region. By progressively expanding the side length of the rectangular region and calculating the ratio of the number of nodes to the area in the rectangular region corresponding to each side length, the node distribution density value of the rectangular region corresponding to each side length is determined. The rectangular region corresponding to the maximum distribution density value is taken as the optimal fitting node rectangular region for the corresponding reference node. Based on the optimal fitting node rectangular regions corresponding to each enhanced feature node of the periodic optimal characteristic detection image data, the type and score of each enhanced feature node in the optimal fitting node division rectangular region are considered. If there are two or more enhanced feature nodes of the same type, they are classified and filtered, and the maximum and minimum scores of the enhanced feature nodes of the same type are retained. The optimal fitting node rectangular regions of each enhanced feature node of the periodic optimal characteristic detection image data are then analyzed. Starting from the center, a full path traversal is performed on heterogeneous enhanced feature nodes. For nodes of the same type, branch traversal is performed separately, outputting the traversal paths of the maximum and minimum values ​​in the rectangular region of the current best-fit node. The node scores of the maximum and minimum value traversal paths are calculated and summed to determine the upper and lower limits of the adaptation defect recognition score for each enhanced feature node in the current period's best feature detection image data, constructing the adaptation defect recognition score interval for the current enhanced feature node. The adaptation defect recognition score intervals of each node on the path of the enhanced feature node in the current period's best feature detection image data are integrated, and the upper and lower limits of each interval are summed and averaged. The average sum of the upper limits is output as the upper limit of the period's comprehensive adaptation defect recognition score interval, and the average sum of the lower limits is output as the lower limit of the period's comprehensive adaptation defect recognition score interval. Thus, the comprehensive adaptation defect recognition score interval for the current part period is constructed.

9. The image enhancement-based automotive parts defect recognition system according to claim 8, characterized in that, The feedback update module includes a defect marking judgment unit and a parameter self-updating unit, wherein: The defect marking judgment unit is based on the comprehensive adaptation defect recognition scoring range within the part cycle. It is used to score and calculate the detection image data and filter it. If the score of the detection image data is in the comprehensive adaptation defect recognition scoring range, it is marked as defect; otherwise, the image data is marked normally. The parameter self-updating unit implements a periodic self-updating cycle, which is used to update and re-analyze the periodic comprehensive adaptation defect identification scoring range by allowing users to set the update time themselves.

Images