Image recognition-based ai server chassis structural component assembly defect detection method

Abstract

The application discloses an AI server case structure assembly defect detection method based on image recognition and belongs to the technical field of case structure image detection. The application is characterized in that the gray scale characteristics of AI server case structure images are analyzed, the AI server case structure image features are extracted after the analysis of the gray scale characteristics, the structure feature extraction eligibility is identified after the extraction of the AI server case structure image features, the AI server case structure assembly defects are identified after the identification of the structure feature extraction eligibility, and the AI server case structure assembly defects are detected based on the obtained identification results, so that the defect detection precision can meet the detection requirements of the precise assembly of the server case structure, and the scene adaptation capability is limited.

Description

Technical Field

[0001] This invention relates to the field of image detection technology for chassis structural components, and more particularly to a method for detecting assembly defects in AI server chassis structural components based on image recognition. Background Technology

[0002] Against the backdrop of the rapid development of AI (Artificial Intelligence) servers towards high computing power and high density, the assembly precision of the server chassis, as the core hardware carrier, directly determines the structural stability of the chassis and the safety of the entire machine operation. However, automated inspection methods face many technical bottlenecks due to the fact that the chassis structural components are mostly made of sheet metal, the assembly parts are complex, and the types of defects are diverse. Therefore, conducting research on high-precision detection of assembly defects in AI server chassis structural components is a key requirement for improving the level of intelligent server manufacturing and ensuring the quality of products leaving the factory.

[0003] The main implementation process for detecting assembly defects in existing chassis structural components is as follows: First, using an industrial camera, a customized ring coaxial light source, and an image acquisition card, multi-view images of the assembled server chassis structural components are acquired at a designated inspection station on the assembly line. Simultaneously, high-speed transmission and initial caching of these images are performed. Next, the acquired raw chassis structural component images undergo preprocessing operations (such as grayscale conversion and noise reduction). Based on feature extraction algorithms (such as edge detection operators and histogram equalization), image features of the chassis structural component assembly scene are extracted. Then, machine vision algorithms, such as NCC (Noise Cross-Correlation) template matching and edge detection algorithms, are used to detect assembly defects in the preprocessed images. The identification results are compared with the initially set chassis structural component assembly process standards and defect judgment benchmarks. Finally, the type of assembly defect (such as missing bolts or deformation), location, and result (such as pass signals and defect detail data) are determined and output.

[0004] Existing technologies for detecting assembly defects in chassis structural components have developed various machine vision-based implementation schemes. For example, the defect detection method and apparatus disclosed in CN114863420B include image detection of the image to be processed to obtain first image information of each object to be tested; determining a second image based on the first image information of each object to be tested; performing image detection on the second image to obtain the target category of the object to be tested; and determining the defect detection result for the object to be tested based on the first image information and target category of each object to be tested.

[0005] Existing chassis structural component assembly defect detection technologies all revolve around image detection and category determination. Their core logic involves first acquiring image information of the object under test, obtaining target-related feature images through image processing, determining the target category through image detection, and finally combining image information and target category to complete defect determination. However, these solutions do not fully consider the material, structure, and production batch, making it difficult to meet the high-precision and comprehensive inspection requirements of precision assembly, thus exposing several technical shortcomings. Specific technical problems are as follows:

[0006] During the assembly defect detection of the internal structural components of the AI ​​server chassis (such as internal screws, module mounting brackets, multi-layer connecting sheet metal, and recessed support clips), including inspections of internal screw fit, module mounting bracket hole alignment, multi-layer connecting sheet metal fit, and recessed support clip installation, it was found that the AI ​​server, in order to adapt to the high-density, high-computing-power module layout, has a large number of compact and hidden structures inside its recesses, at the bottom of deep holes, and at the intersections of multi-layer sheet metal parts (such as side plate fixing screw recesses, deep-hole module mounting holes, and multi-layer sheet metal connecting slots). Furthermore, the metal sheet metal on the inner walls of these structures is mirror-polished, causing the incident detection light to undergo multiple diffuse reflections between the inner walls of these hidden structures. Specular reflections create irregular, high-brightness artifacts (complex light spots) and severe lighting shadows. This type of structural noise is highly mixed with the grayscale, contours, and textures of real assembly features such as screw threads and assembly edges. Existing technologies often use global denoising methods (such as Gaussian denoising and median filtering) when performing image preprocessing (such as denoising, brightness adjustment, and grayscale correction) on such images. This may result in existing technologies being unable to solve the problem of accurately separating structural noise from real subtle features (such as filtering out subtle features such as screw cross grooves and micron-level fitting edges during denoising, leading to feature loss, or failing to completely suppress noise such as complex light spots and shadows, resulting in residual noise interfering with subsequent detection).

[0007] Furthermore, when extracting image features of this part based on feature extraction algorithms (such as edge detection operators, histogram equalization, etc.), due to image blurring and reflection interference, it is impossible to effectively extract the core features of various internal structural components (such as the fitting edge features of screws and grooves, the joint gap features of multi-layer sheet metal, the complete outline features of the support buckle in the groove, etc.). Feature loss and feature misalignment may occur (such as mistakenly extracting the gap features of multi-layer sheet metal as the buckle outline, or mistakenly extracting the fitting edge of screws and grooves as gaps, etc.). In addition, when performing assembly defect detection (such as missing screws, stripped screws, and excessive gaps in multi-layer sheet metal) based on machine vision algorithms (such as normalized cross-correlation template matching, reference value segmentation, etc.), it is impossible to capture various defects with high precision. The subtle features of the defects (such as the deformation features of the cross groove caused by the stripped screw, the incomplete outline features after the support buckle is broken, etc.) cannot be identified, and the blank area features in the groove or mounting position when various structural components are missing (such as the blank area in the groove when the screw is missing, the blank area in the mounting position when the support buckle is missing, etc.) cannot be identified. This leads to the deviation between the defect identification results (such as defect location, defect shape, etc.) and the initial assembly process standards and defect judgment benchmark values ​​(such as misjudging the blurred cross groove caused by the stripped screw as the screw not being tightened properly, and misjudging the feature loss caused by the coverage of reflective bright spots as the screw being missing). Ultimately, this may lead to low accuracy in identifying the defect type (such as micro-cracks in the buckle, stripped screw, missing bracket in the groove, etc.) and the failure to detect hidden defects.

[0008] In summary, existing AI server chassis structural component assembly defect detection methods cannot meet the complex requirements of precision assembly inspection scenarios involving metal sheet metal materials and numerous hidden areas. They lack the adaptive capabilities of image processing and inspection scenarios, the synergistic efficiency of feature extraction and machine vision algorithms is insufficient, and the synergistic adaptability of multi-view acquisition and feature extraction has not been specifically explored. As AI servers widely develop towards higher precision, problems such as image acquisition distortion, incomplete feature extraction, and defect identification bias gradually accumulate and are amplified, leading to a continuous decline in detection accuracy and comprehensiveness. This makes it difficult to meet the inspection needs of precision assembly of server chassis structural components, limiting their scenario adaptability. Summary of the Invention

[0009] To address the accumulating and amplifying problems of image acquisition distortion, incomplete feature extraction, and defect identification bias in existing technologies, which lead to a continuous decline in detection accuracy and comprehensiveness and make it difficult to meet the inspection requirements of precision assembly of server chassis structural components, as well as limited scene adaptability, this invention provides an image recognition-based method for detecting assembly defects in AI server chassis structural components. This method includes: during the defect detection of internal structural components of an AI server chassis, performing grayscale characteristic analysis of the AI ​​server chassis structural component image, and deciding whether to adopt adaptive grayscale correction of the chassis structural component image based on the obtained analysis results; after the grayscale characteristic analysis, performing AI server chassis structural component image feature extraction to extract the core features of the internal structural components of the AI ​​server chassis, and after the AI ​​server chassis structural component image feature extraction, performing structural component feature extraction qualification judgment to quantify the qualification level of feature extraction; after the structural component feature extraction qualification judgment, performing AI server chassis structural component assembly defect identification, and performing AI server chassis structural component assembly defect detection to verify the accuracy of the identification results based on the obtained identification results.

[0010] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0011] 1. The image recognition-based assembly defect detection method for AI server chassis structural components provided by this invention analyzes the grayscale characteristics of AI server chassis structural component images and decides whether to adopt adaptive grayscale correction based on the obtained analysis results. This helps reduce interference such as uneven brightness, excessive reflection, and excessive shadows in AI server chassis structural component images, improving the imaging quality of structural components in hidden areas. After the grayscale characteristic analysis is completed, feature extraction of the AI ​​server chassis structural component images is performed. After the feature extraction is completed, the passability of the extracted structural component features is judged, which helps ensure that the extracted structural component features are complete, clear, and accurately positioned, eliminating defects caused by unqualified features. To mitigate the risk of misjudgment and improve the stability and accuracy of subsequent defect identification, after the structural component feature extraction and qualification judgment are completed, AI server chassis structural component assembly defect identification is performed, and AI server chassis structural component assembly defect detection is performed based on the acquired identification results. This helps to achieve accurate positioning of defect areas, automatic identification of defect types, and effective verification of defect authenticity, thereby achieving high-precision detection of assembly defects of internal structural components of AI server chassis. This effectively solves the problems of image acquisition distortion, incomplete feature extraction, and defect identification deviation in existing technologies, which gradually accumulate and are amplified, leading to a continuous decline in detection accuracy and comprehensiveness, making it difficult to meet the detection requirements of precision assembly of server chassis structural components, and limiting the adaptability of scenarios.

[0012] 2. This invention, by specifically acquiring the global grayscale mean and standard deviation of the structural components in the grayscale image of the AI ​​server chassis, helps to quantify the overall brightness and uniformity of grayscale distribution of the image. It objectively reflects whether the image has quality defects such as uneven brightness, highlights, shadows, and noise. This overcomes the problem of existing technologies using a single grayscale parameter to judge image quality, which cannot comprehensively characterize the degree of grayscale distribution dispersion and is difficult to accurately identify slight grayscale anomalies. Based on the global grayscale mean and standard deviation of the structural components, and comparing the initial grayscale mean range with the initial grayscale standard deviation benchmark value, it determines whether adaptive grayscale correction processing of the chassis structural components is needed. This helps to achieve automatic image quality assessment and adaptive correction, reduce ineffective and over-correction, and improve the intelligence and stability of image preprocessing. It also helps to solve the problem of accurately separating structural noise from real subtle features when using global denoising methods (such as Gaussian denoising, median filtering, etc.) in image preprocessing (such as denoising, brightness adjustment, grayscale correction, etc.) of images interfered with by structural noise.

[0013] 3. When AI server chassis structural components are produced in customized small batches, with slight dimensional fluctuations between batches, and some core features are located in extremely concealed areas (such as the inner wall features of deep-hole module mounting holes), the standard features in the initial feature set cannot be adapted to all batches of structural component features due to batch size fluctuations. Furthermore, the grayscale gradient of features in extremely concealed areas is weak, the edges are blurred, and the position coordinates are easily affected by noise, potentially leading to misjudgments of qualified structural component features. This can affect the accuracy of subsequent defect detection. Therefore, a parallel implementation method for structural component feature extraction and qualification judgment needs to be implemented. This involves specifically acquiring the structural component feature extraction coverage, edge clarity, and position accuracy of various core features in the AI ​​server chassis structural component feature set. The rate helps quantify the extraction quality of individual core features, clarify the integrity, edge clarity, and positional accuracy of features, and overcome the shortcomings of existing technologies that only use a single matching degree for discrimination, cannot take into account the quality of multi-dimensional features, have poor adaptability to batch size fluctuations and features in extremely hidden areas, and are prone to misjudgment and omission. The obtained structural component feature extraction coverage, structural component feature edge clarity, and structural component feature positional accuracy are input into the initial feature quality fusion evaluation model for fusion to obtain quantitative indicators of structural component feature extraction. This helps to achieve comprehensive judgment of multi-dimensional feature quality parameters, take into account the mutual influence relationship between various indicators, reduce the risk of misjudgment caused by the abnormality of a single indicator, and improve the adaptability to customized small batch size fluctuations and the discrimination accuracy of fuzzy features in extremely hidden areas. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 Flowchart of an image recognition-based AI server chassis structural component assembly defect detection method provided in this application embodiment;

[0016] Figure 2 A flowchart outlining the overall process for detecting assembly defects in AI server chassis structural components based on image recognition, as provided in the embodiments of this application.

[0017] Figure 3 The following is a logic diagram of adaptive grayscale correction for the image of the chassis structure component in the AI ​​server chassis structure component assembly defect detection method based on image recognition provided in the embodiments of this application;

[0018] Figure 4 This is a logic diagram for extracting image features of AI server chassis structural components in the AI ​​server chassis structural component assembly defect detection method based on image recognition provided in the embodiments of this application. Detailed Implementation

[0019] The following provides explanations for some of the terms used in this application. It should be noted that these explanations are for the convenience of those skilled in the art and do not constitute a limitation on the scope of protection claimed in this application.

[0020] The embodiments of this application involve at least one, including one or more; wherein, multiple means two or more. Furthermore, it should be understood that in the description of this specification, terms such as "first," "second," and "third" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or order. In the embodiments of this application, "and / or" merely describes an association relationship, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone.

[0021] The directional terms mentioned in the embodiments of this application, such as "up", "down", "left", "right", "inner", and "outer", are only for reference to the directions in the accompanying drawings. Therefore, the directional terms used are for better and clearer explanation and understanding of the embodiments of this application, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the embodiments of this application.

[0022] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0023] Example 1, as Figure 1 The diagram shows a flowchart of an image recognition-based method for detecting assembly defects in AI server chassis structural components, as provided in this application. The method includes the following steps: S1, image grayscale quality monitoring. During the assembly defect detection process of the internal structural components of the AI ​​server chassis, grayscale characteristics are assessed based on the images of the AI ​​server chassis structural components to quantify the uniformity of grayscale distribution and brightness anomalies. Based on the acquired assessment results, a decision is made on whether to adopt adaptive grayscale correction for the chassis structural component images. Adaptive grayscale correction is used to improve problems such as uneven brightness and grayscale distribution in the chassis structural component images, enhance the clarity of grayscale details in hidden structural areas, and suppress interference from bright spots and deep shadows. By monitoring image grayscale quality, it helps to improve image imaging quality from the source, reduce environmental interference such as lighting, shadows, and reflections, and provide a stable and reliable image foundation for subsequent feature extraction and defect identification.

[0024] S2, Feature Extraction Quality Monitoring: After grayscale characteristic assessment, core features (such as screw cross slots, fitting edges, and connection gaps) of the internal structural components of the AI ​​server chassis are extracted from the image features of the AI ​​server chassis structural components. After the image feature extraction of the AI ​​server chassis structural components is completed, the passability judgment of structural component feature extraction is performed to quantify the passability of feature extraction. Through feature extraction quality monitoring, it is helpful to improve the completeness, clarity, and positioning accuracy of the extracted structural features and reduce the risk of misjudgment caused by unqualified features.

[0025] S3, Defect Identification Accuracy Monitoring: After the structural component feature extraction and qualification judgment are completed, AI server chassis structural component assembly defect identification is performed to identify structural component assembly defects. Based on the acquired identification results, AI server chassis structural component assembly defect detection is performed to verify the accuracy of the identification results. Monitoring the accuracy of defect identification helps to achieve accurate positioning, automatic identification and reliable verification of AI server chassis structural component assembly defects, reducing the rate of missed detections and false detections.

[0026] It should be understood that the AI ​​server chassis structural component assembly defect detection method based on image recognition provided in this application requires the construction and continuous maintenance of a feature parameter and defect knowledge database to support the entire detection process during actual deployment and long-term operation. The data sources of this database are multi-dimensional and scenario-based. It includes benchmark parameters directly set by technical personnel in this field in combination with the structural characteristics of the AI ​​server chassis, assembly process specifications and image acquisition equipment parameters, such as the initial grayscale mean range, the initial grayscale standard deviation benchmark value, the initial gradient benchmark value and other core configuration data. It also integrates real-scene images of chassis assembly from multiple batches, test samples under different lighting conditions and various assembly defect annotation datasets, such as image samples of structural components from different batches, feature imaging data of hidden areas, typical assembly defect cases and records of misjudgment and omission correction, etc., to provide a real and reliable data basis for the setting and optimization of various key parameters in the detection process.

[0027] In terms of data storage and management, a hierarchical storage architecture can be adopted. A structured database is used to store standard features and judgment benchmarks to ensure efficient and stable parameter calls. At the same time, an unstructured database is used to store massive amounts of chassis images, core features, and defect detection results, enabling efficient storage and rapid retrieval of large-capacity visual data. Relevant technical personnel can dynamically iterate and update the standard features and judgment benchmarks in the database based on newly added chassis inspection data, on-site application feedback, and defect recognition results, so that the overall inspection solution can better adapt to the inspection needs of different customized chassis structures, small-batch production batch differences, and complex industrial site environments.

[0028] In this embodiment, by monitoring image grayscale quality, feature extraction quality, and defect identification accuracy, a closed-loop quality control system covering the entire process from image preprocessing to feature extraction and then to defect identification is constructed. Specifically, image grayscale quality monitoring provides high-quality image input for feature extraction quality monitoring, feature extraction quality monitoring provides effective feature support for defect identification accuracy monitoring, and defect identification accuracy monitoring achieves accurate determination of the final defect based on the results of the preceding monitoring. The three are sequentially connected and mutually supportive, jointly achieving high-precision and high-stability defect detection of assembly defects in AI server chassis structural components.

[0029] like Figure 2 The diagram shown is a general overview flowchart of the image recognition-based AI server chassis structural component assembly defect detection method provided in this application embodiment. Figure 2It is known that: grayscale characteristic analysis is performed, and the global grayscale mean and standard deviation of the structural components are obtained. It is determined whether the global grayscale mean of the structural components is within the initially set grayscale mean range, and whether the standard deviation of the structural components is less than the initially set grayscale standard deviation benchmark value. If not, adaptive grayscale correction of the chassis structural component image is performed; otherwise, image feature extraction of the AI ​​server chassis structural component is performed. After the image feature extraction of the AI ​​server chassis structural component is completed, assembly defect identification of the AI ​​server chassis structural component is performed, and assembly defect detection of the AI ​​server chassis structural component is performed based on the obtained assembly defect identification results. It is determined whether the defect confidence of the chassis structural component corresponding to the component assembly defect undetermined area is less than the initially set defect confidence benchmark value. If so, the corresponding component assembly defect undetermined area is marked as the structural component assembly qualified area; otherwise, the defect details information is sent to the defect detection control center.

[0030] Preferably, the specific process for grayscale characteristic assessment is as follows: A1, the acquired original AI server chassis structural component image is converted to grayscale to obtain a grayscale image of the AI ​​server chassis structural component; the grayscale conversion process means converting the original color AI server chassis structural component image into a single-channel grayscale image based on a weighted average grayscale conversion algorithm, which is used to reduce the interference of redundant information in the color channel on the assessment results; the global grayscale mean and grayscale standard deviation of the structural component in the grayscale image of the AI ​​server chassis structural component are obtained. The global grayscale mean is used to characterize the overall brightness of the grayscale image of the AI ​​server chassis structural component, and the grayscale standard deviation is used to characterize the grayscale distribution dispersion of the grayscale image of the AI ​​server chassis structural component, quantifying the brightness uniformity of the grayscale image of the AI ​​server chassis structural component.

[0031] Specifically, the process for obtaining the global grayscale mean of the structural component is as follows: the grayscale values ​​of all pixels in the grayscale image of the AI ​​server chassis structural component monitored by the industrial high-definition area array camera are summed to obtain the grayscale sum value of the structural component image; the ratio of the grayscale sum value of the structural component image to the total number of pixels monitored by the industrial image acquisition instrument is calculated to obtain the global grayscale mean of the structural component.

[0032] The specific formula for the global grayscale mean of structural components is:

[0033]

[0034] Where μ represents the global grayscale mean of the structural component, i = 1, 2, ..., N, i represents the horizontal coordinate number (i.e., column coordinate) of the pixel in the grayscale image of the structural component, w = 1, 2, ..., M, w represents the vertical coordinate number (i.e., row coordinate) of the pixel in the grayscale image of the structural component, x represents the horizontal coordinate of the pixel in the grayscale image of the structural component, y represents the vertical coordinate of the pixel in the grayscale image of the structural component, f(x,y) represents the grayscale value of the pixel with coordinates (x,y) in the grayscale image of the AI ​​server chassis structural component, M represents the width (i.e., total number of columns) of the grayscale image of the AI ​​server chassis structural component, N represents the length (i.e., total number of rows) of the grayscale image of the AI ​​server chassis structural component, and the pixel grayscale value is represented by traversing the pixel grayscale value. The single-channel grayscale quantization of the pixel with coordinates (x, y) in the grayscale image of the AI ​​server chassis structural component extracted by the algorithm should be noted. Before calculating the global grayscale mean of the structural component, a two-dimensional Cartesian coordinate system is pre-constructed for the grayscale image of the AI ​​server chassis structural component. The upper left corner vertex of the grayscale image of the AI ​​server chassis structural component, i.e., the coordinate point (1, 1), is used as the origin. The horizontal direction to the right of the image is set as the horizontal coordinate axis, and the vertical direction to the down of the image is set as the vertical coordinate axis. Each column of pixels on the horizontal coordinate axis is numbered sequentially, and each row of pixels on the vertical coordinate axis is numbered sequentially, so that each pixel in the grayscale image corresponds to a unique horizontal coordinate number and a unique vertical coordinate number.

[0035] Specifically, the grayscale standard deviation of structural components represents the standard deviation of all pixels in the grayscale image of the AI ​​server chassis structural components, obtained based on the standard deviation formula. The specific formula for the grayscale standard deviation of structural components is as follows:

[0036]

[0037] Where σ represents the standard deviation of grayscale of the AI ​​server chassis structural components, and μ represents the global mean grayscale of the structural components.

[0038] A2. Based on the global grayscale mean and standard deviation of the structural components, compare the initial grayscale mean range with the initial grayscale standard deviation benchmark value to determine whether adaptive grayscale correction processing of the chassis structural components is required. If the global grayscale mean of the structural components is within the initial grayscale mean range and the grayscale standard deviation of the structural components is less than the initial grayscale standard deviation benchmark value, then the grayscale characteristics of the AI ​​server chassis structural component grayscale image are determined to meet the judgment requirements. The corresponding AI server chassis structural component grayscale image is marked as a qualified chassis structural component grayscale image, and further processing is performed based on the qualified chassis structural component grayscale image. If the AI ​​server chassis structural component image features are not extracted, it indicates that the grayscale image of the AI ​​server chassis structural component has problems such as uneven brightness and abnormal grayscale distribution. The corresponding AI server chassis structural component grayscale image is marked as a non-compliant chassis structural component grayscale image, and adaptive grayscale correction of the chassis structural component image is performed based on the non-compliant chassis structural component grayscale image. The initial grayscale mean range includes the two endpoints of the initial grayscale mean range, which is set in advance by the preset personnel. The initial grayscale standard deviation benchmark value is represented by the average value of the grayscale standard deviation of the structural components over a historical time period.

[0039] As described above, the analysis of grayscale characteristics helps to quantify the overall brightness and uniformity of grayscale images of AI server chassis structural components, objectively identify quality defects such as uneven brightness, highlights, deep shadows, and grayscale noise in the images, and provide a quantitative basis for subsequent image preprocessing. This enables precise and adaptive preprocessing of chassis structural component images, providing a stable and reliable high-quality image foundation for subsequent analysis.

[0040] like Figure 3 The diagram shown is a logic diagram for adaptive grayscale correction of chassis structural component images in the AI ​​server chassis structural component assembly defect detection method based on image recognition provided in this application embodiment. Figure 3It can be seen that: grayscale adaptive correction is performed on the image of the chassis structural component, and preliminary grayscale adaptive correction processing is performed to obtain a grayscale image of the chassis structural component after preliminary correction. For the grayscale image of the chassis structural component after preliminary correction, noise regions are initially located, and the grayscale gradient value of the grayscale image of the chassis structural component after preliminary correction is obtained. Based on the grayscale gradient value, noise and feature regions are distinguished. It is determined whether the grayscale gradient value is greater than the initially set gradient reference value. If it is, the region where the corresponding pixel is located is initially determined to be a noise interference region of the structural component image; otherwise, the region where the corresponding pixel is located is initially determined to be a true feature region of the chassis structural component image. Based on the noise interference region of the structural component image and the true feature region of the chassis structural component image, grayscale similarity and texture similarity are obtained to determine the grayscale similarity. If the similarity of degree and texture meets the similarity judgment conditions, a preliminary noise region positioning failure prompt is sent; otherwise, the noise interference intensity and the clarity of the real features are calculated, and the noise interference intensity and the clarity of the real features are obtained. It is determined whether the noise interference intensity is greater than the initial noise intensity benchmark value. If so, noise denoising processing is performed on the overall noise interference area of ​​the structural component image. At the same time, it is determined whether the clarity of the real features is less than the initial feature clarity benchmark value. If so, the real region enhancement processing is performed on the overall real feature area of ​​the chassis structural component image. If the noise interference intensity is not greater than the initial noise intensity benchmark value and the clarity of the real features is not less than the initial feature clarity benchmark value, then the AI ​​server chassis structural component image feature extraction is performed.

[0041] Preferably, the specific process of adaptive grayscale correction for chassis structural component images is as follows: B1, based on the adaptive histogram equalization algorithm, perform preliminary adaptive grayscale correction processing on the grayscale images of unqualified chassis structural components to obtain a pre-corrected grayscale image of the chassis structural components. This is used to enhance the grayscale details of hidden structural areas (such as the inside of grooves and the bottom of deep holes) and suppress complex light spots with excessively high brightness and shadows with excessively low brightness in the image; B2, perform preliminary noise region localization on the pre-corrected grayscale image of the chassis structural components. The specific process is as follows: calculate the grayscale gradient value of each pixel in the pre-corrected grayscale image of the chassis structural components. The grayscale gradient value is used to characterize the grayscale change amplitude between the pixel and the surrounding pixels in the pre-corrected grayscale image of the chassis structural components; the grayscale gradient value is represented by the result of differential operation and modulo operation on the grayscale values ​​of the pixel and its neighboring pixels; based on the grayscale gradient value, noise and feature... Region discrimination: If the grayscale gradient value of the corresponding pixel in the grayscale image of the chassis structural component after preliminary correction is greater than the initial gradient reference value, the region where the corresponding pixel is located is initially determined to be a noise interference region (such as complex light spots, lighting shadows, etc.) of the structural component image. Conversely, the region where the corresponding pixel is located is initially determined to be a real feature region of the chassis structural component image. The initial gradient reference value is represented by the average grayscale gradient value of the corresponding pixel in the grayscale image of the chassis structural component after preliminary correction over a historical time period. The region where the pixel is located represents the neighborhood region centered on the pixel and divided according to the initial size (such as 3×3 or 5×5 pixels) to avoid misjudgment of a single pixel and improve the accuracy of region positioning. The initial size is set in advance by the preset personnel. Based on the noise interference region of the structural component image and the real feature region of the chassis structural component image, grayscale similarity and texture similarity are obtained.

[0042] Specifically, the acquisition process for grayscale similarity and texture similarity is as follows: Based on connected component analysis, the noisy interference regions of the structural component image are clustered and merged to obtain the overall noisy interference region of the structural component image. Simultaneously, based on connected component analysis, the true feature regions of the chassis structural component image are clustered and merged to obtain the overall true feature region of the chassis structural component image. Grayscale similarity, used to quantify the difference in grayscale distribution between the two types of regions, is acquired as follows: Grayscale values ​​of the overall noisy interference region and the overall true feature region are extracted separately. Based on the extracted grayscale values, grayscale pixel matrices corresponding to the two types of regions are constructed respectively. Gaussian filtering is then applied to the two constructed grayscale pixel matrices respectively. The filtering preprocessing removes isolated extreme gray-level noise points in both types of regions, preserving the gray-level distribution characteristics of the regions themselves to the maximum extent while eliminating noise. The two gray-level pixel matrix values ​​after Gaussian filtering preprocessing are input into a normalized cross-correlation algorithm to obtain gray-level similarity values. The specific construction process of the gray-level pixel matrix is ​​as follows: using the pixel coordinates of the two types of regions as matrix indices, the gray-level value corresponding to each pixel is filled into the matrix element of the corresponding coordinates one by one, forming a two-dimensional matrix that perfectly matches the region size, ensuring that the matrix can completely reflect the gray-level situation of each pixel in the corresponding region, and that the gray-level situation of each pixel corresponds to its position in the image, and that the gray-level value is within a reasonable range.

[0043] Specifically, texture similarity, used to quantify the differences in texture details between two types of regions, is obtained through the following process: Based on the gray-level co-occurrence matrix (GLCM) algorithm, the GLCMs of the overall noise interference region of the structural component image and the overall true feature region of the chassis structural component image are obtained (including the GLCMs of the overall noise interference region and the overall true feature region). The two GLCMs are then standardized. Standardization involves subtracting the minimum value of all elements in each of the two GLCMs from the value of the element in that matrix, and then dividing by the difference between the maximum and minimum values. This process maps all elements of the two matrices to a fixed, reasonable range, eliminating the dimensional differences between the two matrices and ensuring the rationality of subsequent similarity calculations. Based on weighted averages... The summation method involves fusing the gray-level co-occurrence matrices of the two standardized regions to obtain a texture similarity value. The fusing calculation involves performing correlation operations on the two standardized gray-level co-occurrence matrices. First, the absolute value of the difference between corresponding elements in the two matrices is calculated (i.e., the absolute value of the difference between an element in the gray-level co-occurrence matrix of the noise-affected region and the corresponding element in the gray-level co-occurrence matrix of the true feature region). Then, this absolute value is subtracted from the upper limit of a fixed reasonable range to obtain the similarity contribution value of the corresponding elements in the matrices. Finally, the similarity contribution values ​​of all elements in both matrices are multiplied by their respective weights and summed to obtain the texture similarity value. The obtained gray-level similarity and texture similarity are then compared with the corresponding similarity judgment criteria for structural component image similarity.

[0044] The structural component image similarity comparison involves comparing grayscale similarity and texture similarity with similarity judgment criteria. If grayscale similarity and texture similarity do not meet the similarity judgment criteria, it indicates that the image region of the initial noise region is accurately located. Based on the noise interference region of the structural component image and the real feature region of the chassis structural component image, the noise interference intensity and real feature clarity are calculated. Based on the noise interference intensity and real feature clarity, a decision is made on whether to adopt structured image denoising and enhancement collaborative processing. Conversely, if they do not meet the criteria, it indicates that the image region of the initial noise region is confused, and a noise region initial location failure prompt is sent. The similarity judgment criteria indicate that the grayscale similarity is greater than the initial grayscale similarity benchmark value, and the texture similarity is greater than the initial texture similarity benchmark value. The initial grayscale similarity benchmark value is represented by the average grayscale similarity over a historical time period, and the initial texture similarity benchmark value is represented by the average texture similarity over a historical time period.

[0045] Specifically, the process for calculating noise interference intensity and true feature sharpness is as follows: The noise interference intensity of the overall noise interference region of the structural component image and the true feature sharpness of the overall true feature region are calculated separately. Based on the grayscale standard deviation and grayscale extreme value difference of the overall noise interference region of the structural component image, the noise interference intensity is obtained, which is used to quantify the interference degree of complex light spots and lighting shadows. The specific calculation process is as follows: First, the grayscale values ​​of all pixels in the overall noise interference region of the structural component image are extracted based on pixel extraction algorithms (such as region pixel traversal algorithms, neighborhood grayscale sampling algorithms, etc.). The standard deviation of the gray values ​​of all pixels within the region is used as the gray standard deviation; the difference between the maximum and minimum gray values ​​of all pixels within the region is used as the gray extreme difference; then, the gray standard deviation and the gray extreme difference are normalized, and the weighted average of the two is taken as the noise interference intensity; the normalization process means mapping the gray standard deviation and the gray extreme difference to the interval [0,1], specifically by dividing each parameter value by the global maximum value of the corresponding parameter to eliminate dimensional differences and ensure that the weights of each parameter are balanced during the weighted calculation. The specific formula for the noise interference intensity is:

[0046]

[0047] Where P represents the noise interference intensity, α is the initial noise weighting coefficient, σ represents the gray standard deviation of the overall noise interference area, Gmax represents the maximum gray value of the pixel in the overall noise interference area, and Gmin represents the minimum gray value of the pixel in the overall noise interference area of ​​the structural component image.

[0048] Based on the mean grayscale gradient and edge contrast of the overall region of the true feature area in the chassis structural component image, the true feature sharpness is obtained. This is used to quantify the recognizability of the core features of the structural component (such as screw cross slots, fitting edges, and joint gaps). The specific calculation process is as follows: First, the grayscale gradient values ​​of all pixels within the true feature area of ​​the chassis structural component image are extracted using pixel extraction algorithms (such as region pixel traversal algorithms and neighborhood grayscale sampling algorithms). The average value of all grayscale gradient values ​​is calculated (i.e., the mean grayscale gradient). Then, the average grayscale difference between the edge pixels of the core feature and the surrounding pixels within this area is calculated (i.e., the edge contrast). After normalizing the mean grayscale gradient and the edge contrast, the weighted average of the two is taken as the true feature sharpness. The specific formula for the true feature sharpness is:

[0049]

[0050] Where C represents the true feature sharpness, β is the initial sharpness weighting coefficient, Gmean represents the mean gray-level gradient of the true feature region of the chassis structural component image, and Cmean represents the edge contrast of the true feature region of the chassis structural component image.

[0051] Specifically, the structured image denoising and enhancement collaborative processing means that if the detected noise interference intensity is greater than the initial noise intensity benchmark value, it indicates that the light spot and shadow interference in the noise interference area of ​​the structural component image has not been effectively suppressed, and noise interference area denoising processing needs to be performed on the noise interference area of ​​the structural component image. The initial noise intensity benchmark value is represented by the average noise interference intensity over a historical time period. If the detected true feature clarity is less than the initial feature clarity benchmark value, it indicates that the core features of the true feature area of ​​the chassis structural component image are blurred and cannot meet the defect detection requirements, and true feature area enhancement processing needs to be performed on the true feature area of ​​the chassis structural component image. The initial feature clarity benchmark value is represented by the average true feature clarity over a historical time period. If the noise interference intensity is not greater than the initial noise intensity benchmark value, and the true feature clarity is not less than the initial feature clarity benchmark value, then the corresponding pre-corrected chassis structural component grayscale image is marked as a qualified chassis structural component grayscale image, and AI server chassis structural component image feature extraction is performed based on the qualified chassis structural component grayscale image. Structured image denoising and... After the enhanced collaborative processing is completed, the noise interference intensity and the clarity of the true features are reacquired. If the noise interference intensity is still greater than the initial noise intensity benchmark value, or the clarity of the true features is still less than the initial feature clarity benchmark value, an image collaborative processing failure prompt is sent. Otherwise, the corresponding pre-corrected grayscale image of the chassis structure is marked as a qualified grayscale image of the chassis structure, and AI server chassis structure image feature extraction is performed based on the qualified grayscale image of the chassis structure. The interference region denoising process is based on the bilateral filtering algorithm to suppress noise in the interference region. While suppressing noise interference, it can retain the edge details of the true features around the noise region, avoiding excessive denoising that leads to feature loss. After processing, the denoised noise interference region image is output. The true region enhancement process is based on the Laplacian operator enhancement algorithm to perform second-order difference operations on the grayscale values ​​of the true feature region of the chassis structure image. This is used to enhance the grayscale contrast and edge clarity of core features (screw cross slots, fitting edges, etc.) and weaken the influence of slight residual noise. After processing, the enhanced true feature region image is output.

[0052] It should be understood that this embodiment involves several key parameters used to characterize the degree of image interference and feature difference in the process of calculating texture similarity and noise interference intensity. Before carrying out the relevant quantitative calculations, it is necessary to determine the corresponding weights and weighting coefficients based on the actual impact of each parameter on the final evaluation result. Specifically, image samples covering different lighting environments, different types of interference, and different structural components are collected first, covering the actual values ​​of parameters such as gray-level co-occurrence matrix element distribution, gray-level standard deviation, and gray-level extreme difference, while removing abnormal data caused by factors such as device jitter and instantaneous strong light. Then, through correlation analysis (such as Pearson correlation analysis) and fitting calculation (such as linear regression and ridge regression), the contribution of gray-level standard deviation and gray-level extreme difference to noise interference intensity, as well as the influence of each element of the gray-level co-occurrence matrix on texture similarity, are quantified respectively. This determines the corresponding weights and coefficients in the weighted calculation, providing parameter basis that fits the actual detection scenario for subsequent accurate quantification of noise interference and texture difference.

[0053] As described above, adaptive grayscale correction of chassis structural components helps to specifically improve problems such as uneven brightness, abnormal grayscale distribution, and local over-brightness or under-brightness in images. It reduces the risk of feature loss and misjudgment caused by light interference, shadow occlusion, reflection, etc., and improves the grayscale detail recognition and imaging quality of hidden structural areas such as deep holes, grooves, and gaps inside the AI ​​server chassis. It achieves adaptive optimization and high-quality preprocessing of grayscale images, providing a more reliable, clear, and stable image foundation for subsequent feature extraction, qualification judgment, and defect identification.

[0054] like Figure 4 The diagram shown is a logic diagram for image feature extraction of AI server chassis structural components in the AI ​​server chassis structural component assembly defect detection method based on image recognition provided in this application embodiment. Figure 4 The process involves: extracting image features from AI server chassis structural components and obtaining a feature set for these components; determining the suitability of extracted features based on this feature set and obtaining the feature matching degree; judging whether the feature matching degree is greater than the initial matching degree benchmark value; if so, marking the corresponding single core feature as a qualified extracted structural component feature; otherwise, marking the corresponding single core feature as an unqualified extracted structural component feature; determining the suitability of extracted image features based on the qualified and unqualified structural component features and obtaining the structural component feature extraction pass rate; judging whether the structural component feature extraction pass rate is greater than the initial overall pass threshold; if not, sending a structural component feature extraction failure prompt; otherwise, identifying assembly defects in the AI ​​server chassis structural components.

[0055] Preferably, the AI ​​server chassis structural component image feature extraction represents the extraction of various structural component features inside the AI ​​server chassis based on feature extraction algorithms (such as edge detection operators, histogram equalization, etc.), resulting in an AI server chassis structural component feature set (such as the cross-groove contour features of internal screws, the fitting edge features of screws and grooves, the connection gap features of multi-layer sheet metal, the complete contour features of support buckles in grooves, the hole features of module mounting brackets, etc.); based on the AI ​​server chassis structural component feature set, the pass / fail judgment of structural component feature extraction is performed, specifically as follows: the AI ​​server chassis structural component feature set is used for... The single core feature and the corresponding standard features in the initial feature set (such as the standard features of the internal screw cross groove, the standard features of the module mounting bracket hole position, and the standard features of the support buckle contour in the groove) are input into the normalized cross-correlation algorithm. The output matching degree value (range 0-1) is used as the structural component feature matching degree to characterize the accuracy of feature extraction. The initial feature set is set in advance by the preset personnel. The matching degree value represents the contour shape similarity between the output single core feature and the corresponding standard feature. The closer the value is to 1, the more the extracted single core feature matches the standard feature, and the higher the accuracy of feature extraction.

[0056] The qualification of single core feature extraction is determined based on the structural component feature matching degree. The qualification determination of single core feature extraction means judging whether the structural component feature matching degree of a single core feature is greater than the initial matching degree benchmark value. If it is, the corresponding single core feature is marked as a qualified structural component feature; otherwise, it is marked as an unqualified structural component feature. The initial matching degree benchmark value is represented by the average value of structural component feature matching degrees over a historical time period. The qualification of structural component image feature extraction is judged as follows: Based on the number of qualified structural component features extracted, the structural component feature extraction qualification rate is obtained, which is used to characterize the overall quality of structural component feature extraction. The structural component feature extraction qualification rate is represented by the ratio of the number of qualified structural component features monitored by the industrial vision feature statistics instrument to the total number of single core features in the AI ​​server chassis structural component feature set.

[0057] The pass rate of structural component feature extraction is compared with the initial overall pass threshold. If the pass rate is greater than the initial overall pass threshold, the image feature extraction of the AI ​​server chassis structural components is deemed qualified, and the corresponding AI server chassis structural component feature set is marked as a qualified AI server chassis structural component feature set. Based on the qualified AI server chassis structural component feature set, assembly defects of the AI ​​server chassis structural components are identified. The initial overall pass threshold is represented by the average pass rate of structural component feature extraction over a historical period. If the pass rate is not greater than the initial overall pass threshold, the type of unqualified structural component features (such as internal screw cross slot features, module mounting bracket hole features, groove support buckle contour features, etc.), structural component feature matching degree, and the batch information of the structural component (such as batch number, production date, customized specifications, etc.) are recorded and sent to the defect detection and control center, along with a structural component feature extraction failure prompt.

[0058] As described above, AI server chassis structural component image feature extraction helps to accurately locate and extract key assembly features such as screws, clips, holes, and gaps inside the chassis, reducing the impact of irrelevant background, noise interference, and redundant image information on subsequent inspection processes, improving the effectiveness, relevance, and completeness of feature data, and achieving efficient conversion from complex grayscale images to core structural features.

[0059] Preferably, the AI ​​server chassis structural component assembly defect identification method involves inputting the feature set of qualified AI server chassis structural components into an initial defect identification model (such as a YOLO object detection model, a CNN convolutional neural network model, etc.), outputting the structural component assembly defect identification results, and performing AI server chassis structural component assembly defect detection based on the structural component assembly defect identification results. The structural component assembly defect identification results include the undetermined area of ​​structural component assembly defects in the AI ​​server chassis, the type of chassis structural component defect, and the confidence level of the chassis structural component defect. The undetermined area of ​​structural component assembly defects represents the local area of ​​the structural component that is initially identified by the defect identification model as having anomalies and needs further verification, such as areas with missing screws, misaligned clips, excessive gaps, misaligned sheet metal, misaligned holes, etc. The type of chassis structural component defect represents the specific type of assembly defect identified by the defect identification model, such as missing screws, tilted screws, sheet metal damage, misaligned holes, etc. The component defect confidence score represents the degree of credibility of the defect recognition model in identifying the undetermined areas of structural component assembly defects and the types of defects in chassis structural components. A higher score indicates more reliable identification results and a lower probability of misjudgment. The specific training process for the initial defect recognition model is as follows: First, a sample dataset of AI server chassis structural components containing different lighting environments and different defect types is constructed. The defect areas and types in the sample data are labeled, and the dataset is divided into training and validation sets. Second, the divided training set data is input into the model for iterative training, continuously adjusting the model's internal parameters to gradually learn the distinction between normal and defective features of structural components. Finally, the trained model is tested and validated using the validation set. Model parameters are optimized based on recognition accuracy and recall until the model reaches the preset defect recognition accuracy requirements (e.g., recognition accuracy greater than a preset accuracy threshold and recall greater than a preset recall threshold), completing the training and putting the model into use.

[0060] Specifically, the process for detecting assembly defects in AI server chassis structural components is as follows: The confidence level of the chassis structural component defect is compared with the initial defect confidence benchmark value. If the confidence level of the chassis structural component defect corresponding to the undetermined assembly defect area is less than the initial defect confidence benchmark value, the undetermined assembly defect area is determined to be a false defect area, and the corresponding undetermined assembly defect area is marked as a qualified assembly area. The initial defect confidence benchmark value is represented by the average confidence level of chassis structural component defects over a historical time period. If the confidence level of the chassis structural component defect is not less than the initial defect confidence benchmark value, the undetermined assembly defect area is determined to be a structural component assembly defect area, and the defect details are recorded and sent to the defect detection control center. The defect details information indicates the assembly defects of the AI ​​server chassis structural components. The complete characterization information of assembly defects includes the specific location of the defect area in the structural component assembly, the defect type of the chassis structural component, and the deviation value of the defect characteristic parameters (such as the cross groove contour deviation of a stripped screw, the width difference of sheet metal gap exceeding tolerance, the coordinate difference of bracket offset, etc.). The deviation value of the defect characteristic parameters represents the difference between the actual characteristic parameters of the structural component assembly defect area (such as the contour dimension of the cross groove when the screw is stripped, the width of the gap between multiple layers of sheet metal, the geometric center coordinates of the module mounting bracket, the fracture length of the support buckle in the groove, etc.) and the corresponding initial standard characteristic parameters of the structural component (such as the initial contour dimension reference value, the initial width reference value of the gap between multiple layers of sheet metal, the initial module center coordinate reference value of the module mounting bracket, the initial complete length reference value of the support buckle in the groove, etc.). It is used to quantify the severity of the defect. The initial standard characteristic parameters of the structural component are set in advance by preset personnel.

[0061] As described above, the identification of assembly defects in AI server chassis structural components helps to automatically and accurately locate abnormal areas on the chassis structural components and distinguish the types of defects, thereby improving the accuracy, consistency, and intelligence of assembly defect detection and achieving efficient, stable, and reliable automated detection and judgment of the assembly quality of AI server chassis structural components.

[0062] Example 2, based on Example 1, addresses the issue of AI server chassis structural components being produced in customized small batches with slight dimensional fluctuations between batches. Furthermore, some core features are located in extremely concealed areas (such as the inner wall features of deep-hole module mounting holes). Due to these batch size fluctuations, the standard features in the initial feature set cannot be adapted to all batches of structural component features. Additionally, the grayscale gradient of features in extremely concealed areas is weak, edges are blurred, and position coordinates are easily affected by noise, potentially leading to misjudgments of qualified structural component features. This can affect the accuracy of subsequent defect detection. Therefore, a second alternative scheme for structural component feature extraction qualification judgment needs to be implemented. The specific process is as follows: Obtain the structural component feature extraction coverage, structural component feature edge clarity, and structural component feature position corresponding to each core feature in the AI ​​server chassis structural component feature set. Accuracy; Structural component feature extraction coverage is represented by the ratio of the actual number of extracted pixels for a single core feature to the initial number of pixels for that single core feature. The closer the structural component feature extraction coverage is to 1, the more complete the feature extraction is. Structural component feature edge clarity is represented by the mean gray-level gradient of the edge pixels of a single feature. The higher the mean gray-level gradient, the clearer and more distinguishable the feature edge is. Edge pixels represent the pixels on the outline of the core feature of the AI ​​server chassis structural component obtained based on the edge detection algorithm, used to define the boundary range of the feature. Structural component feature position accuracy is obtained by performing Euclidean distance calculation between the geometric center coordinates of a single feature monitored by an industrial vision coordinate measuring instrument and the initial geometric center coordinates of the feature, and then taking the reciprocal of the result of the Euclidean distance calculation.

[0063] The structural component feature extraction coverage, edge sharpness, and location accuracy, calculated from individual core features, are input into a preliminary feature quality fusion evaluation model (such as a BP neural network model or a support vector machine model) for structural component parameter fusion. This yields a quantitative index for structural component feature extraction used to quantify the quality of individual core feature extraction. The specific training process of the preliminary feature quality fusion evaluation model is as follows: First, structural component feature samples from different scenarios and quality levels over historical time periods are acquired. The structural component feature extraction coverage, edge sharpness, and location accuracy are labeled, and corresponding training datasets are constructed. Then, the structural component feature extraction coverage, edge sharpness, and location accuracy are used as model inputs, and manually labeled feature quality scores are used as output targets. The model is iteratively trained, and the weights are continuously optimized. Finally, the evaluation accuracy of the model is tested using a validation set, and the model parameters are continuously adjusted until a preset stable evaluation effect is achieved, completing the model training.

[0064] The specific process of structural component parameter fusion is as follows: First, the input structural component feature extraction coverage, structural component feature edge clarity, and structural component feature position accuracy are standardized to unify the numerical range and dimensions of each parameter. Second, the standardized parameters are input into the initial feature quality fusion evaluation model, and the model performs weighted calculations and feature fusion on the three parameters. Finally, a single quantitative result after comprehensive weighted calculation is output, which is the structural component feature extraction quantitative index. It is determined whether the structural component feature extraction quantitative index is greater than the initial feature extraction benchmark value. If it is, the corresponding single core feature is marked as a qualified structural component feature; otherwise, the corresponding single core feature is marked as an unqualified structural component feature. The initial feature extraction benchmark value is represented by the average value of the structural component feature extraction quantitative index over a historical time period. Based on the obtained qualified and unqualified structural component features, the qualification of structural component image feature extraction is judged.

[0065] As described above, the qualification judgment of structural component feature extraction helps to adapt to the inspection needs of AI server chassis structural components in customized small-batch production. It effectively alleviates the misjudgment problem caused by small size fluctuations between batches and blurred features in extremely hidden areas, reduces the probability of qualified features being incorrectly judged due to abnormal feature extraction, blurred edges, and positional deviations, improves the stability, accuracy and scenario adaptability of feature quality assessment, provides reliable qualified feature data for subsequent defect identification, and realizes intelligent and refined judgment of the quality of structural component feature extraction.

[0066] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the scope and intent of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and variations.

Claims

1. A method for detecting assembly defects in AI server chassis structural components based on image recognition, characterized in that, Includes the following steps: During the assembly defect detection of the internal structural components of the AI ​​server chassis, the grayscale characteristics of the AI ​​server chassis structural component images are analyzed, and the grayscale adaptive correction of the chassis structural component images is used as a basis for decision-making based on the obtained analysis results. After the grayscale characteristic analysis is completed, the AI ​​server chassis structural component image feature extraction is performed to extract the core features of the internal structural components of the AI ​​server chassis. After the AI ​​server chassis structural component image feature extraction is completed, the structural component feature extraction qualification judgment is performed to quantify the qualification degree of feature extraction. After the structural component feature extraction and qualification judgment are completed, the assembly defects of the AI ​​server chassis structural components are identified, and the assembly defects of the AI ​​server chassis structural components are detected based on the obtained identification results to verify the accuracy of the identification results.

2. The method for detecting assembly defects in AI server chassis structural components based on image recognition as described in claim 1, characterized in that, The specific process for determining grayscale characteristics is as follows: The original images of the AI ​​server chassis structural components were converted to grayscale to obtain grayscale images of the AI ​​server chassis structural components. The grayscale conversion process refers to converting the original color AI server chassis structural component image into a single-channel grayscale image based on a weighted average grayscale conversion algorithm; Obtain the global grayscale mean and standard deviation of the grayscale components of the AI ​​server chassis structural components from grayscale images; The specific process for obtaining the global grayscale mean of the structural component is as follows: Summing the grayscale values ​​of all pixels in the grayscale image of the AI ​​server chassis structural components yields the summed grayscale value of the structural component image. The sum of grayscale values ​​of the structural component image is compared with the total number of pixels to obtain the global average grayscale value of the structural component. Based on the global grayscale mean and grayscale standard deviation of the structural components, compare the initial grayscale mean range with the initial grayscale standard deviation benchmark value to determine whether adaptive grayscale correction processing of the chassis structural components is required. If the global grayscale mean of the structural component is within the initially set grayscale mean range, and the grayscale standard deviation of the structural component is less than the initially set grayscale standard deviation benchmark value, then the corresponding grayscale image of the AI ​​server chassis structural component is marked as a qualified chassis structural component grayscale image, and AI server chassis structural component image feature extraction is performed based on the qualified chassis structural component grayscale image; otherwise, the corresponding AI server chassis structural component grayscale image is marked as an unqualified chassis structural component grayscale image, and chassis structural component image grayscale adaptive correction is performed based on the unqualified chassis structural component grayscale image.

3. The method for detecting assembly defects in AI server chassis structural components based on image recognition as described in claim 2, characterized in that, The specific process of adaptive grayscale correction for the chassis structural components is as follows: The grayscale image of the unqualified chassis structural component is initially corrected by the adaptive histogram equalization algorithm to obtain the grayscale image of the chassis structural component after preliminary correction. For the grayscale image of the chassis structural components after preliminary correction, the noise area is initially located. The specific process is as follows: Calculate the grayscale gradient value of each pixel in the grayscale image of the chassis structural components after preliminary correction; Based on grayscale gradient values, noise and feature regions are distinguished: if the grayscale gradient value of the corresponding pixel in the grayscale image of the chassis structure is greater than the initial gradient reference value, the region where the corresponding pixel is located is initially determined to be the noise interference region of the structure image; otherwise, the region where the corresponding pixel is located is initially determined to be the real feature region of the chassis structure image. Based on the noise interference region of the structural component image and the real feature region of the chassis structural component image, grayscale similarity and texture similarity are obtained.

4. The method for detecting assembly defects in AI server chassis structural components based on image recognition as described in claim 3, characterized in that, The specific process for obtaining grayscale similarity and texture similarity is as follows: Based on the connected component analysis algorithm, the noise interference region of the structural component image is clustered and merged to obtain the overall noise interference region of the structural component image. At the same time, based on the connected component analysis algorithm, the true feature region of the chassis structural component image is clustered and merged to obtain the overall true feature region of the chassis structural component image. The grayscale similarity is used to quantify the difference in grayscale distribution between the two types of regions, and the specific acquisition process is as follows: Extract the gray values ​​of the noise interference region and the true feature region respectively, and construct the gray-level pixel matrix corresponding to the two types of regions based on the extracted gray values. Based on the Gaussian filtering algorithm, the two constructed gray-level pixel matrices are preprocessed by Gaussian filtering to remove isolated extreme gray-level noise points in the two types of regions. While removing noise, the gray-level distribution characteristics of the region itself are preserved to the maximum extent. The two gray-level pixel matrix values, which have been preprocessed by Gaussian filtering, are input into a normalized cross-correlation algorithm to obtain gray-level similarity values. The specific process for obtaining the texture similarity is as follows: The gray-level co-occurrence matrix (GLCM) algorithm was used to obtain the GLCM of the overall noise interference region of the structural component image and the overall true feature region of the chassis structural component image, respectively. Standardize the two gray-level co-occurrence matrices respectively; Based on the weighted summation method, the gray-level co-occurrence matrices of the two types of regions after standardization are fused to obtain the texture similarity value; The obtained grayscale similarity and texture similarity are compared with the corresponding similarity judgment conditions to perform structural component image similarity comparison; The structural component image similarity comparison means comparing grayscale similarity and texture similarity with similarity judgment conditions. If grayscale similarity and texture similarity do not meet the similarity judgment conditions, the noise interference intensity and real feature clarity are calculated based on the noise interference area of ​​the structural component image and the real feature area of ​​the chassis structural component image. Otherwise, it indicates that the image area positioning of the noise area is confused and a noise area positioning failure prompt is sent. The similarity determination criteria indicate that the grayscale similarity is greater than the initial grayscale similarity benchmark value, and the texture similarity is greater than the initial texture similarity benchmark value.

5. The method for detecting assembly defects in AI server chassis structural components based on image recognition as described in claim 4, characterized in that, The specific process for calculating the noise interference intensity and the sharpness of the true feature is as follows: Calculate the noise interference intensity of the overall noise interference region of the structural component image and the clarity of the true features of the overall true feature region, respectively. The noise interference intensity is obtained based on the standard deviation and extreme difference of gray levels of the overall noise interference region in the structural component image. The specific calculation process is as follows: Based on the pixel extraction algorithm, the gray values ​​of all pixels in the noise interference area of ​​the structural component image are extracted, and the standard deviation of the gray values ​​of all pixels in the area is obtained as the gray standard deviation. The difference between the maximum and minimum gray values ​​of all pixels within the region is taken as the gray-level extreme value difference. The standard deviation of gray level and the difference of gray level extreme values ​​are normalized, and the weighted average of the two is taken as the noise interference intensity. The sharpness of the true features is obtained based on the mean gray-level gradient and edge contrast of the overall region of the true feature image of the chassis structural components. The specific calculation process is as follows: First, the gray-level gradient values ​​of all pixels in the real feature area of ​​the chassis structural component image are extracted based on the pixel extraction algorithm. The average value of all gray-level gradient values ​​is calculated. Then, the average gray-level difference between the core feature edge pixels and the surrounding pixels in the region is calculated. After normalizing the average gray-level gradient value and the edge contrast, the weighted average of the two is taken as the real feature sharpness. The decision on whether to adopt structured image denoising and enhancement co-processing is based on the intensity of noise interference and the clarity of true features.

6. The method for detecting assembly defects in AI server chassis structural components based on image recognition as described in claim 5, characterized in that, The structured image denoising and enhancement collaborative processing means that if the detected noise interference intensity is greater than the initial noise intensity benchmark value, the interference region denoising processing is performed on the overall noise interference area of ​​the structural component image. If the detected real feature clarity is less than the initial feature clarity benchmark value, perform real region enhancement processing on the overall real feature area of ​​the chassis structural component image. If the noise interference intensity is not greater than the initial noise intensity benchmark value and the real feature clarity is not less than the initial feature clarity benchmark value, then the corresponding pre-corrected chassis structure grayscale image is marked as a qualified chassis structure grayscale image, and AI server chassis structure image feature extraction is performed based on the qualified chassis structure grayscale image. After the structured image denoising and enhancement collaborative processing is completed, the noise interference intensity and the clarity of the real features are reacquired. If the noise interference intensity is still greater than the initial noise intensity benchmark value, or the clarity of the real features is still less than the initial feature clarity benchmark value, an image collaborative processing failure prompt is sent. Otherwise, the corresponding pre-corrected chassis structure grayscale image is marked as a qualified chassis structure grayscale image, and AI server chassis structure image feature extraction is performed based on the qualified chassis structure grayscale image. The noise reduction processing of the interference region refers to the suppression of noise in the interference region based on the bilateral filtering algorithm. After the processing is completed, the noise-reduced image of the noise interference region is output. The real region enhancement processing refers to performing a second-order difference operation on the gray values ​​of the real feature regions of the chassis structural component image based on the Laplacian operator enhancement algorithm, and outputting the enhanced real feature region image after processing.

7. The method for detecting assembly defects in AI server chassis structural components based on image recognition as described in claim 6, characterized in that, The AI ​​server chassis structural component image feature extraction means extracting features of various structural components inside the AI ​​server chassis based on feature extraction algorithms to obtain an AI server chassis structural component feature set; The following is a process for extracting structural component features and determining their suitability based on the feature set of AI server chassis structural components: Input the single core features from the AI ​​server chassis structural component feature set and the corresponding standard features from the initial feature set into the normalized cross-correlation algorithm, and use the output matching degree value as the structural component feature matching degree used to characterize the accuracy of feature extraction; The qualification of single core feature extraction is determined based on the feature matching degree of structural components; The qualification determination of single core feature extraction means judging whether the structural component feature matching degree of the single core feature is greater than the initial matching degree benchmark value. If it is, the corresponding single core feature is marked as a qualified structural component feature to be extracted; otherwise, the corresponding single core feature is marked as an unqualified structural component feature to be extracted. The specific process for extracting image features and determining the passability of structural components is as follows: The structural component feature extraction pass rate is obtained based on the number of qualified structural component features extracted. The pass rate of structural component feature extraction is compared with the initial overall pass threshold. If the pass rate of structural component feature extraction is greater than the initial overall pass threshold, the corresponding AI server chassis structural component feature set is marked as a qualified AI server chassis structural component feature set. Based on the qualified AI server chassis structural component feature set, assembly defects of AI server chassis structural components are identified. If the pass rate of structural component feature extraction is not greater than the initial overall pass threshold, then record the type of unqualified structural component features, the structural component feature matching degree and the batch information of the structural component to which it belongs, and send it to the defect detection and control center, and at the same time send a structural component feature extraction failure prompt.

8. The method for detecting assembly defects in AI server chassis structural components based on image recognition as described in claim 7, characterized in that, The AI ​​server chassis structural component assembly defect identification means inputting the feature set of qualified AI server chassis structural components into the initial defect identification model and outputting the structural component assembly defect identification result. The structural component assembly defect identification results include the undetermined area of ​​structural component assembly defects of the AI ​​server chassis, the type of chassis structural component defect, and the confidence level of the chassis structural component defect. Detection of assembly defects in AI server chassis structural components based on the results of structural component assembly defect identification.

9. The method for detecting assembly defects in AI server chassis structural components based on image recognition as described in claim 8, characterized in that, The specific process for detecting assembly defects in the structural components of the AI ​​server chassis is as follows: The confidence level of the defects in the chassis structural components is compared with the initial design confidence level benchmark value. If the confidence level of the chassis structural component defect corresponding to the undetermined area of ​​chassis structural component assembly defects is less than the initial defect confidence level benchmark value, the corresponding undetermined area of ​​chassis structural component assembly defects will be marked as a qualified area of ​​structural component assembly. If the confidence level of a defect in a chassis structural component is not less than the initial defect confidence level benchmark value, the undetermined area of ​​the chassis structural component assembly defect is determined as the structural component assembly defect area, and the defect details are recorded and sent to the defect detection and control center.

10. The method for detecting assembly defects in AI server chassis structural components based on image recognition as described in claim 7, characterized in that, The qualification judgment of structural component feature extraction also includes: Obtain the structural component feature extraction coverage, structural component feature edge clarity, and structural component feature position accuracy corresponding to each core feature in the AI ​​server chassis structural component feature set; The structural component feature extraction coverage rate is represented by the ratio of the actual number of extracted pixels for a single core feature to the initial number of pixels for that single core feature. The sharpness of the feature edges of the structural component is represented by the average gray-level gradient of the individual feature edge pixels. The edge pixels refer to the pixels on the core feature contour of the AI ​​server chassis structural components obtained based on the edge detection algorithm; The accuracy of the structural component feature position is obtained by calculating the Euclidean distance between the geometric center coordinates of an individual feature and the initial geometric center coordinates of that feature. The structural component feature extraction coverage, structural component feature edge clarity, and structural component feature position accuracy calculated from individual core features are input into the initial feature quality fusion evaluation model to perform structural component parameter fusion, resulting in a structural component feature extraction quantitative index used to quantify the quality of individual core feature extraction. Determine whether the quantitative index for structural component feature extraction is greater than the initial feature extraction benchmark value. If it is, mark the corresponding single core feature as a qualified structural component feature; otherwise, mark the corresponding single core feature as an unqualified structural component feature. Based on the extracted features of qualified and unqualified structural components, the qualification of structural components is determined by extracting image features.

Images