A method and system for intelligent quality inspection of second-hand mobile phones

Abstract

The application relates to the technical field of image recognition, and discloses a used mobile phone intelligent quality inspection method and system, which comprises the following steps: acquiring a target image sequence to be detected; extracting a physical damage label and a display area state label by using a model; recognizing a key image frame according to the display area state; extracting a global edge feature and an internal semantic anchor point of the display area; using the semantic anchor point as a topological support node, combining the edge feature to construct a topological local grid, resetting pixel sampling according to a mapping relationship, and generating a standardized image; identifying hardware attribute parameters through a multi-modal model, generating a check result by comparing nominal parameters; and generating a loss grade report through comprehensive quantitative calculation. The application couples the semantic anchor point and the local grid, compensates for image distortion caused by physical deformation, and enhances the objectivity of quality inspection by using multi-modal semantic understanding.

Description

Technical Field

[0001] This invention relates to a method and system for intelligent quality inspection of used mobile phones, belonging to the field of image recognition technology. Background Technology

[0002] Currently, image processing technology is commonly used in the industry to identify appearance damage and extract hardware configuration parameters. This method usually acquires image sequences of the device under test, uses neural network models to identify target features, and initially completes automated classification and information verification. However, in non-standardized acquisition environments, due to uneven light intensity, complex shooting backgrounds, and randomness in acquisition angles, the imaging quality of the target object is limited. For devices using micro-curved glass technology, or devices whose internal battery expansion causes physical deformation, the screen produces non-uniform geometric distortion in the optical imaging space.

[0003] While hardware improvements such as optimizing mechanical limits or adding high-precision sensors can alleviate imaging stability issues, software control algorithmic logic flaws limit breakthroughs in quality inspection accuracy. Chinese invention patent CN117237353B discloses a method, device, equipment, and storage medium for detecting defects in mobile phone appearance. This technical solution relies on 3D scanning and infrared thermal imaging to construct a finite element model for defect analysis. The technology is based on the assumption of a rigid structure, using physical modeling to invert material defects. However, in the context of used mobile phone recycling, the equipment faces nonlinear conditions such as 2.5D edge curvature or irregular shell deformation. Existing technologies lack pixel-level correction mechanisms and cannot eliminate semantic breaks caused by imaging distortion at the underlying topological mapping level. This leads to accumulated errors in character recognition and parameter verification, affecting the objectivity of classification in non-cooperative acquisition scenarios. Existing global perspective transformation algorithms are based on the assumption of an ideal rigid plane. When processing images with non-ideal topological features, they easily cause nonlinear stretching of edge pixels, disrupting the continuity of the character space structure. Simply increasing model depth or expanding the dataset size may improve feature recognition probability, but it cannot correct semantic breaks caused by imaging distortion at the underlying physical mapping level.

[0004] Therefore, how to eliminate nonlinear geometric distortion in imaging under non-cooperative acquisition scenarios and achieve topologically consistent configuration parameter extraction has become the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A method for intelligent quality inspection of used mobile phones, comprising the following steps: Step S101: Obtain a multi-angle image sequence of the target to be detected, denoted as S; Step S102: Input the image sequence S into the pre-trained target detection model, extract and output the set of physical damage feature labels and the set of physical state labels of the display area corresponding to each image; Step S103: Based on the set of physical state labels of the display area, identify key image frames containing attribute information of the display interface from the image sequence S, and perform an enhanced feature extraction process: perform orientation correction based on the image classification results of the key image frames, extract global edge features of the display area of ​​the target to be detected, and extract internal semantic anchors from the feature space of the key image frames; use the internal semantic anchors as topological support nodes, combine global edge features to construct a topology-preserving local mesh, and perform pixel resampling according to the node mapping relationship of the topology-preserving local mesh to generate a standardized feature image; Step S104: Input the standardized feature image into the multimodal large language model, identify and extract the hardware attribute feature parameters of the target to be detected, and compare the hardware attribute feature parameters with the pre-stored nominal parameters to generate a consistency verification result. Step S105: Based on the physical damage feature label set, hardware attribute feature parameters, and consistency verification results, a loss level characterization report of the target to be detected is generated through weighted quantization calculation.

[0006] Preferably, before executing the enhanced feature extraction process in step S103, the following steps are also included: calculating the motion blur score and the proportion of the highlight occlusion area of ​​the key image frame; when the motion blur score exceeds a preset blur threshold, or the proportion of the highlight occlusion area exceeds a preset area threshold, step S101 is triggered to perform the image re-acquisition action; wherein, the motion blur score is determined based on the response variance of the key image frame under the Laplacian operator.

[0007] Preferably, the specific method for constructing the topology-preserving local mesh in step S103 is as follows: within the key image frame, the effective display area is determined by semantic segmentation, and interface controls with fixed positions are located within the effective display area as main semantic anchors, and the center point of the text line spacing is located as auxiliary semantic anchors; a set of control points is established based on the main semantic anchors and auxiliary semantic anchors, and the nonlinear mapping rule between the set of control points and the actual physical coordinate system is fitted using the thin plate smoothing spline function to generate a non-uniform topology mesh covering the entire target to be detected.

[0008] Preferably, the set of physical damage feature labels extracted in step S102 includes labels that identify the display area of ​​the target to be detected as broken, the frame as bumped, the body as scratched, the camera module as aged, and the display area as leaking. Each physical damage feature label in the set of physical damage feature labels is associated with corresponding location coordinate information and damaged area data, which serve as quantitative input parameters for the loss level characterization report in step S105.

[0009] Preferably, the control logic for performing pixel reset sampling in step S103 while preserving the topology of the local mesh follows the following rule: utilizing the projection transformation matrix of each local mesh cell. Correcting local distortion, projection transformation matrix The following relationship must be satisfied: ,in, The original feature coordinates of the j-th anchor point within the i-th local mesh cell are given. To correspond to the target feature coordinates in the standardized coordinate system, To maintain the continuity of adjacent grid cells, the topological constraint term is k, which is the total number of internal semantic anchors contained in the local grid cell, and λ is the preset smoothing regularization coefficient.

[0010] Preferably, the specific method for generating the consistency verification result in step S104 is as follows: extract the physical identification code of the target to be detected and the model of the core computing unit as the core identification item; in the process of comparing the hardware attribute feature parameters with the nominal parameters, if the matching degree of the core identification item is less than 100%, generate a low-level hardware change warning label and increase the anti-counterfeiting feature detection strength of the standardized feature image.

[0011] Preferably, the extraction logic of the auxiliary semantic anchor point is as follows: under the standardized coordinate system, the pixel projection density analysis in the vertical direction is performed on the image after preliminary correction, and the peak region in the pixel density distribution is identified as the candidate text line position; the vertical reference of the auxiliary semantic anchor point is determined by calculating the center distance between adjacent peak regions, so as to compensate for the local stretching of the image caused by the change in the physical properties of the target to be detected.

[0012] Preferably, in step S104, the multimodal large language model uses a prompt-guided mechanism to extract hardware attribute feature parameters, including: inputting first modal data containing standardized feature images and second modal instructions containing quality inspection protocol constraints into the multimodal large language model; the second modal instructions are used to constrain the multimodal large language model to output a structured data stream, which includes memory capacity, storage space, energy component attributes, and network protocol parameters. In step S105, the generation logic of the loss level characterization report is as follows: determining the hardware baseline parameter value based on the hardware attribute feature parameters, determining the compliance correction weight based on the consistency verification result, and performing weight superposition calculation by combining the damage quantification value corresponding to the physical damage feature label set; wherein, the display area leakage label and the underlying hardware change warning label have a veto right in the weight superposition calculation, that is, once identified, the rating result of the loss level characterization report is directly determined to be at the scrap level.

[0013] Preferably, the method further includes the following steps: when the physical damage feature label set contains display area scratch labels, by extracting the same feature points in the overlapping field of view and calculating the spatial transformation matrix, performing feature alignment between the macro image and the macro image, the physical depth estimate of the display area scratch is obtained; if the physical depth estimate exceeds a preset depth threshold, a physical performance degradation warning signal for the display area is generated in the loss level characterization report based on the mapping relationship between the physical depth estimate and the physical state of the display area.

[0014] A second-hand mobile phone intelligent quality inspection system includes: The image acquisition module is used to acquire a multi-angle image sequence S of the target to be detected; The detection and extraction module is used to input the image sequence S into the pre-trained target detection model, extract and output the set of physical damage feature labels and the set of physical state labels of the display area corresponding to each image; The image reconstruction module is used to identify key image frames containing attribute information of the display interface from the image sequence S based on the set of physical state labels of the display area, and to perform an enhanced feature extraction process: performing orientation correction based on the image classification results of the key image frames, extracting global edge features of the display area of ​​the target to be detected and internal semantic anchors in the feature space of the key image frames, using the internal semantic anchors as topological support nodes, constructing a topology-preserving local mesh in combination with global edge features, and performing pixel resampling according to the node mapping relationship of the topology-preserving local mesh to generate a standardized feature image; The identification and verification module is used to input standardized feature images into a multimodal large language model, identify and extract hardware attribute feature parameters of the target to be detected, and compare the hardware attribute feature parameters with the pre-stored nominal parameters to generate a consistency verification result. The report generation module is used to generate a loss level characterization report of the target under test by integrating the set of physical damage feature labels, hardware attribute feature parameters and consistency verification results through weighted quantization calculation.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. In the intelligent quality inspection of used mobile phones, the cross-mechanism linkage between the target detection model and the enhanced information extraction process solves the problem of topological deformation in mobile phone screen imaging under non-standard shooting environments. Traditional image correction schemes are based on the assumption of an ideal two-dimensional rigid plane, which is difficult to cope with the non-uniform optical distortion caused by the 2.5D micro-curved glass of modern mobile phones or battery aging and bulging. This invention uses a semantic segmentation model to extract the screen area mask and combines a multi-strategy corner detection algorithm to perform multi-level iterative fitting of the screen contour, realizing the geometric benchmark reconstruction of non-ideal curved surface imaging in the feature space. This processing mechanism avoids the nonlinear stretching of edge pixels caused by traditional global perspective transformation, ensuring that the generated standardized information page image maintains the semantic coherence of the text structure, providing high-fidelity underlying data for the accurate recognition of subsequent multimodal large language models, and eliminating parameter recognition false alarms caused by image geometric distortion.

[0016] 2. Construct a closed-loop control chain based on image classification, quality assessment, and geometric transformation to improve the engineering adaptability of the vision processing system to extreme working conditions. Utilize a pre-trained image classification model to identify the rotation angle of a specific image relative to the standard reading perspective. Before performing perspective transformation matrix calculation, simultaneously introduce maximum tilt angle detection and a blur scoring mechanism based on Laplacian variance. This multi-dimensional pre-quality verification logic can automatically identify and block pathological data with severe perspective distortion, reflective occlusion, or motion blur, preventing low signal-to-noise ratio images from entering the character recognition stage and causing error accumulation. Through dynamic real-time feedback on image acquisition quality, the system effectively reduces the ineffective occupation of downstream computing resources while ensuring the reliability of output results, thereby improving the stability of the overall processing flow.

[0017] 3. An integrated architecture for fine-grained identification of appearance defects and verification of configuration information is achieved. Through the output of a collaborative neural network model, the automation breadth of used mobile phone quality inspection is improved. The target detection model simultaneously extracts the set of appearance defect labels and the set of screen status labels, establishing the association and triggering relationship between physical damage features and system software information pages. This technical solution integrates pixel-level tasks such as micro-scratch detection and screen leakage identification with structured tasks such as information page semantic segmentation and character extraction into a unified visual processing workflow. This multi-task parallel processing mode enables the system to complete the equipment condition grading and nominal parameter consistency verification simultaneously through a single image sequence acquisition, eliminating the information gap between manual input and visual inspection in the traditional quality inspection process, and improving the operational efficiency and grading objectivity in industrial recycling scenarios. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall implementation process of the intelligent quality inspection method for used mobile phones of the present invention; Figure 2 This is a logic diagram of the image enhancement processing of semantic anchor point compensation and local mesh reconstruction in this invention.

[0019] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0021] A smart quality inspection method for used mobile phones includes the following steps: Step S101: Obtain a multi-angle image sequence of the target to be detected, denoted as S; Step S102: Input the image sequence S into the pre-trained target detection model, extract and output the set of physical damage feature labels and the set of physical state labels of the display area corresponding to each image; Step S103: Based on the set of physical state labels of the display area, identify key image frames containing attribute information of the display interface from the image sequence S, and perform an enhanced feature extraction process: perform orientation correction based on the image classification results of the key image frames, extract global edge features of the display area of ​​the target to be detected, and extract internal semantic anchors from the feature space of the key image frames; use the internal semantic anchors as topological support nodes, combine global edge features to construct a topology-preserving local mesh, and perform pixel resampling according to the node mapping relationship of the topology-preserving local mesh to generate a standardized feature image; Step S104: Input the standardized feature image into the multimodal large language model, identify and extract the hardware attribute feature parameters of the target to be detected, and compare the hardware attribute feature parameters with the pre-stored nominal parameters to generate a consistency verification result. Step S105: Based on the physical damage feature label set, hardware attribute feature parameters, and consistency verification results, a loss level characterization report of the target to be detected is generated through weighted quantization calculation.

[0022] Preferably, before executing the enhanced feature extraction process in step S103, the following steps are also included: calculating the motion blur score and the proportion of the highlight occlusion area of ​​the key image frame; when the motion blur score exceeds a preset blur threshold, or the proportion of the highlight occlusion area exceeds a preset area threshold, step S101 is triggered to perform the image re-acquisition action; wherein, the motion blur score is determined based on the response variance of the key image frame under the Laplacian operator.

[0023] Preferably, the specific method for constructing the topology-preserving local mesh in step S103 is as follows: within the key image frame, the effective display area is determined by semantic segmentation, and interface controls with fixed positions are located within the effective display area as main semantic anchors, and the center point of the text line spacing is located as auxiliary semantic anchors; a set of control points is established based on the main semantic anchors and auxiliary semantic anchors, and the nonlinear mapping rule between the set of control points and the actual physical coordinate system is fitted using the thin plate smoothing spline function to generate a non-uniform topology mesh covering the entire target to be detected.

[0024] Preferably, the set of physical damage feature labels extracted in step S102 includes labels that identify the display area of ​​the target to be detected as broken, the frame as bumped, the body as scratched, the camera module as aged, and the display area as leaking. Each physical damage feature label in the set of physical damage feature labels is associated with corresponding location coordinate information and damaged area data, which serve as quantitative input parameters for the loss level characterization report in step S105.

[0025] Preferably, the control logic for performing pixel reset sampling in step S103 while preserving the topology of the local mesh follows the following rule: utilizing the projection transformation matrix of each local mesh cell. Correcting local distortion, projection transformation matrix The following relationship must be satisfied: ,in, The original feature coordinates of the j-th anchor point within the i-th local mesh cell are given. To correspond to the target feature coordinates in the standardized coordinate system, To maintain the continuity of adjacent grid cells, the topological constraint term is k, which is the total number of internal semantic anchors contained in the local grid cell, and λ is the preset smoothing regularization coefficient.

[0026] Preferably, the specific method for generating the consistency verification result in step S104 is as follows: extract the physical identification code of the target to be detected and the model of the core computing unit as the core identification item; in the process of comparing the hardware attribute feature parameters with the nominal parameters, if the matching degree of the core identification item is less than 100%, generate a low-level hardware change warning label and increase the anti-counterfeiting feature detection strength of the standardized feature image.

[0027] Preferably, the extraction logic of the auxiliary semantic anchor point is as follows: under the standardized coordinate system, the pixel projection density analysis in the vertical direction is performed on the image after preliminary correction, and the peak region in the pixel density distribution is identified as the candidate text line position; the vertical reference of the auxiliary semantic anchor point is determined by calculating the center distance between adjacent peak regions, so as to compensate for the local stretching of the image caused by the change in the physical properties of the target to be detected.

[0028] Preferably, in step S104, the multimodal large language model uses a prompt-guided mechanism to extract hardware attribute feature parameters, including: inputting first modal data containing standardized feature images and second modal instructions containing quality inspection protocol constraints into the multimodal large language model; the second modal instructions are used to constrain the multimodal large language model to output a structured data stream, which includes memory capacity, storage space, energy component attributes, and network protocol parameters. In step S105, the generation logic of the loss level characterization report is as follows: determining the hardware baseline parameter value based on the hardware attribute feature parameters, determining the compliance correction weight based on the consistency verification result, and performing weight superposition calculation by combining the damage quantification value corresponding to the physical damage feature label set; wherein, the display area leakage label and the underlying hardware change warning label have a veto right in the weight superposition calculation, that is, once identified, the rating result of the loss level characterization report is directly determined to be at the scrap level.

[0029] Preferably, the method further includes the following steps: when the physical damage feature label set contains display area scratch labels, by extracting the same feature points in the overlapping field of view and calculating the spatial transformation matrix, performing feature alignment between the macro image and the macro image, the physical depth estimate of the display area scratch is obtained; if the physical depth estimate exceeds a preset depth threshold, a physical performance degradation warning signal for the display area is generated in the loss level characterization report based on the mapping relationship between the physical depth estimate and the physical state of the display area.

[0030] A second-hand mobile phone intelligent quality inspection system includes: The image acquisition module is used to acquire a multi-angle image sequence S of the target to be detected; The detection and extraction module is used to input the image sequence S into the pre-trained target detection model, extract and output the set of physical damage feature labels and the set of physical state labels of the display area corresponding to each image; The image reconstruction module is used to identify key image frames containing attribute information of the display interface from the image sequence S based on the set of physical state labels of the display area, and to perform an enhanced feature extraction process: performing orientation correction based on the image classification results of the key image frames, extracting global edge features of the display area of ​​the target to be detected and internal semantic anchors in the feature space of the key image frames, using the internal semantic anchors as topological support nodes, constructing a topology-preserving local mesh in combination with global edge features, and performing pixel resampling according to the node mapping relationship of the topology-preserving local mesh to generate a standardized feature image; The identification and verification module is used to input standardized feature images into a multimodal large language model, identify and extract hardware attribute feature parameters of the target to be detected, and compare the hardware attribute feature parameters with the pre-stored nominal parameters to generate a consistency verification result. The report generation module is used to generate a loss level characterization report of the target under test by integrating the set of physical damage feature labels, hardware attribute feature parameters and consistency verification results through weighted quantization calculation.

[0031] Example 1: For the quality inspection deployment scenario of a high-throughput used mobile phone recycling center, the system faces challenges due to the prevalence of 2.5D micro-curved glass screens on the targets to be inspected, and physical deformation caused by battery aging and bulging in some devices. These non-ideal topological features result in non-uniform geometric distortion of screen imaging, leading to impaired spatial structure continuity of characters on the "About Me" page. This application example uses an intelligent quality inspection method for used mobile phones to address these technical challenges. The image acquisition module acquires a multi-angle image sequence of the target to be inspected, denoted as S. The detection and extraction module inputs the image sequence S into a pre-trained target detection model, extracts and outputs a set of appearance defect labels and a set of screen state labels corresponding to each image. The image reconstruction module identifies key image frames containing attribute information from the image sequence S based on the screen state label set and triggers an enhanced feature extraction process. For this key image frame, the system uses image classification results to determine the rotation angle and reset the image orientation. Within the effective display area defined by the binarized mask, the morphological gradient operator is used to extract the horizontal list separator lines and text alignment baselines of the mobile phone information page. The intersections of these structured features are located as internal semantic anchors, and a hybrid anchor set is constructed by combining the screen edge contour points. The image reconstruction module uses a hybrid anchor point set. As topological support nodes, the Delaunay triangulation algorithm is used to construct an initial topological mesh in the feature space. This transforms a single global rigid transformation into multiple mesh-based local mappings; for each sub-triangle facet, the system uses the original feature coordinates of its vertices. Target feature coordinates in standard coordinate system Calculate the local projection transformation matrix that satisfies the constraints. The local projection transformation matrix The following relationship must be satisfied: ;in, Let k be the local projection transformation matrix, and k be the total number of internal semantic anchor points contained within the local mesh cell. For the target feature coordinates, Here, λ represents the original feature coordinates, and λ is a preset smoothing regularization coefficient. To maintain the continuity of adjacent mesh cells, a topology-preserving local mesh is constructed in step S103. To maintain the geometric continuity of adjacent mesh cells, the calculation method is to project the transformation matrix of adjacent mesh cells. Perform differential operations and calculate the sum of Frobenius norms, where the deviation between the subscript i and the adjacent index is limited to a unit grid step. Adjust the deformation intensity by a preset smoothing regularization coefficient λ. When the standard deviation of the detected feature point location is greater than 0.5 pixels, λ is incremented in the range of 0.12 to 0.18. Filter out optical texture noise-induced pseudo-deformation by constraining the gradient of the grid node displacement vector.

[0032] Through this local reset sampling logic, the system generates a standardized feature image that removes interference from surface distortion and physical deformation. The serial number and model characters, originally distorted due to screen curvature, are restored to a linear arrangement in this standardized feature image. The identification and verification module inputs the standardized feature image into a multimodal large language model to identify and extract the hardware attribute feature parameters of the target to be detected. These parameters are then compared with pre-stored nominal parameters to generate a consistency verification result. In this embodiment, because image distortion can be offset at the pixel level, the multimodal large language model maintains a stable recognition rate for key fields, eliminating false alarms in parameter extraction caused by edge stretching. The report generation module executes a specific weighted quantization algorithm: setting a base score of 100 points, the identified cosmetic damage and scratches are totaled... The area (in 1.0 square millimeter increments) multiplied by a decay coefficient of 0.75 is deducted from the total score; simultaneously, the measured percentage deviation of memory capacity and battery health is multiplied by a weight coefficient of 0.6 for a second deduction; if the consistency verification result shows that the matching degree of core identification items is less than 95%, an additional 15 points of compliance weight penalty are added. The report generation module integrates the set of appearance defect labels, hardware attribute feature parameters, and consistency verification results to generate a report reflecting the true wear level of the target under test; this local mesh reconstruction mechanism based on UI semantic anchor points translates the optical distortion caused by the physical deformation of curved surfaces into a node offset problem in the feature space, and uses the stable topological structure inside the image to offset the geometric distortion in the non-cooperative acquisition environment.

[0033] Example 2: This example verifies the contribution of topology-preserving local mesh generation of standardized feature images to the accuracy of hardware attribute feature parameter recognition. It includes a detection box equipped with a diffuse LED light source, where the internal illumination intensity is maintained between 450 and 550 lux. The image acquisition terminal uses an industrial camera with a resolution of 4000 pixels by 3000 pixels. The experimental dataset includes 1200 images of used mobile phone screens covering 2.5D edge curvature and 1-3 mm of micro-deformation in the casing due to internal component aging. This dataset actively superimposes Gaussian white noise with a signal-to-noise ratio of 25 dB and high-brightness flickering disturbance with an illumination non-uniformity of 15% to simulate real sensor thermal noise and ambient light interference. The value of the smoothing regularization coefficient λ is controlled by the hybrid anchor point set. The localization variance, a parameter that balances image reconstruction smoothness with local deformation fitting accuracy, when the image sequence When the signal-to-noise ratio is below 30dB, in order to improve the initial topology mesh To ensure structural stability, the smoothing regularization coefficient λ tends towards the upper limit of its range; in the experimental group, this parameter was set to 0.15; where λ is the smoothing regularization coefficient. Let S be a set of mixed anchor points, and S be an image sequence. As the initial topological mesh, a control group was set up to process the original image using a global perspective transformation algorithm, while the experimental group used a local pixel resampling method based on internal semantic anchor point topological support. With the original input data captured at a 30-degree tilt angle, the corrected image generated by the control group exhibited pixel stretching at character edges, resulting in an 18.4% deviation in the horizontal and vertical aspect ratio of characters on the "About This Page" page. The experimental group constructed the initial topological mesh by extracting internal semantic anchor points. Under the above conditions, the local projection transformation matrix The average mapping error for local triangular patches remains at 0.23 pixels, and the topological constraint term... It remains at a numerical level of 0.052; among which, The local projection transformation matrix is... For topological constraints; the character recognition accuracy of the control group under the above conditions was 82.5%, and the false alarm rate of model parameters due to edge stretching was 12.3%.

[0034] The standardized feature images generated by the experimental group achieved a character recognition accuracy of 98.4% under a multimodal large language model, while reducing the false alarm rate of hardware attribute feature parameter extraction to 0.75%. Examining the gradient characteristics of performance changes with physical deformation intensity, when the deformation of the phone casing increased from 1mm to 5mm, the recognition accuracy of the control group showed a non-linear accelerating decline. After the deformation exceeded 3.5mm, the recognition accuracy dropped below 60% due to the global perspective model's inability to fit non-planar topological changes. In contrast, the experimental group maintained a stable character recognition accuracy above 95.5% when the deformation was between 1 and 4mm. After the deformation reached the performance inflection point of 4.2mm, the accuracy declined due to physical optical occlusion effects. These experimental results demonstrate that by deconstructing the global geometric transformation into a local mesh mapping based on semantic anchors, the non-linear pixel displacement caused by non-ideal topological features can be offset, thus providing a geometrically consistent data foundation for hardware attribute extraction. This method utilizes a local projection transformation matrix. The established local mapping mechanism demonstrates error compensation accuracy under shell deformation conditions within 4mm, transforming the image recognition risk caused by changes in physical properties into a coordinate transformation problem.

[0035] Example 3: In the automated quality inspection line of a used mobile phone recycling platform, the system processes localized highlight interference caused by fingerprint residue on the screen surface. This interference causes breakage of image edge features during binarization, disrupting the initial topological mesh. The structural integrity of the image causes pixel displacement during pixel resampling using traditional methods. This application example uses a method that compensates for the displacement deviation caused by the above interference by adding auxiliary semantic anchors. Before extracting the internal semantic anchors, the image reconstruction module determines that the resolution of the key image frame to be processed is 4000 pixels by 3000 pixels, and the image bit depth is 8 bits. Within the effective display area, the processor uses a rectangular structuring element with a size of 5 pixels by 5 pixels to calculate the morphological gradient. The calculation formula is as follows: I=(I⊕B)-(I B); among which, I represents the morphological gradient image, I' represents the original grayscale image, B represents the rectangular structuring element, and ⊕ represents the dilation operator. For the erosion operator; when the screen status label set returned by the detection and extraction module contains a local occlusion marker, the system initiates the auxiliary semantic anchor point extraction process. The processor, in a standardized coordinate system, statistically analyzes the vertical pixel projection density of the orientation-corrected image, calculates the cumulative grayscale value of each row of pixels, identifies local peak regions in the pixel density distribution curve as candidate text line positions, and determines the vertical reference of the auxiliary semantic anchor point by calculating the center distance between adjacent peak regions. The vertical reference The statistical average value of the interface text alignment baseline is used to maintain the topological shape of the local grid when the main semantic anchor point is lost due to specular highlights. The internal semantic anchor point extraction logic is based on the prior recognition of the operating system interface topology structure. The display area is divided into an initial retrieval array of N rows and M columns. The Sobel operator is used to extract the text edge gradient of each sub-region and Harris corner detection is performed. The k feature points with the highest response values ​​are selected as candidate anchor points. When the anchor point is missing due to reflection occlusion, the auxiliary anchor point compensation procedure is executed. The pixel grayscale of the preliminary correction image is accumulated in the vertical direction. The local maxima points in the projection curve that conform to the periodic distribution characteristics are identified as the reference line of the text line center. Combined with the preset interface control relative coordinate distribution map, the geometric node coordinates of the disturbed area are completed by linear interpolation to generate a global non-uniform topological grid covering the target to be detected.

[0036] The image reconstruction module will use auxiliary semantic anchors and hybrid anchor sets. Merge and calculate the local projection transformation matrix using the updated node coordinates. To maintain the smoothness of the mesh transformation, the system adjusts the smoothing regularization coefficient λ, and the processor calculates the smoothing regularization coefficient based on the variance of the pixel gradient within the local mesh cell. The rule for calculating this coefficient is defined as follows: If Greater than the preset gradient threshold If λ is reduced to enhance the mesh's fit to deformation, then... Less than or equal to Then λ is kept at 0.15; where λ is the smoothing regularization coefficient. This represents the variance of the pixel gradient. The preset gradient threshold is used. The recognition and verification module receives the standardized feature image output by the image reconstruction module. The processor inputs first modal data containing the standardized feature image and second modal instructions containing quality inspection protocol constraints into the multimodal large language model. The second modal instructions are used to constrain the multimodal large language model to output a structured data stream, which includes memory capacity, storage space, energy component attributes, and network protocol parameters. The model converts the visual feature vector into structured data based on the pixel layout features in the standardized feature image. The recognition and verification module inputs the second modal instructions into the multimodal large language model, using a fixed structure prompt word sequence, including hardware information extraction task definition, structured data stream output format requirements, and missing parameter filling rules, restricting the model to extract key-value pairs only for pixel layout features of the standardized feature image, in order to establish a low-level physical scale mapping benchmark. Before data acquisition, a black and white checkerboard calibration board with known physical dimensions is used to determine the equivalent focal length and distortion parameters of the camera. The gain coefficient of the industrial camera is adjusted according to the average gray value of the central area of ​​the calibration board to ensure that the image input has consistent brightness response characteristics under different lighting conditions. The report generation module receives the above structured data stream and determines the compliance correction weight based on the consistency verification results. It generates a loss level characterization report by combining the set of appearance defect labels. Due to the introduction of auxiliary semantic anchors and dynamic regularization mechanisms, the local mapping error of the system is reduced from 1.5 pixels to 0.28 pixels under the condition of high light occlusion. The geometric fidelity of the standardized feature image meets the constraint of multimodal semantic recognition on the aspect ratio of characters. This mesh reconstruction mechanism based on morphological gradient and auxiliary anchor compensation transforms visual noise into a constrained topological node adjustment problem and corrects semantic anchors through the physical pixel distribution law.

[0037] Example 4: In the initial deployment of the new quality inspection center, the inconsistent brightness of the acquired images caused by fluctuations in ambient light intensity necessitates a standardized calibration process to establish a physical benchmark for underlying visual perception. The processor controls the image acquisition module to acquire a reference image of a black and white checkerboard calibration board of a preset size, and extracts the pixel coordinates of the checkerboard corner points. And combined with physical length Calculate the camera's equivalent focal length and radial distortion coefficient; where, These are the pixel coordinates of the corner points of the chessboard. The physical length of the calibration board; when the average grayscale value of the central area of ​​the calibration board deviates from the target grayscale threshold of 128, the system adjusts the exposure time of the industrial camera. Until the brightness residual is less than 2%; among which, The exposure time of the camera is used; this pre-calibration process defines the original input brightness benchmark for subsequent feature extraction processes, so that the pixel grayscale distribution under different physical environments has statistical consistency across scenes.

[0038] When the quality inspection system encounters a situation where the local nominal parameter library version lags behind due to the addition of new mobile phone models to the market, the system executes a baseline parameter reconstruction procedure to ensure the compliance of the data source of the identification and verification module. The processor obtains the hardware attribute feature dictionary maintained in the cloud through an authorized interface, and uses the model identification code of the target to be tested as an index to retrieve the baseline parameter vector containing memory capacity, storage space, energy component attributes, and network protocol parameters. ;in, This serves as the baseline parameter vector for hardware attributes; the verification module calculates the hardware attribute feature parameters extracted from the multimodal large language model and compares them with this baseline parameter vector. If the Euclidean distance is greater than the preset logical judgment threshold and the model identification code matching degree reaches 100%, a low-level hardware change warning label is generated. This dynamic dictionary update mechanism based on cloud synchronization eliminates the identification island phenomenon caused by device hardware iteration.

[0039] Example 5: In the initial deployment of the automated quality inspection module, the processor establishes a scale mapping relationship between the feature space and the physical space by acquiring a standardized raster image containing a preset pixel spacing. To achieve an optimal balance between computational load and reset accuracy, the system defines the total number of patches. To optimize the variables, the total number of horizontal pixels in the display area of ​​the target to be detected is used. Total number of vertical pixels Calculate the initial topology mesh Reference value of unit side length The reference value of the side length of this unit The following relationship must be satisfied: ;in, The reference value for the element side length. This represents the total number of horizontal pixels. This represents the total number of vertical pixels. The total number of facets is selected from 256 to 1024. When the target to be detected is a micro-curved screen with a radius of curvature greater than 3mm, the system performs mesh refinement in the 10% area at the edge, so that the vertex distribution density of the local mesh is adjusted to 1.5 times that of the central area. This method adjusts the mesh fineness through physical pixel density, providing geometric accuracy support for edge pixel reset.

[0040] When the quality inspection system is in a multi-source acquisition environment, the processor determines the gradient threshold. An offline optimization procedure was executed, with minimizing the false anchor point detection rate under background noise interference as the performance metric. The illumination intensity of the light source array was adjusted in 10-lux steps within the range of 300 to 800 lux. Fifty test samples containing fingerprint interference were collected, and the global mean of their morphological gradient images was calculated. and standard deviation Therefore, the gradient threshold is set. Calibrated to meet The criteria for judgment; among which, For gradient threshold, The global mean. The standard deviation is used to trigger the downward adjustment of the smoothing regularization coefficient λ when the pixel gradient fluctuation exceeds the background noise. Meanwhile, in the recognition and verification module processing stage, the first modality data received by the multimodal large language model is normalized into a 512-dimensional local visual feature vector, which is mapped to the semantic feature space corresponding to the second modality instruction through a linear projection layer. The processor determines the consistency between the extraction result and the nominal parameters based on the cosine similarity. The above calibration procedure realizes the stable output of the system under non-consistent perturbation.

[0041] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0042] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for intelligent quality inspection of used mobile phones, characterized in that, Includes the following steps: Step S101: Obtain a multi-angle image sequence of the target to be detected, denoted as S; Step S102: Input the image sequence S into the pre-trained target detection model, extract and output the set of physical damage feature labels and the set of physical state labels of the display area corresponding to each image; Step S103: Based on the set of physical state labels of the display area, identify key image frames containing attribute information of the display interface from the image sequence S, and perform an enhanced feature extraction process: perform orientation correction based on the image classification results of the key image frames, extract global edge features of the display area of ​​the target to be detected, and extract internal semantic anchors from the feature space of the key image frames; use the internal semantic anchors as topological support nodes, combine global edge features to construct a topology-preserving local mesh, and perform pixel resampling according to the node mapping relationship of the topology-preserving local mesh to generate a standardized feature image; Step S104: Input the standardized feature image into the multimodal large language model, identify and extract the hardware attribute feature parameters of the target to be detected, and compare the hardware attribute feature parameters with the pre-stored nominal parameters to generate a consistency verification result. Step S105: Based on the physical damage feature label set, hardware attribute feature parameters, and consistency verification results, a loss level characterization report of the target to be detected is generated through weighted quantization calculation.

2. The intelligent quality inspection method for used mobile phones according to claim 1, characterized in that, Before executing the enhanced feature extraction process in step S103, the following steps are also included: calculating the motion blur score and the proportion of the highlight occlusion area of ​​the key image frame; when the motion blur score exceeds the preset blur threshold, or the proportion of the highlight occlusion area exceeds the preset area threshold, step S101 is triggered to perform the image re-acquisition action; wherein, the motion blur score is determined based on the response variance of the key image frame under the Laplacian operator.

3. The intelligent quality inspection method for used mobile phones according to claim 1, characterized in that, The specific method for constructing the topology-preserving local mesh in step S103 is as follows: within the key image frame, the effective display area is determined by semantic segmentation, and the interface control with a fixed position is located within the effective display area as the main semantic anchor point, and the center point of the text line spacing is located as the auxiliary semantic anchor point. A set of control points is established based on the main semantic anchor points and auxiliary semantic anchor points. The nonlinear mapping rule between the control point set and the actual physical coordinate system is fitted by the thin plate smooth spline function to generate a non-uniform topological mesh covering the entire target to be detected.

4. The intelligent quality inspection method for used mobile phones according to claim 1, characterized in that, Step S102 extracts a set of physical damage feature labels, including labels for identifying broken display areas, bumped edges, scratches on the body, aging of camera modules, and liquid leakage in the display area of ​​the target to be detected. Each physical damage feature label in the set of physical damage feature labels is associated with corresponding location coordinate information and damaged area data, which serve as quantitative input parameters for the loss level characterization report in step S105.

5. The intelligent quality inspection method for used mobile phones according to claim 1, characterized in that, The control logic for performing pixel reset sampling in step S103, which preserves the topology of the local mesh, follows these rules: It utilizes the projection transformation matrix of each local mesh cell. Correcting local distortion, projection transformation matrix The following relationship must be satisfied: ,in, The original feature coordinates of the j-th anchor point within the i-th local mesh cell are given. To correspond to the target feature coordinates in the standardized coordinate system, To maintain the continuity of adjacent grid cells, the topological constraint term is k, which is the total number of internal semantic anchors contained in the local grid cell, and λ is the preset smoothing regularization coefficient.

6. The intelligent quality inspection method for used mobile phones according to claim 1, characterized in that, The specific method for generating the consistency verification result in step S104 is as follows: extract the physical identification code of the target to be detected and the model of the core computing unit as the core identification item; in the process of comparing the hardware attribute feature parameters with the nominal parameters, if the matching degree of the core identification item is less than 100%, generate the underlying hardware change warning label and increase the anti-counterfeiting feature detection strength of the standardized feature image.

7. The intelligent quality inspection method for used mobile phones according to claim 3, characterized in that, The extraction logic of auxiliary semantic anchor points is as follows: Under the standardized coordinate system, the pixel projection density analysis in the vertical direction is performed on the image after preliminary correction, and the peak regions in the pixel density distribution are identified as candidate text line positions; the vertical reference of the auxiliary semantic anchor points is determined by calculating the center distance between adjacent peak regions, so as to compensate for the local stretching of the image caused by the change in the physical properties of the target to be detected.

8. The intelligent quality inspection method for used mobile phones according to claim 1, characterized in that, In step S104, the multimodal large language model uses a prompt-guided mechanism to extract hardware attribute feature parameters, including: inputting first modal data containing standardized feature images and second modal instructions containing quality inspection protocol constraints into the multimodal large language model; the second modal instructions are used to constrain the multimodal large language model to output a structured data stream, which includes memory capacity, storage space, energy component attributes, and network protocol parameters. In step S105, the generation logic of the loss level characterization report is as follows: determining the hardware baseline parameter value based on the hardware attribute feature parameters, determining the compliance correction weight based on the consistency verification result, and performing weight superposition calculation by combining the damage quantification value corresponding to the physical damage feature label set; among them, the display area leakage label and the underlying hardware change warning label have a veto right in the weight superposition calculation, that is, once identified, the rating result of the loss level characterization report is directly determined to be at the scrap level.

9. The intelligent quality inspection method for used mobile phones according to claim 4, characterized in that, The method also includes the following steps: when the physical damage feature label set contains display area scratch labels, by extracting the same feature points in the overlapping field of view and calculating the spatial transformation matrix, performing feature alignment between the macro image and the macro image, the physical depth estimate of the display area scratch is obtained; if the physical depth estimate exceeds the preset depth threshold, a physical performance degradation warning signal for the display area is generated in the loss level characterization report based on the mapping relationship between the physical depth estimate and the physical state of the display area.

10. A second-hand mobile phone intelligent quality inspection system, used to implement the second-hand mobile phone intelligent quality inspection method according to claim 1, characterized in that, include: The image acquisition module is used to acquire a multi-angle image sequence S of the target to be detected; The detection and extraction module is used to input the image sequence S into the pre-trained target detection model, extract and output the set of physical damage feature labels and the set of physical state labels of the display area corresponding to each image; The image reconstruction module is used to identify key image frames containing attribute information of the display interface from the image sequence S based on the set of physical state labels of the display area, and to perform an enhanced feature extraction process: performing orientation correction based on the image classification results of the key image frames, extracting global edge features of the display area of ​​the target to be detected and internal semantic anchors in the feature space of the key image frames, using the internal semantic anchors as topological support nodes, constructing a topology-preserving local mesh in combination with global edge features, and performing pixel resampling according to the node mapping relationship of the topology-preserving local mesh to generate a standardized feature image; The identification and verification module is used to input standardized feature images into a multimodal large language model, identify and extract hardware attribute feature parameters of the target to be detected, and compare the hardware attribute feature parameters with the pre-stored nominal parameters to generate a consistency verification result. The report generation module is used to generate a loss level characterization report of the target under test by integrating the set of physical damage feature labels, hardware attribute feature parameters and consistency verification results through weighted quantization calculation.

Images