An image analysis-based cold-rolled strip surface quality detection system

Abstract

This invention relates to the field of industrial visual inspection technology and discloses an image analysis-based surface quality inspection system for cold-rolled strip steel. In this system, an image acquisition and synchronization module simultaneously acquires images of the entire width of the strip steel; an online defect analysis module uses a defect recognition model to detect and classify defects, and innovatively performs periodic analysis of similar defects, automatically matching the roll circumference to locate suspected source stands; a defect display and correction module supports manual visualization correction of the recognition results; a model self-optimization module incrementally optimizes the defect recognition model based on the corrected data; and a comprehensive service and output module provides multi-source data fusion query and strip steel surface image playback functions. This invention realizes a closed-loop intelligent inspection mode covering the entire process from defect detection, intelligent diagnosis, human-machine collaborative optimization to deep quality traceability.

Description

Technical Field

[0001] This invention relates to the field of industrial visual inspection technology, and in particular to a surface quality inspection system for cold-rolled strip steel based on image analysis. Background Technology

[0002] Existing surface quality inspection technologies for cold-rolled strip steel generally adopt a vision inspection scheme based on the synchronization of a line scan camera and an encoder. However, in practical applications, such systems only judge periodic defects and lack the ability to automatically match the defect period in the image with the physical circumference of the specific roll and locate the source of the defect. Defect diagnosis heavily relies on the experience comparison of process personnel, which is inefficient and highly subjective.

[0003] Existing detection systems generally use fixed learning models, which are difficult to adapt to new defect types that appear on the production line. Although some studies have attempted to introduce incremental learning, under the stringent stability requirements of industry, the models are prone to catastrophic forgetting of historical defects, resulting in fluctuations in model performance and a lack of industrial stability.

[0004] The existing system's defect classification and grading rules are mostly pre-coded and set by developers, while the defect identification experience and grading rules accumulated by on-site process experts cannot be effectively fed back to the detection system. When the production line process is adjusted or new materials are put into production for the first time, the existing detection system cannot quickly respond to the dynamic changes in the production status.

[0005] The existing detection system suffers from data silos in its quality analysis and traceability process. Data such as defect images, identification results, and production process parameters are scattered across different subsystems, requiring manual cross-system queries, correlations, and comparisons, which is time-consuming and labor-intensive. Furthermore, it is difficult to quickly retrieve complete production context information at the time of defect occurrence during defect review, which severely restricts the closed-loop quality control and decision-making efficiency. Summary of the Invention

[0006] This invention provides an image analysis-based surface quality inspection system for cold-rolled strip steel to solve existing technical problems.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides an image analysis-based surface quality inspection system for cold-rolled strip steel, comprising: The image acquisition and synchronization module is used to synchronously acquire image sequences covering the upper and lower surfaces of the strip across the entire width of the strip based on the strip's position encoding, and to obtain the corresponding strip running speed. The online defect analysis module is used to identify defect areas and extract their type, location and size information based on the image sequence through a defect recognition model, and to calculate the defect cycle value based on the strip running speed and the spatiotemporal distribution of the defect area. The defect display and correction module is used to visually display the defect identification results, receive manual corrections from users, and support users to define new defect types and severity grading rules. The model self-optimization module is used to adaptively optimize the defect identification model based on the manually corrected defect identification results using an adaptive elastic weight solidification algorithm based on data tracing. The integrated service and output module is used to generate multi-dimensional statistical reports based on current and historical defect identification results, and provides a function to play back the original surface image based on the strip position code.

[0008] The beneficial effects of the technical solution provided by this invention include at least the following: This invention innovatively introduces a periodic defect intelligent diagnosis and automatic rack matching mechanism, which effectively solves the problem that traditional detection systems rely on manual experience for source location. The system can perform cluster analysis based on the similarity of defect visual features, extract the spatial periodic value of defects by combining speed fluctuation compensation, and automatically match it with a pre-stored roll circumference library to output the suspected source rack and matching degree score, which greatly improves the accuracy and timeliness of equipment maintenance.

[0009] This invention addresses the contradiction between continuous model learning and industrial stability by designing an adaptive elastic weight solidification algorithm and a model update gateway based on data traceability. It protects important parameters of the model regarding historical defects through the Fisher information matrix and combines a dynamic knowledge distillation strategy to balance the learning of new and old knowledge, ensuring that the system can maintain overall performance stability while updating and iterating, thus achieving safe and controllable model self-evolution.

[0010] This invention efficiently transforms the experience of process experts into system intelligence by constructing an open user-defined interface. All manual corrections are versioned, traced, and filtered into high-quality incremental samples to drive model optimization. This breaks through the system's rigidity bottleneck, giving it flexible adaptability and growth potential, and achieving a virtuous cycle of human-machine collaboration.

[0011] This invention system uses strip steel location coding as a link to deeply integrate multi-source data such as images, defects, and processes. Users can obtain full-dimensional correlation information of defects with one click, replay strip steel surface images, and simultaneously overlay process parameter curves. This achieves visualized root cause analysis with simultaneous image, defect, and process analysis, providing three-dimensional and scenario-based data support for production decisions. Attached Figure Description

[0012] 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.

[0013] Figure 1 This is a system structure diagram provided in an embodiment of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0015] This embodiment provides an image analysis-based surface quality inspection system for cold-rolled strip steel. Please refer to [link / reference]. Figure 1 This is a system structure diagram provided in an embodiment of the present invention.

[0016] The image analysis-based surface quality inspection system for cold-rolled strip steel in this embodiment includes: The image acquisition and synchronization module is used to synchronously acquire image sequences covering the upper and lower surfaces of the strip across the entire width of the strip based on the strip's position encoding, and to obtain the corresponding strip running speed. The online defect analysis module is used to identify defect areas and extract their type, location and size information based on image sequences and a defect recognition model, and to calculate the defect cycle value based on the strip running speed and the spatiotemporal distribution of the defect area. The defect display and correction module is used to visually display the defect identification results, receive manual corrections from users, and support users to define new defect types and severity grading rules. The model self-optimization module is used to adaptively optimize the defect identification model based on the manually corrected defect identification results using an adaptive elastic weight solidification algorithm based on data traceability. The integrated service and output module is used to generate multi-dimensional statistical reports based on current and historical defect identification results, and provides a function to play back the original surface image based on the strip position code.

[0017] In one embodiment of the present invention, the image acquisition and synchronization module includes the following sub-modules: The light source and line scan camera sub-modules are deployed at the inspection stations on the upper and lower surfaces of the strip. The fields of view of adjacent line scan cameras are calibrated and stitched together by a preset overlap rate to cover the full width of the strip. The signal acquisition submodule is used to acquire rotary encoder signals synchronized with the main drive shaft of the production line in real time and convert them into strip position pulses; The trigger control submodule is used to receive strip position pulses and generate synchronous trigger signals according to the preset sampling interval, which are then sent to the line scan camera to drive it to acquire image sequences.

[0018] It should be noted that in this embodiment, the light source is a high-frequency LED linear light source, which is arranged on both sides of the upper and lower surface inspection station of the strip steel to ensure that the strip steel surface can still obtain uniform and shadow-free illumination under high-speed movement; the linear scan camera is a high-resolution industrial-grade camera (such as a resolution of 8192 pixels or higher), and one or more cameras are deployed at each upper and lower surface inspection station to collect a maximum image width of 755mm. The fields of view of adjacent cameras are stitched together by calibration algorithm, and the field of view overlap rate is not less than 10%. In addition, compressed air is used for cooling and dust removal of the camera and light source equipment.

[0019] The rotary encoder is installed on the main drive shaft of the production line. Its output pulse frequency is proportional to the running speed of the strip. The signal acquisition submodule receives the encoder signal through a high-speed data acquisition card and converts it into strip position pulses in real time. The signal is then transmitted to the trigger control submodule via Ethernet or fieldbus. The latter generates a trigger signal according to a preset sampling interval (e.g., 1ms) to drive all line scan cameras to acquire images synchronously. Each frame of the image is accompanied by a strip position code and a timestamp. The sequence of multiple frames of images is transmitted to the image storage server through a high-speed network and stored in segments according to the strip number and its position code for easy subsequent analysis and playback.

[0020] In addition, the system supports online self-calibration, which periodically calibrates the camera field of view, light source intensity, and encoder signal to ensure detection accuracy. If a camera or light source fails, it will automatically switch to backup equipment (if available) or adjust the overlap rate of the field of view of adjacent cameras to ensure detection continuity, and at the same time send maintenance notices to management personnel.

[0021] In one embodiment of the present invention, the online defect analysis module is connected to the image acquisition and synchronization module, and includes the following sub-modules: The defect localization submodule is used to run the defect recognition model, identify defect regions in the image sequence, and output their location parameters, including pixel coordinates and cross-frame start position encoding. The feature extraction and defect classification submodule is used to calculate the visual features of the defect region, including morphological features, texture features and grayscale statistical features, and determine the defect type based on the extracted visual features; The period value calculation submodule is used to analyze the spatial distribution pattern of defects identified as being of the same type and calculate their period values ​​based on the starting position code of the defect area, timestamp, and the running speed of the strip.

[0022] It should be noted that in this embodiment, the defect recognition model adopts a deep learning-based convolutional neural network structure, such as the improved YOLOv7 or Mask R-CNN model. During training, the defect recognition model uses a sample library of surface defects of strip steel, covering a variety of typical defect types such as scratches, oxidation, pitting, and roll marks, and combines data augmentation technology to improve generalization ability.

[0023] The defect localization result is output as the coordinates of the defect bounding box and its encoded value relative to the starting position of the strip. It supports cross-frame tracking to ensure continuous identification of elongated defects. Preferably, for each identified defect region, the following multi-dimensional visual features are extracted: Morphological characteristics: area, perimeter, aspect ratio, and convexity; Texture features: Contrast and energy based on the gray-level co-occurrence matrix; Gray-scale statistical characteristics: mean, variance, skewness, and kurtosis.

[0024] By using a feature classifier based on support vector machines or lightweight fully connected neural networks, combined with the aforementioned visual features, the defect type is determined, and the defects are initially grouped. Subsequently, feature similarity calculation and periodic analysis are only performed within the same type of defect.

[0025] In this embodiment, the sub-modules of the online defect analysis module adopt a multi-threaded parallel processing architecture and support GPU-accelerated inference to ensure real-time analysis and response under high-speed strip operation. When computing resources are scarce, the image resolution or feature extraction complexity can be dynamically adjusted to ensure continuous system operation.

[0026] In one embodiment of the present invention, the periodic value calculation submodule is specifically used for: For defects of the same type and with visual feature similarity greater than a preset similarity threshold, the starting position encoding sequence is used, a sliding window Fourier transform based on adaptive bandwidth is adopted, and dynamic compensation is performed to address the impact of strip speed fluctuations on spatial sampling intervals, and period estimation is performed in the spatial domain. The estimated results are matched with the pre-stored roll circumferences of each stand to identify the suspected source stand that caused the periodic defect. The suspected source stand identifier with matching score and its roll circumference are output as the defect period value.

[0027] It should be noted that in this embodiment, the preset similarity threshold is set to 0.85. For each group of defects of the same type identified from the same coil of steel, the system further performs cluster analysis within each group based on visual feature similarity (such as using cosine similarity or Euclidean distance), and only retains defects with similarity ≥ 0.85 as candidate defects for periodic analysis, so as to exclude the interference of occasional defects on periodic analysis.

[0028] Subsequently, the spatial position sequence of candidate defects is constructed based on the starting position encoding of the candidate defects on the strip. At the same time, the strip running speed signal is received in real time. The system sets a preset speed fluctuation threshold (for example, set to ±1.5% of the set speed of the strip). If the fluctuation of the real-time speed relative to the set speed exceeds this threshold, it is determined that the speed fluctuation is significant. Dynamic resampling and interpolation compensation of the defect position sequence are required to eliminate the spatial sampling interval distortion caused by the uneven speed.

[0029] The system employs a sliding window mechanism to segment and analyze the spatial sequence. The window length is adaptively adjusted based on the defect density and distribution. Within each window, a fast Fourier transform is performed to extract frequency domain features. The spectral resolution is dynamically set according to the window length and defect distribution to ensure good detection sensitivity for both short-periodic and long-periodic defects. Finally, the spatial periodicity of the defect is estimated by analyzing the frequency components corresponding to the spectral peaks of each window.

[0030] The estimated spatial period value is matched with the theoretical circumference of each stand roll pre-stored in the system (preferably, the matching process takes into account the small changes in circumference caused by roll wear, allowing an error range of ±2%). Based on the period matching error, the uniformity of defect distribution and the significance of spectral peaks, a matching degree score is calculated for each matching result. The stand number with the highest matching degree score and its corresponding roll circumference are output as the suspected source of the periodic defect.

[0031] In addition, the system supports operators to manually confirm or correct the periodic analysis results. The corrected results will be fed back to the periodic analysis model as training samples to optimize the subsequent matching algorithm. For frequently occurring periodic defects, the system can automatically generate alarms and push them to the equipment maintenance system to prompt the corresponding frame to perform roller surface inspection or replacement.

[0032] In one embodiment of the present invention, the defect display and correction module is connected to the online defect analysis module, and includes the following sub-modules: The panoramic visualization submodule is used to generate a strip surface unfolded map based on the location parameters of the defects, and to plot the defect area in the unfolded map in the form of visual primitives. The details display submodule is used to respond to user interactions and display the type, size, and two-dimensional position coordinates of the selected defect on the strip surface; The interactive correction submodule provides users with a visual correction toolbar for correcting defect identification results. It associates and stores correction operations with user ID, operation time, and defect identification model version information. Corrections include defect reclassification, adjusting defect area boundaries, and deleting and adding defect entries.

[0033] It should be noted that the system dynamically generates a two-dimensional unfolded diagram of the strip surface based on the strip position code and width information, and presents it in a length-width coordinate system. Defects are marked in the two-dimensional unfolded diagram of the strip surface with visual elements of different colors and shapes. In this embodiment, different colors represent different defect types (e.g., red for scratches and blue for roll marks), and different shapes represent different degrees of severity (e.g., circles for minor defects and triangles for severe defects).

[0034] Users can zoom in and out of the view using the mouse wheel and drag to pan and view any section of the strip. The system supports filtering by defect type, severity level, and location range in a list view. When a user clicks on any defect element in the expanded view or list view, the system displays detailed information about the defect in a pop-up window, including defect type and confidence score, defect dimensions (length, width, area, unit: millimeters), precise two-dimensional location coordinates (distance from strip head, distance from strip edge), strip coil number, inspection time, rack position, and other production-related information, as well as a magnified view of the original defect image area.

[0035] In this embodiment, the system also provides a visual correction toolbar interface, through which users can perform the following operations on defects: (1) Defect reclassification: Select the type to be modified from a preset list of defect types, and then the system will automatically update the display color and display shape of the defect; (2) Adjust the boundary of the defect area: Adjust the detection range of the defect area by dragging the control anchor point of the defect area boundary box; (3) Remove defects: Remove false detections or irrelevant markers. You can fill in / select the reason for deletion as needed (such as "non-defect" or "duplicate marker"). (4) Add new defects: Select the area in the expanded view and manually add the corresponding defect entry. You must also manually specify the defect type and defect level. All the above correction operations record the operator ID, timestamp, pre-operation status, and post-operation status, and are associated with the current system version and the defect identification model version to form a traceable correction log. Manual correction operations are synchronized to the system backend database in real time and trigger the sample collection process of the model self-optimization module.

[0036] In one embodiment of the present invention, the defect display and correction module further includes the following sub-modules: The periodic defect annotation submodule is used to highlight periodic defects in the strip surface unfolded view and display their periodic value and suspected source frame location information. The critical defect alarm submodule is used to highlight defects that reach a preset severity level in the strip surface unfolded diagram and trigger an alarm, while also displaying the corresponding defect identification results and handling instructions. The classification system customization submodule provides authorized users with a classification system management interface, allowing them to expand the defect type dictionary, including defining new defect type names, uploading typical defect sample images, and customizing defect visual feature descriptions. The classification rule customization submodule provides authorized users with a classification rule configuration interface, allowing them to customize or modify the classification rules for defect severity levels based on one or more dimensions such as defect type, size, location, and periodicity.

[0037] It should be noted that in this embodiment, for defects identified as having periodicity, a flashing or bright border, which is different from that of conventional defects, is used for visual marking in the strip surface unfolding diagram. When the user clicks on such a defect, the interface will display its periodicity analysis details, including the periodicity estimate (unit: meters) and confidence level, the matching suspected source frame number, the roll position and circumference information, and the distribution histogram of the same type of periodic defects on the whole coil of strip. In contrast to regular defects, users can also identify, correct, or mark periodic defects as "non-periodic." Correction operations are synchronized to the system's backend database in real time and trigger the sample collection process of the model self-optimization module.

[0038] In this embodiment, the system monitors and alerts in real time to defects whose severity level reaches the alarm threshold according to preset or user-defined grading rules. In the unfolded diagram, severe defects are marked with flashing red icons, and an alarm window automatically pops up, displaying the defect type, size, location and image, suggested handling measures (such as: shutdown inspection, downgrade processing or key re-inspection), and relevant process parameters. The alarm information is simultaneously pushed to the terminal equipment of the relevant positions, and the alarm time, the person confirming the alarm and the handling status are recorded.

[0039] In this embodiment, a graphical configuration interface is provided to authorized users with system administrator or engineer privileges. Authorized users are allowed to perform the following operations through this graphical configuration interface: (1) Add a new defect type and assign it a unique type code, display name and default color; (2) Upload a typical sample image of this type of defect; (3) Select and fill in the characteristic description template for this type of defect (such as morphological characteristics, common causes and process correlations); Once the content added or modified through the graphical configuration interface takes effect, it should be synchronized to the training process of the defect identification model.

[0040] In this embodiment, a rule configuration panel is also provided to authorized users with system administrator or engineer privileges. Authorized users are allowed to define defect severity levels through this panel by dynamically combining multiple dimensions such as defect type, size range, location information, and periodic characteristics. Some example configuration rules are provided, including: Defect type: "Scratches" type defects are automatically upgraded; Size range: Length greater than 50mm or area greater than 100mm² 2 The defect level has been upgraded; Location information: Enhanced monitoring of defects located at the edge of the strip or in the joint area; Periodic characteristics: Periodic defects are automatically upgraded to the "high-priority concern" level; The system supports setting rule priorities, simulation testing, and historical data backtesting to ensure consistency in system judgments after rule changes.

[0041] In one embodiment of the present invention, the model self-optimization module is connected to the defect display and correction module, and includes the following sub-modules: The incremental learning sample quality management submodule is used to perform quality filtering and version tracing on the manual correction results from the defect display and correction module. Only when the correction operation comes from an authorized user and the difference between the defect identification results before and after the correction exceeds the preset reliability threshold, it is accepted as a valid training sample. All valid training samples are automatically associated with the defect identification model version number, system software version number and operation user ID that triggered the manual correction, forming an incremental sample set.

[0042] It should be noted that in this embodiment, the system maintains a list of authorized users and only receives correction operations submitted by users in this list. User identity is double-verified through login token and permission level, and the following traceability information is automatically captured and associated for each manual correction: Defect identification model version number: Records the model version identifier that triggered this fix; System software version number: Records the current version of the front-end interface and back-end services; User ID and Role: Record the user ID and permission level of the user who performed the modification; Correction timestamp and production line status: Record the time when the correction occurred and the production process parameters at that time; All traceability information is stored in a structured format along with the defect images and coordinate information before and after correction, forming traceable incremental sample units.

[0043] For each correction, the system calculates the difference in recognition results before and after the correction. The difference is evaluated based on the following factors: (1) Has the defect type changed after the correction? (2) Whether the change in the boundary overlap rate (hereinafter referred to as IoU) of the defect area after correction exceeds the preset overlap rate threshold; (3) Whether the difference between the original recognition confidence level and the corrected user-specified confidence level exceeds the preset difference threshold; In a preferred embodiment, the manual correction result is considered a valid sample only when the modified defect type changes or the IoU changes by ≥30%, and the difference between the original identification confidence and the modified user-specified confidence is ≥20%. The valid samples are then organized into incremental sample sets in chronological order, and each incremental sample set is associated with a model training cycle.

[0044] The system generates a unique ID for each training sample entry in the incremental sample set and establishes an index, supporting multi-dimensional queries and filtering by time, user, defect type, and model version.

[0045] In addition, the system regularly performs statistical analysis on the distribution uniformity of incremental samples and the consistency of user corrections, and performs conflict detection to generate sample quality reports. In the event that multiple authorized users make inconsistent corrections to the same type of defect, the system can trigger an arbitration process and prompt the superior engineer for final confirmation.

[0046] In one embodiment of the present invention, the model self-optimization module further includes the following sub-modules: The constrained incremental training submodule, connected to the incremental learning sample quality management submodule, executes an adaptive elastic weight solidification algorithm based on data source attribution, including: (a) Based on the historical model version number associated with the incremental sample set, trace the distribution of training samples of the corresponding version, and calculate the importance weight of the model parameters of the current defect identification model to each historical defect type based on the Fisher information matrix under the Bayesian inference framework. (b) The importance weight is used as a regularization constraint, and the weight of the knowledge distillation loss term is dynamically adjusted according to the defect type associated with the new sample to optimize the defect identification model. (c) Based on the independent validation set, evaluate the performance of the optimized defect identification model. Evaluation metrics include the retention rate of historical defect types, the identification rate of new defect types, and the change in the overall false alarm rate. Only when all evaluation metrics are not lower than the corresponding preset thresholds, a new version number is automatically generated and deployed.

[0047] It should be noted that in this embodiment, the system retrieves the corresponding historical version training sample set from the version library based on the defect identification model version number recorded in the incremental sample set, and reconstructs the sample distribution characteristics of each defect type.

[0048] In this embodiment, the system is based on the Bayesian inference framework. It calculates the second derivative of the predicted log-likelihood of the defect assessment model for various historical defect samples under the current model parameters, constructs the Fisher information matrix, diagonalizes the Fisher information matrix, extracts the importance score of each model parameter for each type of defect, and normalizes it to obtain the weight vector of each model parameter for each type of defect, which is used as the importance weight.

[0049] During incremental training, the loss function Designed as follows: In the formula, To address the cross-entropy loss for newly added samples, For model parameters, Assigning importance weights to each parameter. These are the initial values ​​of the current model parameters before training begins. The balancing hyperparameter is dynamically adjusted based on the number of various defects in the newly added samples; For tasks involving the integration of new and old knowledge, a knowledge distillation loss term is also introduced. : In the formula, The fitness coefficient is used to adaptively adjust the new samples based on whether they contain old category samples. It is inversely correlated with the number of old category samples in the new samples.

[0050] The optimized defect identification model is comprehensively evaluated on an independent validation set. In this embodiment, the validation set includes representative samples of historical defect types, samples of newly added defect types, and easily confused non-defect samples. The evaluation metrics include: historical defect type retention rate, new defect type identification rate, and overall false alarm rate change. In a preferred embodiment, the three metrics that meet the preset thresholds are as follows: The retention rate of historical defect types is ≤3%, meaning that the decrease in accuracy for identifying historical categories is no more than 3% compared to the original model; The new defect type identification rate is ≤85%, meaning the average identification rate for new categories is not less than 85%. The overall false alarm rate change is ≤1%, meaning the increase in the false alarm rate does not exceed 1%. At this point, the system automatically generates a new version number and starts the deployment process, which includes exporting model parameters, hot updating services, and recording version logs. If the new model fails the evaluation, the system automatically rolls back to the previous stable version and records the reason for the failure. In addition, the system supports operators to manually adjust the evaluation threshold, re-trigger training, or manually review the sample quality.

[0051] In one embodiment of the present invention, the integrated service and output module is connected to the image acquisition and synchronization module, the defect display and correction module, and the model self-optimization module, and includes the following sub-modules: The data association retrieval submodule is used to respond to queries based on strip location code, defect number, time range, or defect type, and to perform collaborative retrieval and information association from image sequences stored in the image acquisition and synchronization module, defect identification and periodic location results generated by the online defect analysis module, and production process parameters associated with the defect sample database module. The multi-source information fusion and publishing submodule is connected to the data association and retrieval submodule. It is used to fuse the retrieved multi-source information and publish it as reports and image playback.

[0052] It should be noted that in this embodiment, the system establishes a unified index for the following data sources: Image sequence library: segmented storage and indexing based on strip coil number, location code, and acquisition time; Defect Information Database: Stores defect number, type, location, size, periodic analysis results, and manual correction records; Production process parameter library: Real-time production process parameters such as rolling speed, rolling force, tension, temperature and lubrication status.

[0053] The system supports any combination of the following query methods: Location-driven query: Enter the strip location code (e.g., "coil number A, location 1200-1500m"), and return all defect information, corresponding images and production process parameters for that section; Defect ID lookup: Enter the unique defect ID to obtain the complete information chain of the defect, including the original image, recognition results, correction history, cycle analysis and related process parameters; Time range query: Specify a time range (e.g., "2025-10-01 08:00 to 12:00") to retrieve all defect records and related data within that time period; Defect type query: Filter by defect type (such as "scratches" or "roller marks"), and support multiple type combination queries.

[0054] Real-time integration of related information: Search results are returned in a structured format, and each data item is automatically linked through strip location codes and timestamps to ensure the consistency and integrity of information.

[0055] In this embodiment, the system has built-in various report templates, such as "Defect Classification Statistics Table", "Periodic Defect Analysis Report" or "Serious Defect Traceability Report", etc. Users can customize the report content, select display fields, statistical dimensions (such as by shift, steel type, rack) and chart types (bar chart, trend chart, distribution chart), and the generated reports can be exported in real time to common office document formats.

[0056] In this embodiment, the image playback function includes: dynamically playing back surface image sequences of any segment based on strip position coding, with adjustable playback speed. During playback, defect locations are superimposed as highlighted boxes, and users can click to view detailed identification information of defects displayed in the playback image sequence. In addition, the system synchronously plays the key process parameter curves (such as rolling force and speed) corresponding to the segment, realizing synchronous visualization analysis of "image-defect-process".

[0057] The system-generated reports and playback links can be pushed to relevant positions via web interface, email or message middleware (such as MQTT, Kafka), and support integration with production management systems such as MES and ERP to achieve automatic synchronization of defect data and quality judgment results.

[0058] In addition, all retrieval and publishing operations are bound to the user ID of the user and are verified and authorized to ensure data access security. It supports encrypted storage of sensitive data (such as images of serious defects and their process parameters) and only de-identifies the data for unauthorized users.

[0059] Furthermore, it should be noted that the present invention can be provided as a method, apparatus, or computer program product. Therefore, embodiments of the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, embodiments of the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code.

[0060] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0061] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxesFigure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing terminal equipment to cause a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0062] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes the element.

[0063] Finally, it should be noted that the above description represents a preferred embodiment of the present invention. It should be pointed out that although preferred embodiments have been described, those skilled in the art, once they understand the basic inventive concept of the present invention, can make various improvements and modifications without departing from the principles described herein. These improvements and modifications should also be considered within the scope of protection of the present invention. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the present invention.

Claims

1. A surface quality inspection system for cold-rolled strip steel based on image analysis, characterized in that, include: The image acquisition and synchronization module is used to synchronously acquire image sequences covering the upper and lower surfaces of the strip across the entire width of the strip based on the strip's position encoding, and to obtain the corresponding strip running speed. The online defect analysis module is used to identify defect areas and extract their type, location and size information based on the image sequence through a defect recognition model, and to calculate the defect cycle value based on the strip running speed and the spatiotemporal distribution of the defect area. The defect display and correction module is used to visually display the defect identification results, receive manual corrections from users, and support users to define new defect types and severity grading rules. The model self-optimization module is used to adaptively optimize the defect identification model based on the manually corrected defect identification results using an adaptive elastic weight solidification algorithm based on data tracing. The integrated service and output module is used to generate multi-dimensional statistical reports based on current and historical defect identification results, and provides a function to play back the original surface image based on the strip position code.

2. The image analysis-based surface quality inspection system for cold-rolled strip steel according to claim 1, characterized in that: The image acquisition and synchronization module includes the following sub-modules: The light source and line scan camera sub-modules are deployed at the inspection stations on the upper and lower surfaces of the strip. The fields of view of adjacent line scan cameras are calibrated and stitched together by a preset overlap rate to cover the full width of the strip. The signal acquisition submodule is used to acquire rotary encoder signals synchronized with the main drive shaft of the production line in real time and convert them into strip position pulses; The trigger control submodule is used to receive the strip position pulse and generate a synchronous trigger signal according to the preset sampling interval, which is then sent to the line scan camera to drive it to acquire image sequences.

3. The image analysis-based surface quality inspection system for cold-rolled strip steel according to claim 1, characterized in that: The online defect analysis module is connected to the image acquisition and synchronization module, and includes the following sub-modules: The defect localization submodule is used to run the defect recognition model, identify defect regions in the image sequence, and output their location parameters, including pixel coordinates and cross-frame start position encoding. The feature extraction and defect classification submodule is used to calculate the visual features of the defect region, including morphological features, texture features and grayscale statistical features, and determine the defect type based on the extracted visual features; The period value calculation submodule is used to analyze the spatial distribution pattern of defects identified as being of the same type and calculate their period values ​​based on the starting position code of the defect area, timestamp, and the running speed of the strip.

4. The image analysis-based surface quality inspection system for cold-rolled strip steel according to claim 3, characterized in that: The periodic value calculation submodule is specifically used for: For defects of the same type and with visual feature similarity greater than a preset similarity threshold, the starting position encoding sequence is used, a sliding window Fourier transform based on adaptive bandwidth is adopted, and dynamic compensation is performed to address the impact of strip speed fluctuations on spatial sampling intervals, and period estimation is performed in the spatial domain. The estimated results are matched with the pre-stored roll circumferences of each stand to identify the suspected source stand that caused the periodic defect. The suspected source stand identifier with matching score and its roll circumference are output as the defect period value.

5. The image analysis-based surface quality inspection system for cold-rolled strip steel according to claim 1, characterized in that: The defect display and correction module is connected to the online defect analysis module and includes the following sub-modules: The panoramic visualization submodule is used to generate a strip surface unfolded map based on the location parameters of the defects, and to plot the defect area in the unfolded map in the form of visual primitives. The details display submodule is used to respond to user interactions and display the type, size, and two-dimensional position coordinates of the selected defect on the strip surface; The interactive correction submodule provides users with a visual correction toolbar for correcting defect identification results. It associates and stores correction operations with user ID, operation time, and defect identification model version information. The corrections include defect reclassification, adjusting defect area boundaries, and deleting and adding defect entries.

6. The image analysis-based surface quality inspection system for cold-rolled strip steel according to claim 5, characterized in that: The defect display and correction module also Includes the following sub-modules: The periodic defect marking submodule is used to highlight periodic defects in the unfolded view of the strip surface and display their periodic value and suspected source frame positioning information. The critical defect alarm submodule is used to highlight defects that reach a preset severity level in the unfolded diagram of the strip surface and trigger an alarm, while also displaying the corresponding defect identification results and handling instructions. The classification system customization submodule provides authorized users with a classification system management interface, allowing them to expand the defect type dictionary, including defining new defect type names, uploading typical defect sample images, and customizing defect visual feature descriptions. The classification rule customization submodule provides authorized users with a classification rule configuration interface, allowing them to customize or modify the classification rules for defect severity levels based on one or more dimensions such as defect type, size, location, and periodicity.

7. The image analysis-based surface quality inspection system for cold-rolled strip steel according to claim 1, characterized in that: The model self-optimization module is connected to the defect display and correction module. Includes the following sub-modules: The incremental learning sample quality management submodule is used to perform quality filtering and version tracing on the manual correction results from the defect display and correction module. Only when the correction operation comes from an authorized user and the difference between the defect identification results before and after the correction exceeds the preset reliability threshold, it is accepted as a valid training sample. All valid training samples are automatically associated with the defect identification model version number, system software version number and operation user ID that triggered the manual correction, forming an incremental sample set.

8. The image analysis-based surface quality inspection system for cold-rolled strip steel according to claim 7, characterized in that: The model self-optimization module also includes the following sub-modules: A constrained incremental training submodule, connected to the incremental learning sample quality management submodule, executes an adaptive elastic weight solidification algorithm based on data source tracing, including: (a) Based on the historical model version number associated with the incremental sample set, trace the distribution of training samples of the corresponding version, and calculate the importance weight of the model parameters of the current defect identification model to each historical defect type based on the Fisher information matrix under the Bayesian inference framework. (b) Using the aforementioned importance weights as regularization constraints, and dynamically adjusting the weights of the knowledge distillation loss term based on the defect types associated with the newly added samples, the defect identification model is constrained and optimized. (c) Based on the independent validation set, evaluate the performance of the optimized defect identification model. Evaluation metrics include the retention rate of historical defect types, the identification rate of new defect types, and the change in the overall false alarm rate. Only when all evaluation metrics are not lower than the corresponding preset thresholds, a new version number is automatically generated and deployed.

9. The image analysis-based surface quality inspection system for cold-rolled strip steel according to claim 1, characterized in that: The integrated service and output module is connected to the image acquisition and synchronization module, the defect display and correction module, and the model self-optimization module, and includes the following sub-modules: The data association retrieval submodule is used to respond to queries based on strip location code, defect number, time range, or defect type by performing collaborative retrieval and information association from the image sequence stored in the image acquisition and synchronization module, the defect identification and periodic location results generated by the online defect analysis module, and the production process parameters associated with the defect sample database module. The multi-source information fusion and publishing submodule is connected to the data association and retrieval submodule and is used to fuse the retrieved multi-source information and publish it as reports and image playback.

Images