Belt conveyor anti-deviation vision recognition system based on image processing

Abstract

This invention discloses a visual recognition system for preventing belt misalignment in belt conveyors based on image processing, specifically relating to the field of image recognition. It includes an image acquisition and preprocessing module, a dual-feature hierarchical extraction module, a frame interference adaptive removal module, an illumination interference dynamic correction module, a medium interference multi-dimensional compensation module, a belt feature multi-dimensional fusion and quantization module, a misalignment feature depth analysis module, and a feature data temporal management module. The image acquisition and preprocessing module performs camera calibration, condition-adaptive acquisition, and multi-algorithm collaborative preprocessing. The dual-feature hierarchical extraction module performs image region segmentation, multi-scale feature extraction, feature validity verification, and filters and transmits valid feature parameters. This invention significantly improves the quality of preliminary data for image recognition through condition-adaptive acquisition and multi-algorithm collaborative preprocessing techniques in the image acquisition and preprocessing module.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically, to a visual recognition system for preventing belt conveyor deviation based on image processing. Background Technology

[0002] Belt conveyors are core equipment for continuous conveying. Belt misalignment can directly lead to problems such as belt wear, material spillage, equipment jamming, and even shutdown. Therefore, real-time detection and identification of belt misalignment is crucial for ensuring stable equipment operation. Currently, the industry mainstream adopts non-contact machine vision inspection solutions. Relying on core technologies such as image recognition and image processing, these solutions use industrial cameras to capture images of the belt in operation. Combined with algorithms such as image segmentation, edge detection, and centerline fitting, visual features of the belt are extracted, enabling automated determination of misalignment. This replaces the traditional contact-based misalignment switch detection method and has become the mainstream technology for belt conveyor misalignment prevention.

[0003] Existing machine vision-based visual recognition technology for belt conveyor misalignment prevention, while possessing advantages such as non-contact operation with no mechanical wear, fast response speed, and direct acquisition of visual information about belt operation, still suffers from several core technical shortcomings in complex industrial applications, making it difficult to meet the requirements for high-precision and high-stability recognition. Firstly, the lack of preprocessing to adapt to the operating conditions after image acquisition, followed by direct feature extraction, results in the inability to effectively eliminate noise and distortion in the acquired images, significantly reducing the accuracy of subsequent feature extraction. Secondly, the extraction of belt body features and environmental interference features lacks a clear hierarchical strategy, employing only a single algorithm for global feature extraction. False features formed by environmental interference such as frame obstruction, lighting changes, dust, and water mist can mix with the belt body features, making it impossible to effectively eliminate false misalignment signals and leading to a high false judgment rate in image recognition. Thirdly, the handling of environmental interference only uses a single global correction method, failing to consider the different types of interference from the frame, lighting, and media. The design of targeted processing methods based on the same formation mechanism results in poor interference correction, and the image processing error increases significantly with changes in operating conditions. Fourth, the extraction of belt body features is limited to a single scale level, without combining the characteristic manifestations of belt misalignment faults to conduct multi-scale feature mining. The sensitivity to fine-scale features such as edge micro-deformation and texture micro-distortion caused by early slight misalignment is low, making it impossible to achieve early warning of misalignment faults. Fifth, the determination of misalignment status relies solely on threshold comparison of a single feature parameter, without constructing a multi-dimensional feature fusion and quantification system. There is a lack of in-depth quantitative analysis of the degree of misalignment, and the scientific rigor and precision of the determination method are insufficient, failing to provide accurate basis for graded early warning of misalignment faults. Sixth, the management of feature data is merely simple storage, without dynamic updates and benchmark iterations based on the time sequence of belt operation. This results in the inability to solve the feature benchmark drift problem caused by equipment aging and changes in operating conditions during long-term operation, leading to poor long-term stability of image recognition.

[0004] To address the shortcomings of existing machine vision-based visual recognition technologies for belt conveyor misalignment prevention, such as lack of image preprocessing, absence of hierarchical feature extraction strategies, lack of targeted handling of environmental interference, insufficient multi-scale feature mining, lack of feature fusion and quantization, and lack of dynamic iteration in data management, a belt conveyor misalignment prevention visual recognition system based on image processing is needed. This system would utilize image recognition, image processing, and machine vision as core technologies, constructing an integrated technical architecture encompassing image acquisition and preprocessing, dual-feature hierarchical extraction, adaptive removal of frame interference, dynamic correction of illumination interference, multi-dimensional compensation for media interference, multi-dimensional fusion and quantization of belt features, in-depth analysis of misalignment features, and time-series management of feature data. Through condition-adaptive image preprocessing, multi-scale hierarchical feature extraction, targeted handling of environmental interference, multi-dimensional feature fusion and quantization, and dynamic data management, the system would achieve high-precision and high-stability identification of belt misalignment, resolving various core deficiencies of existing technologies and improving the application system of machine vision technology in the field of belt conveyor misalignment detection. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides a visual recognition system for preventing belt conveyor deviation based on image processing, which solves the problems mentioned in the background art through the following solution.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a visual recognition system for preventing belt conveyor deviation based on image processing, comprising: Image acquisition and preprocessing module: Completes camera calibration, performs condition-adaptive acquisition, and executes multi-algorithm collaborative preprocessing; Dual-feature hierarchical extraction module: performs image region segmentation, multi-scale feature extraction, verifies feature validity, and filters and transmits valid feature parameters; Rack interference adaptive culling module: Generates pixel-level masks based on effective feature parameters, reconstructs edge feature parameters, and transmits the corrected edge feature parameters; Illumination interference dynamic correction module: performs local brightness adaptive equalization based on relevant characteristic parameters, dynamically corrects centerline characteristic parameters, and transmits the corrected centerline characteristic parameters; Multi-dimensional compensation module for media interference: Based on relevant feature parameters, it performs image deblurring and reconstruction, restores texture feature parameters, and transmits the corrected texture feature parameters; The belt feature multi-dimensional fusion quantization module: calls the corrected feature parameters to obtain the mean radius of curvature, constructs multi-dimensional quantization indicators, and transmits the fused quantization indicator parameters; Deep analysis module for deviation characteristics: Constructs a comprehensive deviation quantitative index based on multi-dimensional quantitative indicators and transmits the comprehensive deviation quantitative index; Feature data time-series management module: provides data retrieval services, performs time-series data storage, carries out dynamic benchmark updates, and performs invalid data cleaning.

[0007] Preferably, the camera calibration operation employs a camera calibration algorithm to calibrate the camera's intrinsic and extrinsic parameters, establishes a dynamic mapping relationship between the image pixel coordinate system and the physical coordinate system, and obtains the total number of pixels in a single frame image through an image segmentation algorithm. Effective pixel ratio of belt area and pixel reference physical size The parameters obtained through the image segmentation algorithm are transmitted to the feature data time-series management module for storage. The working condition adaptation acquisition uses an industrial line scan camera to continuously acquire real-time images of the belt conveyor throughout its operation. The camera's line frequency is adaptively adjusted according to the belt speed to achieve precise matching between the acquired image and the belt's operating state. The multi-algorithm collaborative preprocessing includes using an adaptive median denoising algorithm based on neighborhood grayscale statistics to eliminate salt-and-pepper noise and Gaussian noise in industrial working conditions, using an adaptive contrast enhancement algorithm with the Laplacian operator to improve the visual contrast between the belt and the background, and using a perspective transformation algorithm to correct image perspective distortion caused by camera installation, resulting in a noise-free, distortion-free, and high-contrast preprocessed image. The image noise grayscale variance is extracted during the preprocessing process. Image contrast enhancement coefficient Simultaneously, the preprocessed image, image noise grayscale variance, and image contrast enhancement coefficient are transmitted to the dual-feature layer extraction module and stored in the feature data time-series management module.

[0008] Preferably, the image region segmentation is based on an image segmentation algorithm to divide the preprocessed image into two independent regions: the belt body feature region and the environmental interference feature region; the multi-scale feature extraction is used to extract the radius of curvature of one side of the belt's edge from the belt body feature region using fine-scale edge detection. and the vertical distance between the belt edge and the baseline Coarse-scale texture analysis was used to extract the lateral deviation between the real-time centerline and the theoretical centerline of the belt. Change in the slope of the belt centerline and the mean deviation of the direction angle of the belt surface texture For environmental interference feature regions, a feature clustering algorithm is used to extract the number of overlapping pixels between the frame and belt edges. Global illumination gradient value of the image Image medium scattering blur radius Light transmittance in areas blocked by the medium The feature validity verification uses historical benchmark feature parameters in the feature data time-series management module as a reference to obtain the deviation rate between the real-time extracted parameters and the benchmark parameters. The validity of the parameters is determined based on the deviation rate range. If the deviation rate is within a reasonable range, the parameters are considered valid; if the deviation rate exceeds the reasonable range, the parameters are considered invalid and are re-extracted. The filtering and transmission of valid feature parameters is used to transmit all the filtered valid feature parameters to the feature data time-series management module for storage. At the same time, the belt body feature parameters are transmitted to the rack interference adaptive rejection module, the light interference dynamic correction module, and the medium interference multi-dimensional compensation module, respectively. The environmental interference feature parameters are transmitted to the corresponding interference processing modules.

[0009] Preferably, the effective feature parameters include the radius of curvature of the single-sided edge of the belt, retrieved from the feature data time-series management module. Vertical distance between belt edge and baseline Number of pixels overlapping between the frame and belt edges At the same time, call the total number of pixels of a single frame image. Image contrast enhancement coefficient The pixel-level mask generation is based on the number of overlapping pixels between the rack and conveyor belt edges to generate a pixel-level mask for the rack occlusion area, and the distorted edge areas covered by the mask are marked and masked; the reconstructed edge feature parameters are reconstructed using a nonlinear formula to obtain the corrected edge curvature radius after removing rack interference. Vertical spacing with correction edge The transmitted and corrected edge feature parameters are used to transmit the corrected edge feature parameters to the feature data time-series management module for storage, and at the same time to the belt feature multi-dimensional fusion quantization module.

[0010] Preferably, the relevant feature parameters include the lateral deviation value of the belt centerline retrieved from the feature data time-series management module. Change in the slope of the belt centerline Global illumination gradient value of the image Simultaneously call the pixel reference physical size and image noise grayscale variance The local brightness adaptive equalization is based on the global illumination gradient value of the image to perform local brightness adaptive equalization on the belt centerline region, eliminating the centerline feature blurring problem caused by uneven illumination; the dynamic correction of centerline feature parameters uses the reference illumination gradient value in the feature data temporal management module. For reference, the characteristic parameters of the equalized centerline are dynamically corrected using a nonlinear formula to obtain the lateral deviation value of the corrected centerline after eliminating illumination interference. Change in slope of the correction centerline The corrected centerline feature parameters are used to transmit the corrected centerline feature parameters to the feature data time-series management module for storage, and at the same time to the belt feature multi-dimensional fusion quantization module.

[0011] Preferably, the relevant feature parameters include the average deviation of the direction angle of the belt surface texture, retrieved from the feature data time-series management module. Image medium scattering blur radius Light transmittance in areas blocked by the medium Simultaneously call the image noise grayscale variance Effective pixel ratio of belt area The image deblurring and reconstruction is performed using a Gaussian deblurring algorithm based on the image medium scattering blur radius to reconstruct the belt texture region, eliminating the texture blurring problem caused by medium scattering. The restored texture feature parameters are then multi-dimensionally restored using a nonlinear formula to obtain the mean value of the corrected texture direction angle deviation after eliminating medium interference. The transmitted and corrected texture feature parameters are used to transmit the corrected texture feature parameters to the feature data temporal management module for storage, and at the same time to the belt feature multi-dimensional fusion quantization module.

[0012] Preferably, the invocation of the corrected feature parameters includes invoking the corrected edge curvature radius from the feature data temporal management module. Correcting the vertical spacing of the edges 1. Correction centerline lateral deviation value , Change in slope of the correction centerline and the mean value of corrected texture orientation angle deviation Simultaneously, the number of segments representing the belt edge in a single frame image is retrieved; the average radius of curvature is determined based on the number of segments representing the belt edge in a single frame image. Obtain the mean radius of curvature of the edges of the corresponding segments. The construction of multi-dimensional quantitative indicators includes the construction of edge deformation quantitative indicators based on the feature fusion concept of machine vision. Centerline offset quantitative indicator and texture distortion metrics This enables in-depth quantification of different dimensions of belt misalignment characteristics; the transmitted and fused quantified index parameters are used to transmit the calculated multi-dimensional quantified indexes to the feature data time-series management module for storage, and simultaneously to the misalignment feature in-depth analysis module.

[0013] Preferably, the construction of the comprehensive belt misalignment quantification index is achieved by fusing edge deformation quantification index, centerline offset quantification index, and texture torsion quantification index using a nonlinear composite formula. This enables full-domain quantification of belt misalignment. The comprehensive misalignment quantification index is used to transmit the calculated comprehensive misalignment quantification index to the feature data time-series management module for storage, providing a quantitative basis for graded early warning and status monitoring of belt conveyor misalignment faults.

[0014] Preferably, the data retrieval service provides a fast retrieval service for historical benchmark data and real-time acquired data for all other modules of the system, meeting the parameter usage needs of each module; the execution of time-series data storage uses a high-speed time-series database to store all parameters of each module of the system in time series, including camera calibration parameters, preprocessing parameters, original feature parameters, corrected feature parameters, fusion quantization indicators, and comprehensive deviation quantization index, realizing data retrieval and traceability by timestamp; the dynamic benchmark update performs statistical analysis on historical feature parameters at fixed time intervals, calculates new feature benchmark parameters after removing outliers, and synchronously updates them to the parameter reference library of each module; the invalid data cleaning uses an outlier detection algorithm to clean invalid data of real-time acquired and calculated parameters, automatically removing invalid data caused by hardware failures and extreme operating conditions.

[0015] The technical effects and advantages of this invention are as follows: 1. This invention eliminates noise and distortion in the acquired images and improves the contrast between the belt and the background by using the working condition-adaptive acquisition and multi-algorithm collaborative preprocessing technology of the image acquisition and preprocessing module. It solves the problem of low feature extraction accuracy caused by direct acquisition without preprocessing in the existing technology, lays a high-precision image foundation for subsequent feature extraction, and greatly improves the quality of the early data for image recognition. 2. This invention achieves accurate separation and multi-scale mining of belt body features and environmental interference features through multi-scale layered extraction and feature validity verification technology of dual feature layered extraction module. At the same time, invalid feature parameters are screened out and eliminated, solving the problems of mixed feature extraction and failure to eliminate false deviation signals in the existing technology. It improves the pertinence and accuracy of feature extraction from the source and reduces the misjudgment rate of image recognition. 3. This invention addresses the different formation mechanisms of environmental interference in three types: rack, illumination, and medium. It designs three independent interference processing modules and adopts innovative technologies such as pixel-level masking, brightness adaptive equalization, and image deblurring and reconstruction. Combined with self-developed nonlinear formulas, it achieves targeted interference processing, solves the problem of poor interference processing effect caused by single global correction in existing technologies, and greatly improves the stability of image processing under complex industrial conditions. 4. This invention constructs a comprehensive deviation quantification index by using the multi-index fusion analysis and deviation degree quantification technology of the deviation feature deep analysis module. It abandons the traditional simple threshold comparison judgment method, solves the problem that the existing technology cannot accurately quantify the degree of deviation, realizes the technical upgrade of belt deviation from "qualitative judgment" to "quantitative evaluation", and provides accurate quantitative basis for the graded early warning of deviation faults. 5. This system uses multi-feature dimension mapping and nonlinear fusion quantization technology of the belt feature multi-dimensional fusion quantization module to mine the belt deviation features in three dimensions: edge, center line, and texture. It constructs independent quantification indicators, solves the problems of single-scale feature extraction and low sensitivity to early slight deviation recognition in existing technologies, realizes early warning of deviation faults, and improves the system's recognition foresight. 6. This invention solves the problems of simple data management and feature benchmark drift during long-term operation in existing technologies by using dynamic benchmark updates and invalid data cleaning technology in the feature data time-series management module. It ensures the long-term recognition stability of the system under equipment aging and changing operating conditions, and realizes time-series data traceability and efficient interaction, providing complete and effective dataset support for the subsequent optimization of machine vision algorithms. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall structure of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] As attached Figure 1 The image processing-based visual recognition system for preventing belt misalignment of conveyors shown includes an image acquisition and preprocessing module, a dual-feature hierarchical extraction module, a frame interference adaptive rejection module, an illumination interference dynamic correction module, a medium interference multi-dimensional compensation module, a belt feature multi-dimensional fusion and quantization module, a misalignment feature in-depth analysis module, and a feature data time-series management module.

[0019] The image acquisition and preprocessing module: completes camera calibration, performs condition-adaptive acquisition, and executes multi-algorithm collaborative preprocessing; Specifically, it should be noted that the camera calibration operation uses a camera calibration algorithm to calibrate the camera's intrinsic and extrinsic parameters, establishes a dynamic mapping relationship between the image pixel coordinate system and the physical coordinate system, and obtains the total number of pixels in a single frame image through an image segmentation algorithm. Effective pixel ratio of belt area and pixel reference physical size The parameters obtained through the image segmentation algorithm are transmitted to the feature data time-series management module for storage. The working condition adaptation acquisition uses an industrial line scan camera to continuously acquire real-time images of the belt conveyor throughout its operation. The camera's line frequency is adaptively adjusted according to the belt speed to achieve precise matching between the acquired image and the belt's operating state. The multi-algorithm collaborative preprocessing includes using an adaptive median denoising algorithm based on neighborhood grayscale statistics to eliminate salt-and-pepper noise and Gaussian noise in industrial working conditions, using an adaptive contrast enhancement algorithm with the Laplacian operator to improve the visual contrast between the belt and the background, and using a perspective transformation algorithm to correct image perspective distortion caused by camera installation, resulting in a noise-free, distortion-free, and high-contrast preprocessed image. The image noise grayscale variance is extracted during the preprocessing process. Image contrast enhancement coefficient Simultaneously, the preprocessed image, image noise grayscale variance, and image contrast enhancement coefficient are transmitted to the dual-feature hierarchical extraction module and stored in the feature data temporal management module. Specifically, this module, with an industrial linear scan camera and algorithm processing unit as the main execution entities, continuously operates throughout the entire operation of an industrial belt conveyor in a vision detection scenario where the conveyor belt is running. Its core tasks include camera calibration, condition-adaptive acquisition, multi-algorithm collaborative preprocessing, and the extraction and transmission of preprocessing-related parameters and images. The algorithm processing unit first calibrates the camera's intrinsic and extrinsic parameters using a camera calibration algorithm, establishing a dynamic mapping relationship between the image pixel coordinate system and the physical coordinate system. After obtaining the relevant calibration parameters, they are directly transmitted to the feature data time-series management module for storage. The industrial line scan camera continuously acquires real-time images of the entire belt conveyor's operation, while adaptively adjusting the line frequency according to the actual belt speed to ensure that the acquired images accurately match the belt's operating state, avoiding image ghosting and blurring. The algorithm processing unit performs multi-algorithm collaborative preprocessing on the acquired raw images, ultimately obtaining noise-free, distortion-free, and high-contrast preprocessed images. During the above preprocessing process, the algorithm processing unit simultaneously extracts relevant preprocessing parameters and transmits the preprocessed images and extracted parameters together to the dual-feature layered extraction module. At the same time, these preprocessing parameters are also stored in the feature data time-series management module to provide data support for subsequent steps. In addition, this module applies an AI-adaptive image processing algorithm, including an adaptive median denoising sub-algorithm and an adaptive contrast enhancement sub-algorithm. This algorithm is specifically designed for the complex industrial conditions of belt conveyors and is deeply integrated with the actual scenarios of belt conveyor operation, such as salt-and-pepper noise, Gaussian noise, perspective distortion, and low contrast between the belt and the background. It addresses image noise issues caused by dust, equipment vibration, and motor interference in industrial production, as well as perspective distortion caused by camera installation angles and low visual distinction between the belt and the frame background. The algorithm does not require manual setting of fixed processing parameters and can adaptively adjust the processing strategy based on the real-time image characteristics of the belt operation. The input data of this algorithm is the original operating image of the belt conveyor acquired in real time by an industrial line scan camera, and the output data is a noise-free, distortion-free, high-contrast pre-processed image of the belt. The inherent relationship between input and output is that the algorithm intelligently analyzes and identifies the noise type, noise intensity, contrast difference, and distortion degree of the original input image, and then activates the corresponding denoising, enhancement, and distortion correction processing logic accordingly. This ensures that the output pre-processed image accurately preserves the visual features of the belt itself, eliminates meaningless image interference, and lays a precise image foundation for subsequent feature extraction.

[0020] The dual-feature hierarchical extraction module: performs image region division, performs multi-scale feature extraction, performs feature validity verification, and filters and transmits valid feature parameters; Specifically, it should be noted that the image region segmentation is based on an image segmentation algorithm that divides the preprocessed image into two independent regions: the belt body feature region and the environmental interference feature region; the multi-scale feature extraction is used to extract the radius of curvature of one side of the belt's edge using fine-scale edge detection in the belt body feature region. and the vertical distance between the belt edge and the baseline Coarse-scale texture analysis was used to extract the lateral deviation between the real-time centerline and the theoretical centerline of the belt. Change in the slope of the belt centerline and the mean deviation of the direction angle of the belt surface texture For environmental interference feature regions, a feature clustering algorithm is used to extract the number of overlapping pixels between the frame and belt edges. Global illumination gradient value of the image Image medium scattering blur radius Light transmittance in areas blocked by the medium The feature validity verification uses historical benchmark feature parameters in the feature data time-series management module as a reference to obtain the deviation rate between the real-time extracted parameters and the benchmark parameters. The validity of the parameters is determined based on the deviation rate range; parameters with deviation rates within a reasonable range are considered valid, while those exceeding the reasonable range are considered invalid and re-extracted. The filtering and transmission of valid feature parameters involves transmitting all filtered valid feature parameters to the feature data time-series management module for storage. Simultaneously, the belt body feature parameters are transmitted to the frame interference adaptive rejection module, the illumination interference dynamic correction module, and the medium interference multi-dimensional compensation module, respectively. Environmental interference feature parameters are transmitted to their respective interference processing modules. Specifically, this module, executed by the algorithm processing unit, continuously performs real-time operations in the core feature extraction stage of industrial belt conveyor misalignment visual recognition after receiving the pre-processed image and related parameters transmitted from the image acquisition and pre-processing module. The core tasks include image region division, multi-scale feature extraction, feature validity verification, and the filtering and transmission of valid feature parameters. The algorithm processing unit first uses an image segmentation algorithm to accurately divide the received preprocessed image into two independent regions: the belt body feature region and the environmental interference feature region. Then, it performs multi-scale feature extraction for both regions. For the belt body feature region, fine-scale edge detection is used to extract relevant belt edge feature parameters, while coarse-scale texture analysis is used to extract relevant belt centerline and surface texture feature parameters. For the environmental interference feature region, a feature clustering algorithm is used to extract environmental interference feature parameters related to the rack, illumination, and medium. Next, using historical baseline feature parameters stored in the feature data time-series management module as a reference, the deviation rate between the real-time extracted feature parameters and the baseline feature parameters is calculated. The validity of the parameter is determined based on the deviation rate range; if the deviation rate is within a reasonable range, it is considered a valid parameter; otherwise, it is considered an invalid parameter, and feature extraction is performed again. Finally, all the selected valid feature parameters are transmitted to the feature data time-series management module for storage. Simultaneously, the belt body feature parameters are transmitted to the rack interference adaptive elimination module, the illumination interference dynamic correction module, and the medium interference multi-dimensional compensation module, respectively. The environmental interference feature parameters are transmitted to the corresponding interference processing modules, achieving accurate distribution of feature parameters.In addition, the core applications of this module include AI image segmentation algorithms, AI multi-scale feature extraction algorithms, AI feature clustering algorithms, and AI feature validity verification algorithms. These multiple algorithms are deeply integrated with the specific scenario of "mixed belt body features and environmental interference features in the visual recognition of belt conveyor anti-deviation: for the continuous belt, fixed rigid structure of the frame, dynamic and irregular visual feature differences in light / medium interference, and the scenario where early slight deviation is only manifested as fine-scale edge / texture micro-deformation, multi-scale and regional feature extraction logic is designed. At the same time, a verification model is built by combining the benchmark feature data stored in the system's history to avoid invalid feature parameters from entering the subsequent processing stage. The unified input data for all algorithms is the preprocessed image of the conveyor belt output by the image acquisition and preprocessing module. The output data consists of the conveyor belt feature parameters and environmental interference feature parameters after validity verification. The inherent relationship between input and output is as follows: the artificial intelligence image segmentation algorithm first performs intelligent region division on the input preprocessed image, accurately separating the conveyor belt feature region from the environmental interference feature region; the artificial intelligence multi-scale feature extraction algorithm then uses fine-scale detection to extract edge and texture micro-features from the conveyor belt feature region, and uses coarse-scale analysis to extract macro-features of the center line; for the environmental interference feature region, the artificial intelligence feature clustering algorithm extracts similar interference features of the frame, lighting, and medium; finally, the artificial intelligence feature validity verification algorithm intelligently compares and analyzes all the feature parameters extracted in real time with the system's historical benchmark feature parameters, determines the validity of the parameters and filters them, and finally outputs the valid feature parameters, ensuring the accuracy and relevance of feature extraction.

[0021] The rack interference adaptive rejection module generates a pixel-level mask based on effective feature parameters, reconstructs edge feature parameters, and transmits the corrected edge feature parameters. Specifically, it should be noted that the effective feature parameters include the radius of curvature of the single-sided edge of the belt, retrieved from the feature data temporal management module. Vertical distance between belt edge and baseline Number of pixels overlapping between the frame and belt edges At the same time, call the total number of pixels of a single frame image. Image contrast enhancement coefficient The pixel-level mask generation is based on the number of overlapping pixels between the rack and conveyor belt edges to generate a pixel-level mask for the rack occlusion area, and the distorted edge areas covered by the mask are marked and masked; the reconstructed edge feature parameters are reconstructed using a nonlinear formula to obtain the corrected edge curvature radius after removing rack interference. Vertical spacing with correction edge The transmitted and corrected edge feature parameters are used to transmit the corrected edge feature parameters to the feature data time-series management module for storage, and simultaneously to the belt feature multi-dimensional fusion and quantization module. Specifically, this module, executed by the algorithm processing unit, performs real-time operations upon receiving the relevant valid feature parameters transmitted by the dual-feature layered extraction module during the frame interference processing stage of the industrial belt conveyor misalignment visual recognition. The core tasks include calling relevant feature parameters, generating pixel-level masks, reconstructing edge feature parameters, and transmitting the corrected edge feature parameters. The algorithm processing unit first retrieves the corresponding belt edge feature parameters and rack-related environmental interference feature parameters from the feature data temporal management module, and simultaneously retrieves single-frame image-related parameters from the camera calibration stage and contrast enhancement-related parameters from the preprocessing stage. Then, based on the relevant parameters of the overlapping pixels between the rack and belt edges, a pixel-level mask for the rack-occluded area is generated. The distorted edge areas covered by the mask are precisely marked and masked, eliminating invalid feature data in these areas. Next, the effective edge feature parameters after masking are reconstructed using a nonlinear formula to obtain corrected edge feature parameters after removing rack interference, achieving accurate edge feature correction. Finally, the corrected edge feature parameters are transmitted to the feature data temporal management module for storage, and simultaneously transmitted to the belt feature multi-dimensional fusion and quantization module to provide corrected effective data for subsequent feature fusion and quantization. In addition, this module applies an AI pixel-level mask generation algorithm and an AI nonlinear feature reconstruction algorithm. These two algorithms are deeply integrated with the specific scenario of a belt conveyor frame being fixed and obscuring the belt edge. For scenarios where the frame is a fixed structure, the obscuring area is a local edge of the belt, and the degree of obscuring changes dynamically with slight belt swaying, the algorithm does not require manual marking of the obscuring area. It can intelligently generate a mask based on the overlapping features of the frame and belt edges, and intelligently reconstruct and correct parameters based on the normal distribution pattern of belt edge features, thereby achieving accurate removal of frame interference. The algorithm's input data consists of the effective belt edge feature parameters and effective rack interference feature parameters output by the dual-feature hierarchical extraction module, as well as the camera calibration parameters and contrast enhancement parameters from the image acquisition and preprocessing module. The output data consists of the corrected belt edge feature parameters after rack interference has been removed. The intrinsic relationship between the input and output is as follows: the AI ​​pixel-level mask generation algorithm intelligently analyzes the input rack interference feature parameters to accurately identify the position and range of the rack obstructing the belt edge, adaptively generates the corresponding pixel-level mask, and masks the distorted edge areas. The AI ​​nonlinear feature reconstruction algorithm then intelligently analyzes the masked effective belt edge feature parameters, performs nonlinear reconstruction calculations based on the normal distribution law of belt edge features, and finally outputs corrected feature parameters that can truly reflect the state of the belt edge.

[0022] The dynamic correction module for illumination interference performs local brightness adaptive equalization based on relevant feature parameters, dynamically corrects the centerline feature parameters, and transmits the corrected centerline feature parameters. Specifically, it should be noted that the relevant feature parameters include the lateral deviation value of the belt centerline retrieved from the feature data time-series management module. Change in the slope of the belt centerline Global illumination gradient value of the image Simultaneously call the pixel reference physical size and image noise grayscale variance The local brightness adaptive equalization is based on the global illumination gradient value of the image to perform local brightness adaptive equalization on the belt centerline region, eliminating the centerline feature blurring problem caused by uneven illumination; the dynamic correction of centerline feature parameters uses the reference illumination gradient value in the feature data temporal management module. For reference, the characteristic parameters of the equalized centerline are dynamically corrected using a nonlinear formula to obtain the lateral deviation value of the corrected centerline after eliminating illumination interference. Change in slope of the correction centerline The corrected centerline feature parameters are transmitted to the feature data time-series management module for storage, and simultaneously to the belt feature multi-dimensional fusion quantization module. Specifically, this module, executed by the algorithm processing unit, performs real-time operations upon receiving relevant valid feature parameters from the dual-feature layered extraction module during the illumination interference processing stage of visual recognition for industrial belt conveyor misalignment. The core tasks include retrieving relevant feature parameters, adaptive equalization of local brightness, dynamic correction of centerline feature parameters, and transmission of the corrected centerline feature parameters. The algorithm processing unit first retrieves the corresponding belt centerline feature parameters and illumination-related environmental interference feature parameters from the feature data temporal management module, and simultaneously retrieves the pixel reference physical size parameters from the camera calibration stage and the image noise-related parameters from the preprocessing stage. Then, based on the relevant parameters of the global illumination gradient of the image, it specifically performs a local brightness adaptive equalization operation on the belt centerline region to eliminate the blurring of the centerline features caused by uneven illumination and restore clear centerline features. Next, using the reference illumination gradient parameters stored in the feature data temporal management module as a reference, it performs dynamic correction calculations on the equalized centerline feature parameters using a nonlinear formula to obtain the corrected centerline feature parameters after eliminating illumination interference. Finally, the corrected centerline feature parameters are transmitted to the feature data temporal management module for storage, and simultaneously transmitted to the belt feature multi-dimensional fusion and quantization module to provide corrected effective data for subsequent feature fusion and quantization. In addition, this module applies an artificial intelligence-based local brightness adaptive equalization algorithm and an artificial intelligence-based dynamic feature correction algorithm. These two algorithms are deeply integrated with the specific scenario of uneven lighting and random changes in strong / weak light on the belt conveyor: In view of the scenario where the changes in lighting are mostly concentrated in the center line area of ​​the belt and changes in lighting gradient can cause blurring or displacement of the center line features, the algorithm abandons the global brightness adjustment method and only performs targeted brightness equalization on the center line area of ​​the belt. At the same time, it combines the correlation between lighting changes and center line feature deviations for dynamic correction to avoid distortion of center line features caused by lighting interference.The algorithm's input data consists of the effective belt centerline feature parameters and effective illumination interference feature parameters output by the dual-feature hierarchical extraction module, as well as the camera calibration parameters and image noise parameters from the image acquisition and preprocessing module. The output data consists of the corrected belt centerline feature parameters after eliminating illumination interference. The intrinsic relationship between the input and output is as follows: the AI-powered local brightness adaptive equalization algorithm intelligently analyzes the input illumination interference feature parameters, identifies the degree of illumination gradient change in the belt centerline region, and performs targeted brightness adaptive equalization processing on this region to eliminate the blurring of the centerline features caused by uneven illumination. The AI-powered dynamic feature correction algorithm then uses the system's historical baseline illumination parameters as a reference to intelligently and dynamically correct the equalized centerline feature parameters, adjusting the correction strategy based on the correlation between illumination changes and centerline feature deviations, and finally outputting corrected feature parameters that truly reflect the state of the belt centerline.

[0023] The multi-dimensional compensation module for media interference performs image deblurring and reconstruction based on relevant feature parameters, restores texture feature parameters, and transmits the corrected texture feature parameters. Specifically, it should be noted that the relevant feature parameters include the average deviation of the direction angle of the belt surface texture, retrieved from the feature data time-series management module. Image medium scattering blur radius Light transmittance in areas blocked by the medium Simultaneously call the image noise grayscale variance Effective pixel ratio of belt area The image deblurring and reconstruction is performed using a Gaussian deblurring algorithm based on the image medium scattering blur radius to reconstruct the belt texture region, eliminating the texture blurring problem caused by medium scattering. The restored texture feature parameters are then multi-dimensionally restored using a nonlinear formula to obtain the mean value of the corrected texture direction angle deviation after eliminating medium interference. The transmitted and corrected texture feature parameters are used to transmit the corrected texture feature parameters to the feature data temporal management module for storage, and simultaneously to the belt feature multi-dimensional fusion and quantization module. Specifically, this module, executed by the algorithm processing unit, performs real-time operations upon receiving the relevant valid feature parameters transmitted by the dual-feature layered extraction module in the media interference processing stage of visual recognition for industrial belt conveyor misalignment. The core tasks include calling relevant feature parameters, image deblurring and reconstruction, texture feature parameter restoration, and transmitting the corrected texture feature parameters. The algorithm processing unit first retrieves the corresponding belt surface texture feature parameters and medium-related environmental interference feature parameters from the feature data temporal management module. It also retrieves image noise-related parameters from the preprocessing stage and the effective pixel ratio parameters of the belt region from the camera calibration stage. Then, based on the image medium scattering blur parameters, a Gaussian deblurring algorithm is used to reconstruct the belt texture region, eliminating texture blur caused by scattering from media such as dust and water mist. Next, a nonlinear formula is used to perform multi-dimensional restoration calculations on the compensated texture feature parameters to obtain the corrected texture feature parameters after eliminating medium interference. Finally, the corrected texture feature parameters are transmitted to the feature data temporal management module for storage and simultaneously transmitted to the belt feature multi-dimensional fusion and quantization module to provide corrected effective data for subsequent feature fusion and quantization. In addition, this module applies the AI ​​Gaussian deblurring algorithm and the AI ​​texture feature restoration algorithm to the specific scenario of dust, water mist and other media scattering and occlusion on the belt conveyor: In response to the scenario where dust and water mist will cause the belt texture area image to be blurred, thus causing the texture feature to be distorted, the algorithm first performs intelligent deblurring and reconstruction of the texture area, and finally combines the original rules of the texture feature to perform multi-dimensional restoration, so as to achieve the comprehensive elimination of media interference. The algorithm's input data consists of the effective belt texture feature parameters and effective medium interference feature parameters output by the dual-feature hierarchical extraction module, as well as the image noise parameters from the image acquisition and preprocessing module and the effective pixel parameters of the belt region calibrated by the camera. The output data consists of the corrected belt texture feature parameters after eliminating medium interference. The intrinsic relationship between the input and output is as follows: the artificial intelligence Gaussian deblurring algorithm intelligently analyzes the input medium interference feature parameters to identify the degree of blurring in the belt texture region, and uses Gaussian deblurring logic to intelligently reconstruct the texture region image, eliminating the blurring problem caused by medium scattering. Finally, the artificial intelligence texture feature restoration algorithm performs intelligent multi-dimensional restoration calculation on the compensated texture feature parameters, and corrects the distortion features by combining the normal distribution law of belt texture features, ultimately outputting corrected feature parameters that can truly reflect the state of the belt texture.

[0024] The belt feature multi-dimensional fusion quantization module: calls the corrected feature parameters to obtain the mean radius of curvature, constructs multi-dimensional quantization indicators, and transmits the fused quantization indicator parameters; Specifically, it should be noted that the invocation of the corrected feature parameters includes invoking the corrected edge curvature radius from the feature data temporal management module. Correcting the vertical spacing of the edges 1. Correction centerline lateral deviation value , Change in slope of the correction centerline and the mean value of corrected texture orientation angle deviation Simultaneously, the number of segments representing the belt edge in a single frame image is retrieved; the average radius of curvature is determined based on the number of segments representing the belt edge in a single frame image. Obtain the mean radius of curvature of the edges of the corresponding segments. The construction of multi-dimensional quantitative indicators includes the construction of edge deformation quantitative indicators based on the feature fusion concept of machine vision. Centerline offset quantitative indicator and texture distortion metrics This system achieves in-depth quantification of different dimensions of belt misalignment characteristics. The fused and quantized index parameters are used to transmit the calculated multi-dimensional quantized indicators to the feature data time-series management module for storage, and simultaneously to the misalignment feature in-depth analysis module. Specifically, this module, executed by the algorithm processing unit, operates in real-time upon receiving the corrected feature parameters from each interference processing module in the core stage of visual recognition of industrial belt conveyor misalignment. The core tasks include calling the corrected feature parameters, calculating the mean radius of curvature, constructing multi-dimensional quantized indicators, and transmitting the fused and quantized index parameters. The algorithm processing unit first retrieves all belt body feature parameters corrected by the frame, illumination, and media interference processing modules from the feature data temporal management module. Simultaneously, it retrieves parameters related to the number of belt edge segments in a single-frame image stored in the feature data temporal management module. Then, based on these parameters, it statistically calculates the corrected curvature radii of multiple edge segments to obtain the average edge curvature radius for each segment. Next, based on the feature fusion concept of machine vision, it constructs edge deformation quantification indicators, centerline offset quantification indicators, and texture distortion quantification indicators by combining corrected feature parameters from different dimensions. Through multi-dimensional quantitative analysis, it accurately characterizes the belt misalignment in different dimensions. Finally, it transmits the calculated multi-dimensional quantification indicators to the feature data temporal management module for storage and simultaneously to the misalignment feature depth analysis module, providing a quantitative basis for subsequent depth determination of misalignment features. In addition, this module applies artificial intelligence feature fusion algorithms and artificial intelligence multi-dimensional quantitative indicators to build a model. The two are deeply integrated with the specific scenarios of multi-dimensional feature manifestations of belt conveyor misalignment faults: For the scenario where belt misalignment is simultaneously manifested as edge deformation, centerline offset, and texture distortion, and the three types of features are inherently related and a single feature cannot fully represent the misalignment state, the model learns the correlation between the three types of features and misalignment faults, and integrates the independent corrected feature parameters into multi-dimensional quantitative indicators to achieve in-depth quantification of misalignment features. The input data for the algorithm / model consists of all corrected belt body feature parameters output by the rack, lighting, and media interference processing modules, as well as the belt edge segmentation parameters stored in the system for a single frame image. The output data consists of edge deformation quantification index, centerline offset quantification index, and texture distortion quantification index. The intrinsic relationship between the input and output is as follows: the artificial intelligence feature fusion algorithm first performs intelligent correlation analysis on all corrected belt body feature parameters to uncover the intrinsic relationship between the three types of features: edge, centerline, and texture, achieving effective fusion of single features. The artificial intelligence multi-dimensional quantification index construction model then transforms the fused feature data into quantification indicators that can accurately characterize the deviation features in each dimension based on the actual scene patterns of belt deviation. The index values ​​are positively correlated with the degree of deviation in the corresponding dimension, and finally, three independent multi-dimensional quantification indicators with clear physical meaning are output.

[0025] Deep analysis module for deviation characteristics: Constructs a comprehensive deviation quantitative index based on multi-dimensional quantitative indicators and transmits the comprehensive deviation quantitative index; Specifically, it should be noted that the construction of the comprehensive belt misalignment quantification index is achieved by fusing edge deformation quantification index, centerline offset quantification index, and texture distortion quantification index using a nonlinear composite formula. This module enables full-domain quantification of belt misalignment. The comprehensive misalignment quantification index is transmitted to the feature data time-series management module for storage, providing a quantitative basis for graded early warning and status monitoring of belt conveyor misalignment faults. Specifically, this module, executed by the algorithm processing unit, performs real-time operations upon receiving the multi-dimensional quantification index transmitted from the belt feature multi-dimensional fusion quantification module during the final judgment stage of misalignment feature visual recognition for industrial belt conveyors. Its core functions include calling the multi-dimensional quantification index, constructing the comprehensive misalignment quantification index, and transmitting the comprehensive misalignment quantification index. The algorithm processing unit first retrieves edge deformation quantification indicators, centerline offset quantification indicators, and texture distortion quantification indicators from the feature data time-series management module, while also retrieving parameters related to the number of belt edge segments and the effective pixel ratio of the belt area. Then, using a nonlinear composite formula, it fuses these three different quantification indicators to construct a comprehensive belt misalignment quantification index. This index enables a comprehensive quantitative assessment of the belt misalignment degree, transforming the degree of misalignment from a qualitative description to a quantitative representation. Finally, the calculated comprehensive misalignment quantification index is transmitted to the feature data time-series management module for storage. This index will serve as the core quantitative basis for belt conveyor misalignment fault classification early warning and operational status monitoring, providing data support for subsequent equipment maintenance. In addition, this module applies an artificial intelligence multi-indicator fusion analysis model, which is deeply integrated with the specific scenario of quantitative assessment and graded early warning of belt conveyor deviation: for the scenario in industrial field where belt deviation detection not only needs to determine whether there is deviation, but also needs to accurately quantify the degree of deviation to support graded early warning and equipment operation and maintenance, the model abandons the simple threshold comparison method, and achieves full-domain quantitative assessment of deviation by learning the correspondence between multi-dimensional quantitative indicators and actual deviation degree. The model's input data consists of edge deformation quantification indicators, centerline offset quantification indicators, and texture distortion quantification indicators output by the belt feature multi-dimensional fusion quantification module, as well as belt edge segmentation parameters and effective pixel parameters of the belt region stored in the system. The output data is the belt comprehensive deviation quantification index. The intrinsic relationship between the input and output is that the model intelligently fuses and analyzes the three input multi-dimensional quantification indicators, combines them with scene parameters such as belt width and effective pixel ratio, and explores the intrinsic correlation between each quantification indicator and the actual deviation degree. Through nonlinear fusion calculation, the multi-dimensional indicators are transformed into a comprehensive quantification index that can fully characterize the overall deviation degree of the belt. The index value is positively correlated with the belt deviation degree. The final output index can provide accurate quantitative basis for graded early warning and status monitoring of deviation faults.

[0026] Feature data time-series management module: provides data retrieval services, performs time-series data storage, carries out dynamic benchmark updates, and performs invalid data cleaning.

[0027] Specifically, the data retrieval service provides a fast access service for historical benchmark data and real-time acquired data to all other modules of the system, meeting the parameter usage needs of each module; the execution of time-series data storage uses a high-speed time-series database to store all parameters of each module in time series, including camera calibration parameters, preprocessing parameters, original feature parameters, corrected feature parameters, fusion quantization indicators, and comprehensive deviation quantization index, enabling data retrieval and traceability by timestamp; the dynamic benchmark update performs statistical analysis on historical feature parameters at fixed time intervals, calculates new feature benchmark parameters after removing outliers, and synchronously updates them to the parameter reference library of each module; the invalid data cleaning uses an outlier detection algorithm to clean invalid data from real-time acquired and calculated parameters, automatically removing invalid data caused by hardware failures or extreme operating conditions. In detail, this module, with a high-speed time-series database and a data processing unit as its main execution components, provides full-process data support for the visual recognition of industrial belt conveyor misalignment. It operates continuously throughout the entire visual recognition system after startup, primarily providing data retrieval services, storing time-series data, updating dynamic benchmarks, cleaning invalid data, and supporting data interaction across all modules. The high-speed time-series database provides rapid access to historical benchmark data and real-time acquired data for all other modules in the system, accurately distributing relevant data according to the operational needs of each module to meet their parameter usage requirements. The high-speed time-series database uses a time-series storage method to uniformly store all parameters generated by each module, covering camera calibration parameters, preprocessing parameters, raw feature parameters, corrected feature parameters, fusion quantization indicators, and comprehensive misalignment quantization indexes. This enables time-stamped retrieval and traceability of all data, facilitating subsequent historical data review and analysis. The data processing unit performs comprehensive statistical analysis of the stored historical feature parameters at fixed time intervals, eliminating data caused by hardware failures or extreme operating conditions. After an outlier is detected, new feature reference parameters are calculated and synchronously updated to the parameter reference library of each module to solve the feature reference drift problem caused by equipment aging and changes in operating conditions. The data processing unit uses an outlier detection algorithm to clean up invalid data from the parameters collected and calculated in real time by each module, automatically removing invalid data to avoid interference with subsequent feature processing and quantitative analysis, and ensuring the validity and accuracy of system data. The high-speed time-series database and the data processing unit work together to ensure real-time one-way data flow and two-way interaction between the modules of the system, so that the operation of each module forms a data closed loop, providing comprehensive data analysis and interaction support for the stable and efficient operation of the entire visual recognition system.In addition, this module applies artificial intelligence time-series data management algorithms, artificial intelligence dynamic benchmark update models, and artificial intelligence outlier detection algorithms. These multiple algorithms / models are deeply integrated with the specific scenarios of long-term continuous operation of belt conveyors, dynamic changes in operating conditions, and gradual aging of equipment. For the characteristics of long-term system operation, such as belt wear, frame loosening, and changes in lighting environment leading to drift of feature benchmark parameters, and hardware failures and extreme operating conditions generating invalid data, the module achieves intelligent storage and retrieval of parameters through time-series data management, adapts to changes in operating conditions and equipment through dynamic benchmark update models, and removes invalid data through outlier detection algorithms, ensuring the recognition stability of the system during long-term operation. The unified input data for each algorithm / model is the full set of parameters generated by all other modules in the system, including camera calibration parameters, preprocessing parameters, raw feature parameters, corrected feature parameters, fusion quantization index, and comprehensive deviation quantization index. The output data consists of cleaned, valid time-series data and updated feature benchmark parameters. The inherent relationship between input and output is as follows: the AI ​​time-series data management algorithm intelligently stores and classifies the full set of input parameters in a time-series manner, enabling rapid retrieval and traceability of data by timestamp; the AI ​​outlier detection algorithm intelligently analyzes the real-time acquisition and calculation parameters of the input, identifying and eliminating invalid data caused by hardware failures or extreme operating conditions, ensuring the validity of the data; the AI ​​dynamic benchmark update model intelligently statistically analyzes the historical feature parameters of the input at fixed time intervals, eliminates outliers, and automatically calculates and updates the feature benchmark parameters based on the changing patterns of operating conditions and equipment, synchronizing them to the parameter reference library of each module to solve the feature benchmark drift problem and ensure the recognition accuracy of the system in long-term operation.

[0028] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A vision recognition system for preventing belt misalignment in belt conveyors based on image processing, characterized in that, include: Image acquisition and preprocessing module: Completes camera calibration, performs condition-adaptive acquisition, and executes multi-algorithm collaborative preprocessing; Dual-feature hierarchical extraction module: performs image region segmentation, multi-scale feature extraction, verifies feature validity, and filters and transmits valid feature parameters; Rack interference adaptive culling module: Generates pixel-level masks based on effective feature parameters, reconstructs edge feature parameters, and transmits the corrected edge feature parameters; Illumination interference dynamic correction module: performs local brightness adaptive equalization based on relevant characteristic parameters, dynamically corrects centerline characteristic parameters, and transmits the corrected centerline characteristic parameters; Multi-dimensional compensation module for media interference: Based on relevant feature parameters, it performs image deblurring and reconstruction, restores texture feature parameters, and transmits the corrected texture feature parameters; The belt feature multi-dimensional fusion quantization module: calls the corrected feature parameters to obtain the mean radius of curvature, constructs multi-dimensional quantization indicators, and transmits the fused quantization indicator parameters; Deep analysis module for deviation characteristics: Constructs a comprehensive deviation quantitative index based on multi-dimensional quantitative indicators and transmits the comprehensive deviation quantitative index; Feature data time-series management module: provides data retrieval services, performs time-series data storage, carries out dynamic benchmark updates, and performs invalid data cleaning.

2. The image processing-based visual recognition system for preventing belt misalignment in belt conveyors according to claim 1, characterized in that: The camera calibration operation employs a camera calibration algorithm to calibrate the camera's intrinsic and extrinsic parameters, establishing a dynamic mapping relationship between the image pixel coordinate system and the physical coordinate system. The total number of pixels in a single frame is then obtained through an image segmentation algorithm. Effective pixel ratio of belt area and pixel reference physical size The parameters obtained through the image segmentation algorithm are transmitted to the feature data time-series management module for storage; the working condition adaptation acquisition uses an industrial line array camera to continuously acquire real-time images of the belt conveyor throughout its operation, and adaptively adjusts the camera line frequency according to the belt speed. The multi-algorithm collaborative preprocessing includes using an adaptive median denoising algorithm based on neighborhood gray-level statistics to eliminate salt-and-pepper noise and Gaussian noise in industrial conditions; using an adaptive contrast enhancement algorithm with the Laplacian operator to improve the visual contrast between the belt and the background; and using a perspective transformation algorithm to correct image perspective distortion caused by camera mounting, resulting in a noise-free, distortion-free, and high-contrast preprocessed image. The image noise gray-level variance is also extracted during the preprocessing process. Image contrast enhancement coefficient Simultaneously, the preprocessed image, image noise grayscale variance, and image contrast enhancement coefficient are transmitted to the dual-feature layer extraction module and stored in the feature data time-series management module.

3. The image processing-based visual recognition system for preventing belt misalignment in belt conveyors according to claim 1, characterized in that: The image region segmentation is based on an image segmentation algorithm, dividing the preprocessed image into two independent regions: the belt body feature region and the environmental interference feature region. The multi-scale feature extraction is used to extract the radius of curvature of one side of the belt's edge from the belt body feature region using fine-scale edge detection. and the vertical distance between the belt edge and the baseline Coarse-scale texture analysis was used to extract the lateral deviation between the real-time centerline and the theoretical centerline of the belt. Change in the slope of the belt centerline and the mean deviation of the direction angle of the belt surface texture For environmental interference feature regions, a feature clustering algorithm is used to extract the number of overlapping pixels between the frame and belt edges. Global illumination gradient value of the image Image medium scattering blur radius Light transmittance in areas blocked by the medium ; The feature validity verification uses historical benchmark feature parameters in the feature data time-series management module as a reference to obtain the deviation rate between the real-time extracted parameters and the benchmark parameters. The validity of the parameters is determined based on the deviation rate range. If the deviation rate is within a reasonable range, the parameters are considered valid; if the deviation rate exceeds the reasonable range, the parameters are considered invalid and are re-extracted. The filtering and transmission of valid feature parameters is used to transmit all the filtered valid feature parameters to the feature data time-series management module for storage. At the same time, the belt body feature parameters are transmitted to the rack interference adaptive rejection module, the light interference dynamic correction module, and the medium interference multi-dimensional compensation module, respectively. The environmental interference feature parameters are transmitted to the corresponding interference processing modules.

4. The image processing-based visual recognition system for preventing belt misalignment in belt conveyors according to claim 1, characterized in that: The effective feature parameters include the radius of curvature of the single-sided edge of the belt, retrieved from the feature data temporal management module. Vertical distance between belt edge and baseline Number of pixels overlapping between the frame and belt edges At the same time, call the total number of pixels of a single frame image. Image contrast enhancement coefficient The pixel-level mask generation is based on the number of overlapping pixels between the rack and conveyor belt edges to generate a pixel-level mask for the rack occlusion area, and the distorted edge areas covered by the mask are marked and masked; the reconstructed edge feature parameters are reconstructed using a nonlinear formula to obtain the corrected edge curvature radius after removing rack interference. Vertical spacing with correction edge The transmitted and corrected edge feature parameters are used to transmit the corrected edge feature parameters to the feature data time-series management module for storage, and at the same time to the belt feature multi-dimensional fusion quantization module.

5. The image processing-based visual recognition system for preventing belt misalignment in belt conveyors according to claim 1, characterized in that: The relevant feature parameters include the lateral deviation value of the belt centerline retrieved from the feature data time-series management module. Change in the slope of the belt centerline Global illumination gradient value of the image Simultaneously call the pixel reference physical size and image noise grayscale variance The local brightness adaptive equalization is based on the global illumination gradient value of the image to perform local brightness adaptive equalization on the belt centerline region, eliminating the centerline feature blurring problem caused by uneven illumination; the dynamic correction of centerline feature parameters uses the reference illumination gradient value in the feature data temporal management module. For reference, the characteristic parameters of the equalized centerline are dynamically corrected using a nonlinear formula to obtain the lateral deviation value of the corrected centerline after eliminating illumination interference. Change in slope of the correction centerline The corrected centerline feature parameters are used to transmit the corrected centerline feature parameters to the feature data time-series management module for storage, and at the same time to the belt feature multi-dimensional fusion quantization module.

6. The image processing-based visual recognition system for preventing belt misalignment in belt conveyors according to claim 1, characterized in that: The relevant feature parameters include the mean value of the direction angle deviation of the belt surface texture retrieved from the feature data time-series management module. Image medium scattering blur radius Light transmittance in areas blocked by the medium Simultaneously call the image noise grayscale variance Effective pixel ratio of belt area The image deblurring and reconstruction is performed using a Gaussian deblurring algorithm based on the image medium scattering blur radius to reconstruct the belt texture region, eliminating the texture blurring problem caused by medium scattering. The restored texture feature parameters are then multi-dimensionally restored using a nonlinear formula to obtain the mean value of the corrected texture direction angle deviation after eliminating medium interference. The transmitted and corrected texture feature parameters are used to transmit the corrected texture feature parameters to the feature data temporal management module for storage, and at the same time to the belt feature multi-dimensional fusion quantization module.

7. The image processing-based visual recognition system for preventing belt misalignment in belt conveyors according to claim 1, characterized in that: The process of calling the corrected feature parameters includes calling the corrected edge curvature radius from the feature data temporal management module. Correcting the vertical spacing of the edges 1. Correction centerline lateral deviation value , Change in slope of the correction centerline and the mean value of corrected texture orientation angle deviation Simultaneously, the number of segments representing the belt edge in a single frame image is retrieved; the average radius of curvature is determined based on the number of segments representing the belt edge in a single frame image. Obtain the mean radius of curvature of the edges of the corresponding segments. The construction of multi-dimensional quantitative indicators includes the construction of edge deformation quantitative indicators based on the feature fusion concept of machine vision. Centerline offset quantitative indicator and texture distortion metrics The transmitted fused and quantized index parameters are used to transmit the calculated multi-dimensional quantized indexes to the feature data time-series management module for storage, and at the same time to the deviation feature depth analysis module.

8. The image processing-based visual recognition system for preventing belt misalignment in belt conveyors according to claim 1, characterized in that: The comprehensive belt misalignment quantification index is constructed by fusing edge deformation quantification index, centerline offset quantification index, and texture torsion quantification index using a nonlinear composite formula. The comprehensive deviation quantification index is used to transmit the calculated comprehensive deviation quantification index to the feature data time-series management module for storage, providing a quantitative basis for graded early warning and status monitoring of belt conveyor deviation faults.

9. The image processing-based visual recognition system for preventing belt misalignment in belt conveyors according to claim 1, characterized in that: The data retrieval service provides a fast access service for historical benchmark data and real-time acquired data to all other modules of the system, meeting the parameter usage needs of each module. The execution of time-series data storage uses a high-speed time-series database to store all parameters of each module in time series, including camera calibration parameters, preprocessing parameters, original feature parameters, corrected feature parameters, fusion quantization indicators, and comprehensive deviation quantization index, enabling data retrieval and traceability by timestamp. The dynamic benchmark update performs statistical analysis on historical feature parameters at fixed time intervals, calculates new feature benchmark parameters after removing outliers, and synchronously updates them to the parameter reference library of each module. The invalid data cleaning uses an outlier detection algorithm to clean invalid data from real-time acquired and calculated parameters, automatically removing invalid data caused by hardware failures and extreme operating conditions.

Images