Railway supervision based on ai visual identification construction material quality acceptance method
By using multispectral imaging and non-uniform illumination correction technology, combined with multidimensional feature extraction and adaptive convolutional neural networks, the problems of detection accuracy and environmental adaptability in the quality acceptance of railway construction materials have been solved, achieving efficient, traceable, and intelligent acceptance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 王鹏
- Filing Date
- 2026-01-20
- Publication Date
- 2026-06-09
AI Technical Summary
Existing railway construction material quality acceptance technologies are inadequate in terms of testing accuracy, environmental adaptability, defect quantification, and data traceability, and cannot meet the modern railway engineering's demand for efficient, objective, accurate, and traceable intelligent acceptance.
Four-dimensional data is acquired using multispectral imaging equipment. Combined with non-uniform illumination correction, multi-dimensional feature extraction, adaptive convolutional neural network, and three-dimensional morphology reconstruction, a material texture-defect mapping model is constructed to generate a digital acceptance report and achieve tamper-proof traceability through blockchain hash value.
It has enabled comprehensive, automated, and high-precision acceptance of railway construction materials in complex on-site environments, improved testing accuracy, environmental adaptability, and data management level, and formed a standardized and traceable digital reporting system.
Smart Images
Figure CN122176495A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent engineering supervision technology, specifically to a method for quality acceptance of construction materials based on AI visual recognition in railway supervision. Background Technology
[0002] The quality of railway construction materials directly affects the safety and service life of railway lines. Material acceptance is the primary step in ensuring project quality during construction supervision. Traditional railway construction material acceptance relies mainly on manual visual inspection and simple instrument measurements, which has significant limitations. First, manual inspection is greatly affected by the inspector's experience, fatigue level, and subjective judgment; the results of judgments from different personnel may vary significantly, lacking a unified standard. Second, visual inspection struggles to detect minute defects on the material surface or abnormal internal textures, especially for railway-specific high-strength steel, sleepers, and ballast; issues such as minute cracks, corrosion, color differences, and uneven textures are easily overlooked. Furthermore, the complex environment of construction sites and unstable lighting conditions significantly reduce the accuracy of manual inspection under strong light, shadow, or nighttime conditions.
[0003] With the expansion of railway construction and the increasing demand for intelligent supervision, the industry has begun to explore the introduction of machine vision technology to assist in material acceptance. However, existing machine vision-based inspection methods mostly employ single-optical-band imaging, capable of capturing two-dimensional planar information only under fixed lighting conditions. This fails to effectively address the varying lighting environments and reflective properties of material surfaces at construction sites. While some methods incorporate multi-view imaging, they lack accurate reconstruction of the three-dimensional morphology of the material surface, resulting in a loss of spatial distribution information on defects and an inability to quantify the impact of defects on material performance. Furthermore, most existing visual inspection algorithms are based on fixed convolutional neural network structures and are not optimized for the unique textures and defect patterns of railway materials, resulting in limited classification accuracy and generalization ability.
[0004] In defect assessment, existing technologies often employ simple threshold judgments or linear scoring, failing to comprehensively consider defect type, spatial location, illumination variations, and stability over time, resulting in scoring results that do not accurately reflect material quality. Furthermore, acceptance reports are mostly generated manually, lacking a unified data structure and tamper-proof mechanisms, which hinders subsequent quality traceability and data analysis.
[0005] Furthermore, existing methods lack multispectral information fusion in the data acquisition stage, failing to capture the differences in material response to different spectra at different wavelengths. This is particularly important when identifying certain types of defects (such as surface oxidation, oil stains, coating aging, etc.). In the acceptance of construction materials, different materials have different absorption and reflection characteristics to the spectrum, and single visible light imaging can easily miss key defect information.
[0006] In summary, existing railway construction material quality acceptance technologies are insufficient in terms of testing accuracy, environmental adaptability, defect quantification, and data traceability, failing to meet the demands of modern railway engineering for efficient, objective, accurate, and traceable intelligent acceptance. Therefore, it is necessary to propose a new method that, in complex field environments, utilizes multispectral imaging, 3D topography reconstruction, time-series analysis, and a proprietary AI visual recognition model to achieve comprehensive, automated, and high-precision acceptance of railway construction material quality, and to establish a standardized and traceable digital reporting system. Summary of the Invention
[0007] The purpose of this invention is to provide a method for quality acceptance of construction materials based on AI visual recognition in railway supervision.
[0008] To achieve the above objectives, the present invention provides the following technical solution: a construction material quality acceptance method based on AI visual recognition in railway supervision. Firstly, a multispectral imaging device is deployed at the construction site. This device can simultaneously acquire multispectral image sequences of the material surface at different wavelengths. ,in Represents pixel coordinates, For wavelength, The acquisition time is used to form four-dimensional data containing spatial, spectral, and temporal information. Subsequently, non-uniform illumination correction is performed on the acquired multispectral image sequence to obtain the corrected image. This eliminates brightness differences caused by variations in the position and intensity of on-site light sources. Next, a multi-dimensional feature extraction module extracts material texture, color distribution, and edge features from the corrected image to construct a high-dimensional feature vector. This provides a sufficient information foundation for subsequent defect identification. Based on this, a material texture-defect mapping model is constructed. Mapping feature vectors to a set of defect categories This enables semantic interpretation of features. Then, an adaptive convolutional neural network branch structure is employed to fuse local and global features, outputting the probability distribution for each defect. Simultaneously, the results of the three-dimensional morphology reconstruction of the material surface were combined. Calculate the spatial impact factor of defects This quantifies the severity of defects in three-dimensional space. Furthermore, a quality scoring function based on probability distribution is introduced. ,in This represents the defect weighting coefficient, reflecting the degree of influence of different defects on material properties. Then, based on the dynamic threshold... The system determines whether materials are qualified and generates a digital acceptance report containing images, features, scores, and judgment results, which is then uploaded to the supervision database, achieving fully traceable and automated acceptance. It covers the entire process from data acquisition to the final acceptance report, forming a complete closed-loop quality control system. Multispectral imaging and time-series acquisition ensure data richness and on-site adaptability; non-uniform illumination correction improves the stability of feature extraction; multi-dimensional feature fusion enhances the comprehensiveness of defect identification; texture-defect mapping models realize the transformation from features to semantics; adaptive convolutional networks consider both local and global information; 3D morphology reconstruction and spatial influence factors quantify the actual hazard of defects; probability scoring and dynamic threshold judgment improve the scientific rigor and flexibility of qualification determination; and digital reporting and database storage ensure the traceability of the acceptance process and the efficiency of data management.
[0009] Furthermore, the wavelength range of the multispectral imaging device was limited to 400nm to 1000nm, covering the visible and near-infrared bands. This allows it to capture the differences in reflection and absorption characteristics of materials under different spectra, aiding in the identification of defects invisible to the naked eye. At least twelve images were acquired within each sampling period T, ensuring the continuity and statistical reliability of the time-series data, which can be used to analyze minute fluctuations in the material surface over time. Illumination angle... The algorithm can switch between preset modes, such as different orientations and elevation angles, to obtain light-insensitive features in subsequent processing, reducing the impact of changes in on-site lighting conditions on the detection results. By expanding the spectral range and increasing the number of sampling frames, the integrity and redundancy of material surface information are significantly improved, making subsequent feature extraction and defect identification more robust. The preset mode of changing the illumination angle effectively mitigates the interference caused by uneven on-site lighting, improves the algorithm's adaptability to environmental changes, and ensures reliable detection data can be obtained under different construction periods and weather conditions.
[0010] Furthermore, the implementation formula for non-uniform illumination correction is specified. ,in This represents the average brightness of the entire image at the same wavelength, used to reflect the overall level of current illumination. The value is an extremely small positive number to prevent calculation errors caused by a zero denominator. This correction method normalizes the pixel brightness of each frame of the image to a unified lighting reference, thereby eliminating brightness differences caused by variations in the position or intensity of the light source, ensuring consistency in subsequent feature extraction under different times and conditions. The formula is simple and efficient, eliminating the effects of uneven lighting while preserving material surface details, guaranteeing the stability and comparability of feature extraction. Compared to complex lighting modeling methods, this scheme has low computational cost, making it suitable for deployment in real-time field systems, and effectively improving the accuracy of defect identification.
[0011] Furthermore, including improvements The operator is used to extract spatiotemporal texture patterns, capturing dynamic texture changes of materials in both time and space dimensions; the mean is calculated in the CIELAB color space. With covariance matrix This reflects the distribution and correlation of material colors; edge density is obtained using Canny multi-scale edge detection. This characterizes the complexity of the material's surface contours. Ultimately, these features are combined into a high-dimensional feature vector. This provides comprehensive information for subsequent defect classification. Multi-dimensional feature fusion characterizes material appearance attributes from different perspectives; texture, color, and edge information complement each other, enabling a more comprehensive representation of the material surface state and improving the robustness and accuracy of defect detection. Improved... The operator enhances the ability to capture spatiotemporal textures. CIELAB spatial features are insensitive to changes in illumination, while multi-scale edge detection can adapt to defect contours at different scales.
[0012] Furthermore, the training method for the material texture-defect mapping model M is limited to a combination of unsupervised clustering and supervised fine-tuning, with the loss function including classification cross-entropy loss. Regression smoothing loss and triplet loss The weights are respectively , , Unsupervised clustering is used to initially explore the natural distribution of the feature space, while supervised fine-tuning performs precise adjustments on the labeled data. Triple loss is used to increase the distance between different defect categories, improving the clarity of classification boundaries. This training strategy balances the needs of data exploration and accurate classification. The design of the composite loss function enables the model to accurately classify known defects while maintaining good inter-class separation, thereby improving the generalization ability to new samples and reducing the risk of misclassification.
[0013] Furthermore, the branching structure of the adaptive convolutional neural network is described in detail, including the use of 3×3 convolutional kernels to extract fine-grained features in local branches, the use of dilated convolutions to expand the receptive field in global branches, and the weighted fusion layer through an attention mechanism. and ,get ,in Learned by the network, this structure can automatically adjust the ratio of local to global information based on input features. It can acquire global contextual information while preserving local defect details, and the attention mechanism makes the fusion process adaptive, allowing for optimized feature combinations based on different materials and defect types, thus improving detection accuracy and model adaptability.
[0014] Furthermore, the method for reconstructing the three-dimensional morphology of the material surface is limited to a combination of structured light and multi-view stereo matching. After generating point cloud data, the surface equation is fitted. And based on the curvature of the surface Calculate the spatial influence factor of defects Three-dimensional topography reconstruction provides information on the spatial distribution of defects, and surface curvature quantifies the degree to which defects alter the geometry of the material surface. This allows quality scoring to consider not only the type of defect but also its actual spatial impact, making the evaluation more scientific and reasonable.
[0015] Furthermore, the quality scoring function was extended to a time series model, defined as follows: ,in The number of time sampling points, For a moment The probability of defects is calculated. Introducing time series averaging reduces the impact of random factors on the scoring, making the quality assessment more stable and reliable, better reflecting the quality status of materials over a period of time, and avoiding misjudgments caused by instantaneous fluctuations.
[0016] Furthermore, a dynamic threshold was defined. This represents the historical average quality score. Standard deviation The adjustment coefficient is dynamically adjusted based on the material type. The dynamic threshold can automatically optimize the judgment criteria based on historical data, adapt to quality fluctuations of different materials and batches, improve the accuracy and flexibility of conformity judgment, and reduce misjudgments caused by fixed thresholds.
[0017] Furthermore, the content and storage format of the digital acceptance report are specified, including original and corrected images, feature vectors, defect probability distribution, 3D topographic map, spatial influence factor, quality score and judgment results, timestamp, equipment number, operator ID, etc., stored in encrypted JSON format, and tamper-proof traceability is achieved through blockchain hash value. The report is structurally complete and information-rich; the encrypted storage and blockchain tamper-proof mechanism ensure the authenticity and security of the data, facilitate long-term preservation and traceability, and provide a reliable basis for engineering quality management and accountability.
[0018] This invention provides a method for quality acceptance of construction materials based on AI visual recognition in railway supervision, which has the following beneficial effects: First, in the data acquisition phase, this invention employs a multispectral imaging device to acquire multispectral image sequences covering a wavelength range of 400nm to 1000nm, obtaining no fewer than twelve images in each sampling cycle, while simultaneously controlling the illumination angle to vary according to a preset pattern. This design effectively overcomes problems such as uneven lighting and strong reflections at construction sites, ensuring the acquisition of rich material surface information under different spectral conditions, providing a sufficient data foundation for subsequent feature extraction. Compared to single visible light imaging, multispectral data can reveal differences in absorption and reflection of materials across different wavelength bands, thus making it easier to identify specific defects such as surface oxidation, oil stains, and coating aging.
[0019] Secondly, this invention employs a unique normalization formula in the non-uniform illumination correction stage. Each frame of the image is divided by the average brightness of the entire image at the same wavelength, with a very small constant added to prevent division by zero, thereby eliminating the influence of illumination intensity differences on feature extraction. This correction method ensures the stability and consistency of subsequent feature extraction, enabling the algorithm to maintain high accuracy under different times and weather conditions.
[0020] In terms of feature extraction, this invention designs a multi-dimensional feature extraction module that integrates the improved LBP-TOP spatiotemporal texture pattern, the mean and covariance features of the CIELAB color space, and multi-scale Canny edge density features to form a high-dimensional feature vector. This multi-dimensional feature fusion method can simultaneously capture the material's texture structure, color distribution, and edge contour information, significantly improving the ability to identify complex defects.
[0021] This invention also constructs a material texture-defect mapping model, which combines unsupervised clustering and supervised fine-tuning training strategies and is optimized using a composite loss function that includes classification cross-entropy, regression smoothing, and triplet loss. This training method enables the model to not only accurately classify defect types but also maintain reasonable inter-class distances, improving its robustness to classifying new samples.
[0022] In terms of neural network architecture design, this invention proposes an adaptive convolutional neural network branch structure, which includes a local fine-grained feature extraction branch and a global dilated convolution branch, and dynamically fuses the outputs of the two through an attention mechanism. This design preserves the detailed information of local defects while taking into account the contextual relationships of the global structure, enabling the network to maintain high accuracy when dealing with large-sized material images.
[0023] This invention incorporates the three-dimensional morphology reconstruction results of the material surface into defect assessment. By fitting the surface equation with point cloud data and calculating the curvature, a spatial influence factor of the defect is obtained. This factor incorporates the three-dimensional distribution characteristics of the defect into the scoring system, enabling the score to not only reflect the presence or absence of the defect but also its potential impact on material properties.
[0024] The quality scoring function is another core innovation of this invention. It comprehensively considers the probability distribution of defect types, spatial influence factors, and stability over time, forming a weighted average dynamic scoring model. Compared with traditional single-threshold judgment, this scoring function can more accurately reflect the material quality status and reduce misjudgments and omissions.
[0025] The dynamic threshold determination process adaptively adjusts the threshold based on the mean and standard deviation of historical acceptance data, so that the determination criteria can be optimized as the material type and batch change, avoiding the applicability problem of fixed thresholds in different periods and under different conditions.
[0026] Regarding the generation of acceptance reports, this invention uses encrypted JSON format to store all relevant data and employs blockchain hash values for tamper-proof traceability. This mechanism ensures the authenticity, integrity, and long-term traceability of the acceptance data, providing a reliable guarantee for subsequent project quality management and accountability.
[0027] In summary, this invention achieves full automation, high precision, strong adaptability, and traceability in the quality acceptance of railway construction materials through multiple technological innovations, including multispectral imaging, non-uniform illumination correction, multi-dimensional feature fusion, a unique texture-defect mapping model, adaptive convolutional network structure, three-dimensional topography reconstruction, dynamic scoring and threshold determination, and blockchain-based tamper-proof report generation. Compared to existing technologies, this invention significantly improves detection accuracy, environmental adaptability, defect quantification capabilities, and data management levels, meeting the high standards of intelligent supervision in modern railway engineering and possessing broad application prospects and promotional value. Attached Figure Description
[0028] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.
[0029] Figure 1 This is the overall flowchart of the present invention; Figure 2 This is a flowchart of the data acquisition and correction process of the present invention; Figure 3 This is a flowchart of the feature extraction and defect identification process of the present invention; Figure 4 This is a flowchart of the conformity assessment and report generation process for this invention. Detailed Implementation
[0030] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses consistent with some aspects of this disclosure as detailed in the appended claims.
[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0032] How to use:
[0033] Multispectral image sequence acquisition: Multispectral imaging equipment was deployed at the railway construction site, with wavelengths ranging from 400nm to 1000nm, to acquire multispectral image sequences of construction materials. ,in For pixel coordinates, For wavelength, The sampling time is specified. At least 12 frames of images are acquired in each sampling cycle, and the illumination angle is controlled to vary according to a preset mode to obtain data that is not sensitive to illumination.
[0034] Non-uniform illumination correction: Non-uniform illumination correction is performed on the acquired multispectral image sequence using the formula: ; in The average brightness of the entire image at the same wavelength. This is a very small constant used to prevent division by zero. The corrected image is obtained. .
[0035] Multidimensional feature extraction: A multidimensional feature extraction module is used to process the corrected image, extracting material texture, color distribution, and edge features to form feature vectors. Specifically, this includes improvements. Spatiotemporal texture mode, CIELAB color space mean With covariance matrix Multi-scale Canny edge density The combination is: ; Material Texture-Defect Mapping Model Construction: Constructing a material texture-defect mapping model ,in This represents the set of defect categories. The model training employs a combination of unsupervised clustering and supervised fine-tuning, with the loss function being: ; in For classification cross-entropy loss, To smooth the regression loss, This is the loss of the triplet.
[0036] Adaptive Convolutional Neural Network Branch Structure Feature Fusion: An adaptive convolutional neural network branch structure is used. Local branches employ 3×3 convolutional kernels to extract fine-grained features, while global branches use dilated convolutions to expand the receptive field. The fusion layer uses an attention mechanism for weighted fusion. and ,get: ; in Obtained through network learning. Output defect probability distribution. .
[0037] Material surface 3D topography reconstruction: Combining the results of material surface 3D topography reconstruction Point cloud data is generated using structured light and multi-view stereo matching, and the surface equation is fitted. ; Calculate the spatial impact factor of defects: ; in For the curvature of the surface.
[0038] Quality scoring function calculation based on probability distribution: Introducing the quality scoring function: ; in The number of time sampling points, For a moment The probability of defects, This is the defect weighting coefficient.
[0039] Dynamic threshold for determining compliance: Based on historical acceptance data, a dynamic threshold is calculated. ; in This represents the historical average quality score. Standard deviation This is an adjustment factor. The quality score... and Compare and determine whether the materials are qualified.
[0040] Digital Acceptance Report Generation: Generates a report containing the original multispectral image, the corrected image, and feature vectors. Defect probability distribution 3D topography, spatial influence factor, quality score The digital acceptance report, containing information such as the judgment result, timestamp, equipment number, and operator ID, is stored in encrypted JSON format.
[0041] Full-process traceable automated acceptance: The generated acceptance report is uploaded to the supervision database, and tamper-proof traceability is achieved through blockchain hash value, completing the automated acceptance of railway construction material quality and ensuring that the data throughout the process is traceable, verifiable, and tamper-proof.
[0042] Example 1: Multispectral image sequence acquisition in an open-air railway station environment In the construction material storage area of the open-air railway station, multispectral imaging equipment was set up according to the acceptance plan. The equipment was installed on a rotatable bracket, covering a wavelength range of 400 nanometers to 1,000 nanometers. In each sampling cycle, the illumination angle was changed in a preset mode. Multispectral image sequences of construction materials were collected by combining natural light and auxiliary light sources. The sequence contains spatial coordinates, wavelength and time dimension information, ensuring that light-insensitive and information-rich raw data are obtained in an open environment with frequent changes in illumination, providing a reliable basis for the subsequent acceptance process.
[0043] Example 2: Non-uniform illumination correction in temporary lighting environments during tunnel construction In the material storage area of the tunnel construction section, due to the fixed position of the lighting fixtures and the uneven distribution of light, after acquiring the image sequence using multispectral imaging equipment, non-uniform illumination correction is immediately performed. Each frame of the image is divided by the average brightness of the entire image at the same wavelength and a very small constant is added to eliminate the brightness difference caused by local strong light and shadow in the tunnel, so as to obtain the corrected image. This ensures that the subsequent feature extraction remains stable and consistent in the closed and unevenly lit environment of the tunnel, and avoids feature distortion from affecting defect identification.
[0044] Example 3: Multidimensional Feature Extraction in Nighttime Construction Environment on Elevated Bridges In the nighttime construction area of the elevated bridge deck, after acquiring and correcting images using supplementary lighting devices and multispectral imaging equipment, a multidimensional feature extraction module is applied. This module combines improved spatiotemporal texture patterns, color space distribution features, and multi-scale edge density features to form feature vectors containing texture, color, and contour information. This process can still comprehensively characterize the surface state of materials under low illumination and complex background conditions at night, providing multidimensional complementary information for defect identification and improving the robustness and accuracy of nighttime detection.
[0045] Example 4: Material Texture-Defect Mapping Model Construction and Adaptive Convolutional Neural Network Branch Structure Feature Fusion in the Indoor Environment of a Large Precast Component Factory In the indoor acceptance area of a large precast component factory, the lighting is controllable and the background is relatively uniform. A mapping model between material texture and defect category is constructed based on the extracted feature vectors. The training method combines unsupervised clustering and supervised fine-tuning to optimize the model parameters. At the same time, an adaptive convolutional neural network branch structure is deployed to extract local fine-grained features and global contextual features respectively. These features are dynamically fused through an attention mechanism to output the defect probability distribution. Under stable indoor lighting conditions, defect recognition can retain details while taking into account the overall structure, thereby improving classification accuracy.
[0046] Example 5: Three-dimensional morphology reconstruction, quality scoring, and dynamic threshold determination of material surface in a rainy and humid temporary stockpile environment. In rainy and humid temporary stockpiles, material surfaces may reflect light and deform due to humidity. First, the three-dimensional morphology reconstruction results of the material surface are combined to fit the surface equation and calculate the spatial influence factor of defects. Then, this factor, together with the defect probability distribution, is input into the quality scoring function. This function comprehensively considers defect type, spatial influence, and temporal stability to obtain the material quality score. Subsequently, the pass / fail determination is made based on the dynamic threshold calculated from historical data, and a digital acceptance report containing images, features, scores, and judgment results is generated, realizing fully traceable automated acceptance in humid and variable environments.
[0047] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A construction material quality acceptance method based on AI visual recognition in railway supervision, characterized in that, Includes the following steps: (1) Deploy multispectral imaging equipment at the railway construction site to collect multispectral image sequences of construction materials. ,in For pixel coordinates, For wavelength, The time of data collection; (2) Perform non-uniform illumination correction on the acquired multispectral image sequence to obtain the corrected image. ; (3) A multi-dimensional feature extraction module is used to extract material texture, color distribution, and edge features to form feature vectors. ; (4) Constructing a material texture-defect mapping model ,in Represents a set of defect categories; (5) Use an adaptive convolutional neural network branch structure to fuse local and global features and output the defect probability distribution. ; (6) Combining the results of three-dimensional morphology reconstruction of material surface Calculate the spatial impact factor of defects ; (7) Introduce a quality scoring function based on probability distribution. ,in This is the defect weighting coefficient; (8) Based on dynamic threshold Determine whether the materials are qualified; (9) Generate a digital acceptance report containing images, features, scores and judgment results, and upload it to the supervision database; (10) Achieve fully traceable automated acceptance.
2. The method for quality acceptance of construction materials based on AI visual recognition in railway supervision according to claim 1, characterized in that, The multispectral imaging device uses a wavelength range during the acquisition process. The light source, and collects no less than [number] samples within each sampling period T. Frame images are used to ensure the continuity of time-series data; illumination angles are collected during data acquisition. The preset pattern changes and is used for subsequent extraction of light-insensitive features.
3. The method for quality acceptance of construction materials based on AI visual recognition in railway supervision according to claim 1, characterized in that, The non-uniform illumination correction uses the following formula: in The average brightness of the entire image at the same wavelength. To prevent division by zero of extremely small constants.
4. The method for quality acceptance of construction materials based on AI visual recognition in railway supervision according to claim 1, characterized in that, The multidimensional feature extraction module includes: Texture features: using improved Operators extract spatiotemporal texture patterns; Color characteristics: Calculate the mean in CIELAB space With covariance matrix ; Edge features: Edge density is obtained using Canny multi-scale edge detection. . The final feature vector is:
5. The method for quality acceptance of construction materials based on AI visual recognition in railway supervision according to claim 1, characterized in that, The material texture-defect mapping model It is trained by combining unsupervised clustering and supervised fine-tuning, and its loss function is: in For classification cross-entropy loss, To smooth the regression loss, Triple loss is used to enhance inter-class distance.
6. The method for quality acceptance of construction materials based on AI visual recognition in railway supervision according to claim 1, characterized in that, The adaptive convolutional neural network branch structure includes: Local branches: Fine-grained features are extracted using 3×3 convolution kernels; Global branch: Employs dilated convolution to expand the receptive field; Fusion layer: Weighted fusion of the output feature maps from the two branches using an attention mechanism. and ,get , Learned online.
7. The method for quality acceptance of construction materials based on AI visual recognition in railway supervision according to claim 1, characterized in that, The three-dimensional morphology reconstruction of the material surface employs a combination of structured light and multi-view stereo matching, generating point cloud data and then fitting the surface equation. Spatial influence factor of defects Based on the curvature of the surface calculate:
8. The method for quality acceptance of construction materials based on AI visual recognition in railway supervision according to claim 1, characterized in that, The quality scoring function further considers the stability of defects over time and is defined as follows: in The number of time sampling points, For a moment The probability of defects.
9. The method for quality acceptance of construction materials based on AI visual recognition in railway supervision according to claim 1, characterized in that, The dynamic threshold Based on historical acceptance data, the calculation formula is as follows: in This represents the historical average quality score. Standard deviation The adjustment coefficient is dynamically adjusted according to the material type.
10. The method for quality acceptance of construction materials based on AI visual recognition in railway supervision according to claim 1, characterized in that, The digital acceptance report includes: Original multispectral image and corrected image; Feature vector and defect probability distribution ; Three-dimensional topography and defect space influencing factors; Quality rating And the judgment result; Timestamp, device number, operator ID; The report is stored in encrypted JSON format and can be traced tamper-proofly via blockchain hash value.