A modified asphalt qi content detection method and system based on image analysis

Abstract

The embodiment of the application provides a modified asphalt QI content detection method and system based on image analysis, and relates to the technical field of image analysis. The method comprises the following steps: obtaining an image sequence by acquiring an image flow of liquid asphalt; obtaining an enhanced image sequence by performing enhancement processing on the image sequence; separating target spots from the enhanced image sequence, and identifying independent target spots from the target spots; extracting surface texture features, edge complexity features and internal gray scale distribution features for the independent target spots; classifying the independent target spots according to classification logic based on the surface texture features, the edge complexity features and the internal gray scale distribution features to obtain a classification result, so as to distinguish solid quinoline insoluble particles and bubbles; and according to the classification result, counting the independent target spots identified as solid quinoline insoluble particles to obtain the quinoline insoluble content. The application can improve the classification accuracy and the robustness of the system.

Description

Technical Field

[0001] This application relates to the field of image analysis technology, and more specifically, to a method and system for detecting the QI content of modified asphalt based on image analysis. Background Technology

[0002] Traditional methods for detecting quinoline insoluble (QI) content in modified pitch, such as manual sampling and offline laboratory analysis, suffer from long detection cycles and slow response times, making them unsuitable for meeting the stringent real-time quality feedback requirements of modern production lines. Furthermore, in actual industrial production, the introduction of microbubbles and the deformation of solid quinoline insoluble particles render image analysis methods based on simple shape judgment ineffective, leading to persistent deviations and random fluctuations in detection results. This severely impacts the reliability of online detection systems and the precise control of production quality. Summary of the Invention

[0003] This application aims to address at least one of the technical problems existing in the prior art. To this end, this application proposes a method and system for detecting the QI content of modified asphalt based on image analysis, thereby improving the accuracy of classification and the robustness of the system.

[0004] This application discloses a method and system for detecting the QI content of modified asphalt based on image analysis. It aims to solve the problems in the prior art where the QI content detection method of modified asphalt is affected by air bubbles, resulting in inaccurate detection results, and where spherical solid quinoline insoluble particles are difficult to distinguish from deformed air bubbles under certain working conditions, thus affecting the reliability of the detection system.

[0005] In a first aspect, this application discloses a method for detecting the QI content of modified asphalt based on image analysis, specifically including the following steps:

[0006] The image stream of liquid asphalt is obtained to produce an image sequence;

[0007] An enhanced image sequence is obtained by performing enhancement processing on the image sequence;

[0008] Target blobs were separated from the enhanced image sequence and individual target blobs were identified from the target blobs, which included solid quinoline insoluble particles and bubbles;

[0009] For independent target spots, extract surface texture features, edge complexity features, and internal gray-level distribution features;

[0010] Based on surface texture features, edge complexity features, and internal grayscale distribution features, independent target spots are classified according to classification logic to obtain classification results, so as to distinguish solid quinoline insoluble particles from bubbles.

[0011] Based on the classification results, the independent target spots identified as solid quinoline insoluble particles were quantified to obtain the quinoline insoluble content.

[0012] Through this technical solution, this application can effectively distinguish between solid quinoline insoluble particles and bubbles in an image by extracting and classifying multi-dimensional features, thereby overcoming the detection error caused by bubble interference in traditional methods and significantly improving the accuracy and reliability of QI content detection in modified asphalt.

[0013] Furthermore, in some implementations, the step of separating the target blob from the enhanced image sequence includes:

[0014] The enhanced image sequence is binarized to generate a binary image;

[0015] When the pixel value of a pixel in a binarized image is less than a preset threshold, the pixel in the binarized image is determined to be a target spot; when the pixel value of a pixel in a binarized image is greater than a preset threshold, the pixel in the binarized image is determined to be an asphalt background.

[0016] Based on this, when the pixel value of a pixel in the binarized image is less than a preset threshold, the step of determining the pixel in the binarized image as the target spot includes:

[0017] Eight-connected component analysis is performed on the pixels identified as target spots in the binarized image, and all connected pixels with the same label are marked as the same connected component;

[0018] Pixels in the same connected region are identified as independent target blobs.

[0019] In a preferred embodiment, the steps of extracting surface texture features, edge complexity features, and internal grayscale distribution features for the independent target spot include:

[0020] For each individual target blob, the image within its region is converted into an eight-bit grayscale image;

[0021] Calculate the gray-level co-occurrence matrix based on the 8-bit grayscale image;

[0022] Surface texture features are extracted from the gray-level co-occurrence matrix. These features include energy, contrast, correlation, and entropy.

[0023] In another preferred embodiment, the step of extracting edge complexity features for an independent target blob includes:

[0024] For individual target spots, an edge detection algorithm is used to extract the edge information of the individual target spot contours;

[0025] Edge complexity features are extracted from edge information. These features include edge length, edge density, and edge roughness.

[0026] To enhance functionality, based on surface texture features, edge complexity features, and internal grayscale distribution features, independent target spots are classified according to classification logic to obtain classification results. The steps to distinguish between solid quinoline insoluble particles and bubbles include:

[0027] Track individual target blobs and obtain the correlation between these blobs in the enhanced image sequence;

[0028] Based on the correlation, the instantaneous deformation rate, deformation recovery rate, motion trajectory smoothness, acceleration, and rotation characteristics of independent target spots are extracted.

[0029] Based on surface texture features, edge complexity features, internal grayscale distribution features, instantaneous deformation rate, deformation recovery rate, motion trajectory smoothness, acceleration, and rotation characteristics, independent target spots are classified according to classification logic to obtain classification results, so as to distinguish between solid quinoline insoluble particles and bubbles.

[0030] Furthermore, in some embodiments, the step of tracking individual target blobs and obtaining the association between individual target blobs in the enhanced image sequence includes:

[0031] For all independent target blobs identified in the current frame, calculate the similarity in position, size, and shape between them and the independent target blobs in the previous frame;

[0032] Based on similarity, the Hungarian algorithm is used for optimal matching to associate the independent target blobs in the current frame with the independent target blobs in the previous frame, thereby obtaining the association relationship of the independent target blobs in the image sequence.

[0033] To improve the scheme, based on surface texture features, edge complexity features, internal grayscale distribution features, instantaneous deformation rate, deformation recovery rate, motion trajectory smoothness, acceleration, and rotational characteristics, independent target spots are classified according to classification logic to obtain classification results. This includes distinguishing between solid quinoline insoluble particles and bubbles, and further including:

[0034] The monitoring and classification results yielded a monitoring statistical distribution.

[0035] When the monitored statistical distribution deviates from the preset statistical distribution under stable operating conditions, an adaptive adjustment mechanism is triggered. The adaptive adjustment mechanism includes:

[0036] Based on the current working condition data, the weights and discrimination thresholds of surface texture features, edge complexity features, internal grayscale distribution features, instantaneous deformation rate, deformation recovery rate, motion trajectory smoothness, acceleration, and rotation characteristics in the classification logic are dynamically adjusted. The working condition data includes flow velocity, temperature, and asphalt composition.

[0037] The classification results are corrected based on the adjusted weights and discrimination thresholds.

[0038] Preferably, the step of obtaining the monitoring statistical distribution by continuously monitoring the classification results includes:

[0039] Continuously monitor the classification results;

[0040] Within a set time window, the classification results of each independent target spot are collected in real time, and the proportion of particles and bubbles classified as solid quinoline insoluble matter within the time window is calculated to obtain short-term statistics.

[0041] Based on short-term statistics, the short-term moving average of the classification results is obtained as the monitoring statistical distribution.

[0042] Secondly, this application also discloses an image analysis-based QI content detection system for modified asphalt, the system comprising:

[0043] The image acquisition module is used to acquire the image stream of liquid asphalt to obtain an image sequence;

[0044] The image enhancement processing module is used to enhance image sequences to obtain enhanced image sequences;

[0045] The identification and separation module is used to separate target spots from the enhanced image sequence and identify individual target spots from the target spots, including solid quinoline insoluble particles and bubbles;

[0046] The feature extraction module is used to extract surface texture features, edge complexity features, and internal grayscale distribution features for independent target spots.

[0047] The classification module is used to classify independent target spots according to classification logic based on surface texture features, edge complexity features, and internal gray-scale distribution features to obtain classification results, so as to distinguish between solid quinoline insoluble particles and bubbles.

[0048] The metering module is used to measure the independent target spots that are identified as solid quinoline insoluble particles based on the classification results, so as to obtain the quinoline insoluble content.

[0049] According to the technical solution of this application embodiment, it has at least the following beneficial effects: by acquiring an image stream of liquid asphalt and performing enhancement processing, target spots in the asphalt can be clearly captured. The method further separates and identifies the target spots, distinguishing between solid quinoline insoluble particles and bubbles. For these independent target spots, this application innovatively extracts surface texture features, edge complexity features, and internal grayscale distribution features. These multi-dimensional features can more comprehensively and precisely describe the essential attributes of the target spots. Based on these features, this application uses classification logic to accurately classify the target spots, thereby effectively distinguishing between solid quinoline insoluble particles and bubbles. Finally, based on the classification results, the independent target spots identified as solid quinoline insoluble particles are quantified to obtain accurate quinoline insoluble content.

[0050] The detection method presented in this application overcomes the detection errors caused by bubble interference in traditional methods. Especially in complex scenarios where it is difficult to distinguish between spherical solid quinoline insoluble particles and deformed bubbles, the method significantly improves classification accuracy and system robustness by introducing multi-dimensional feature analysis. This effectively solves the problem of unreliable data in existing online detection systems, avoiding production interruptions and unnecessary process adjustments due to misjudgments. It allows operators to trust the detection data, thereby achieving refined control over production quality and improving production efficiency and product quality.

[0051] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0052] The accompanying drawings are used to provide a further understanding of the technical solutions of this application and constitute a part of the specification. They are used together with the embodiments of this application to explain the technical solutions of this application and do not constitute a limitation on the technical solutions of this application.

[0053] Figure 1 This is a schematic flowchart of an image analysis-based method for detecting the QI content of modified asphalt, provided in one embodiment of this application. Detailed Implementation

[0054] To make the objectives, technical methods, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0055] It should be noted that the meaning of "multiple" (or "more than") in the description of the embodiments of this application refers to two or more, and "greater than," "less than," "exceeding," etc. are understood to exclude the number itself, while "above," "below," "within," etc. are understood to include the number itself. If "first," "second," etc. are described, they are only for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated or the order of the technical features indicated.

[0056] In the description of this application, unless otherwise expressly defined, terms such as "setup," "installation," and "connection" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this application in conjunction with the specific content of the technical solution.

[0057] Based on the above, this application proposes a method and system for detecting the QI content of modified asphalt based on image analysis. See [link to application]. Figure 1 , Figure 1 This is a flowchart illustrating an image analysis-based method for detecting the QI content of modified asphalt according to one embodiment of this application. The image analysis-based method for detecting the QI content of modified asphalt provided in this embodiment includes, but is not limited to, steps S110 to S150, which are described below.

[0058] Step S110: Obtain the image stream of liquid asphalt to obtain the image sequence;

[0059] Step S120: Perform enhancement processing on the image sequence to obtain an enhanced image sequence;

[0060] Step S130: Separate the target spots from the enhanced image sequence and identify independent target spots from the target spots, including solid quinoline insoluble particles and bubbles;

[0061] Step S140: For independent target spots, extract surface texture features, edge complexity features, and internal grayscale distribution features;

[0062] Step S150: Based on surface texture features, edge complexity features, and internal grayscale distribution features, classify independent target spots according to classification logic to obtain classification results, so as to distinguish solid quinoline insoluble particles from bubbles.

[0063] Step S160: Based on the classification results, measure the independent target spots that are identified as solid quinoline insoluble particles to obtain the quinoline insoluble content.

[0064] An "image stream" refers to a data sequence composed of continuously captured image frames, reflecting the dynamic changes of particles and bubbles in liquid asphalt. An "image sequence" is a series of ordered images formed after digitization and preliminary processing of the image stream. A "target spot" refers to a region in the image that differs significantly in color or brightness from the background asphalt; these regions may contain solid quinoline insoluble particles or bubbles. An "independent target spot" refers to a single, independent entity identified after further processing, not adhering to other spots; it is the basic unit for feature extraction and classification. "Surface texture features" describe the fine structure and grayscale variation patterns of the spot's surface, such as roughness and smoothness. "Edge complexity features" reflect the complexity of the spot's outline, such as the degree of edge curvature and length. "Internal grayscale distribution features" describe the distribution of pixel grayscale values ​​within the spot, such as uniformity and concentration. The comprehensive application of these features aims to provide rich and comprehensive information for subsequent intelligent classification.

[0065] First, an image stream of liquid asphalt needs to be acquired to obtain an image sequence. This can be achieved by placing a transparent observation window on the asphalt pipe and continuously capturing images using a high-speed industrial camera. For example, a frequency of 100 frames per second can be used to acquire color images with a resolution of 1920x1080 pixels. These raw image sequences may be affected by factors such as uneven lighting and noise interference, thus requiring image enhancement processing. One approach is to perform histogram equalization on each frame to enhance image contrast and use median filtering to remove salt-and-pepper noise. Another approach is to use gamma correction to adjust image brightness and combine it with Gaussian filtering to smooth the image, reducing the impact of noise on subsequent processing.

[0066] Next, target blobs are separated from the enhanced image sequence, and individual target blobs are identified from these blobs. Target blobs include solid quinoline insoluble particles and bubbles. One method for separating target blobs is to perform adaptive threshold binarization on the enhanced image sequence, dividing the pixels in the image into foreground (target blobs) and background (asphalt). For example, the Otsu algorithm can be used to automatically determine the binarization threshold, identifying regions with pixel values ​​below the threshold as target blobs. When identifying individual target blobs, connected component analysis can be performed on the binarized image, marking all interconnected pixels as the same connected component, thus identifying each connected component as an independent blob.

[0067] For each individual target blob, it is necessary to extract its surface texture features, edge complexity features, and internal gray-level distribution features. For surface texture feature extraction, one approach is to convert the image within the individual target blob region into an 8-bit grayscale image, then calculate the Gray-Level Co-occurrence Matrix (GLCM), and extract texture features such as energy, contrast, correlation, and entropy from the GLCM. For example, the calculation direction of the GLCM can be set to 0 degrees, 45 degrees, 90 degrees, and 135 degrees, and the average value can be taken as the final texture feature. For edge complexity feature extraction, one approach is to use the Canny edge detection algorithm to extract the edge information of the individual target blob contour, and then calculate the edge length, edge density, and edge roughness from the edge information. For example, the edge length can be obtained by calculating the number of edge pixels, and the edge density can be obtained by the ratio of the edge length to the blob area. For internal gray-level distribution feature extraction, one approach is to calculate the gray-level mean, standard deviation, skewness, and kurtosis of all pixels within the individual target blob region. For example, the gray-level mean can reflect the overall brightness of the blob, and the standard deviation can reflect the dispersion of the gray-level values.

[0068] Finally, based on the extracted surface texture features, edge complexity features, and internal grayscale distribution features, independent target spots are classified according to classification logic to distinguish between solid quinoline insoluble particles and bubbles. The independent target spots identified as solid quinoline insoluble particles are then measured according to the classification results to obtain the quinoline insoluble content. One classification logic involves constructing a Support Vector Machine (SVM) classifier and training it using pre-labeled solid quinoline insoluble particle and bubble samples. After training, the extracted feature vectors are input into the SVM classifier, which outputs the classification result for each independent target spot. For example, when the classifier outputs "solid quinoline insoluble particles," it is included in the quinoline insoluble content. During measurement, the number of spots identified as solid quinoline insoluble particles can be counted, and the quinoline insoluble content can be calculated by combining the area or volume of each spot with a preset conversion factor.

[0069] The image analysis-based method for detecting the QI content of modified asphalt proposed in this application works by employing multi-stage, multi-dimensional image processing and feature analysis to accurately distinguish and measure solid quinoline insoluble particles and bubbles in liquid asphalt. First, image acquisition and enhancement processing provide high-quality image data for subsequent refined analysis. Acquiring image sequences ensures the capture of dynamic processes, while enhancement processing effectively improves image quality, laying the foundation for accurate identification of target spots.

[0070] Subsequently, by separating target spots and identifying independent target spots, potential interferences (such as bubbles) and target objects (such as solid quinoline insoluble particles) in the image are extracted from the background, ensuring that each analyzed object is an independent entity. This step is the foundation for subsequent feature extraction and classification, avoiding misjudgments caused by spot adhesion.

[0071] The key lies in extracting surface texture features, edge complexity features, and internal grayscale distribution features for each individual target spot. These features characterize the physical properties of the spot from different dimensions. Surface texture features reflect the uniformity or roughness of the spot's internal structure; for example, solid particles typically have more complex textures, while bubble surfaces are relatively smooth. Edge complexity features capture the regularity of the spot's outline; for example, bubbles tend to be regularly round in static or stable flow, while solid particles are mostly irregular in shape. Internal grayscale distribution features further reveal the density or transparency differences of the material inside the spot; for example, the grayscale distribution inside solid particles may be more uneven, while the interior of bubbles may exhibit a more uniform low grayscale value. The comprehensive use of these multi-dimensional features allows the system to describe the essential properties of each spot more comprehensively and accurately, thus overcoming the limitations of judging by a single shape feature.

[0072] Finally, based on these multidimensional features, independent target spots are classified using a pre-defined classification logic to distinguish between solid quinoline insoluble particles and bubbles. This classification logic can be a classification model built on a machine learning algorithm, which establishes a mapping relationship between features and categories by learning the features of a large number of known samples. Once the classification is complete, the system can accurately measure the spots identified as solid quinoline insoluble particles based on the classification results, thereby obtaining an accurate quinoline insoluble content. The entire process forms a closed loop; from image input to the final content output, each step is closely linked and works together to ensure the reliability and real-time nature of the detection results.

[0073] The image analysis-based method for detecting the QI content of modified asphalt proposed in this application demonstrates significant technological progress and innovation, addressing the misjudgment problem caused by the deformation of air bubbles and solid quinoline insoluble particles in existing technologies. Traditional methods primarily rely on single shape features, such as roundness, to distinguish between air bubbles and solid particles. However, as described in the background section, the presence of aggregates of spherical solid quinoline insoluble particles and deformed air bubbles in actual production environments renders this single-feature judgment logic ineffective, leading to persistent deviations and random fluctuations in the detection results.

[0074] The core innovation of this application lies in its departure from limiting itself to a single shape feature and instead introducing a multi-dimensional feature extraction strategy, including surface texture features, edge complexity features, and internal grayscale distribution features. By comprehensively analyzing these features, this application can more comprehensively and accurately characterize the essential properties of each individual target spot. For example, even if a bubble exhibits an irregular shape due to deformation under force, its internal grayscale distribution and surface texture may still differ significantly from those of solid particles; similarly, even if an aggregate of solid particles is spherical, its surface texture and edge complexity may differ from those of bubbles. This fusion analysis of multi-dimensional features greatly enhances the system's ability to distinguish between solid quinoline insoluble particles and bubbles.

[0075] Compared with the closest existing technology, the advantages of this application are as follows: First, by acquiring and enhancing the image stream of liquid asphalt, higher quality image data is provided for subsequent refined analysis, reducing the impact of noise and uneven illumination on recognition. Second, separating and identifying independent target spots from the enhanced image sequence ensures that each analyzed object is independent, avoiding misjudgments caused by spot adhesion. Most importantly, this application constructs a more robust and accurate classification model by extracting surface texture features, edge complexity features, and internal grayscale distribution features. This multi-feature fusion method enables the system to effectively cope with the problem of bubble and solid particle deformation under complex working conditions, significantly improving the accuracy and reliability of QI content detection. Therefore, this application can provide more stable and reliable online detection data, providing solid technical support for refined control of the production process and improvement of product quality, overcoming the drawbacks of unreliable and untrustworthy data in existing technologies.

[0076] In some embodiments, the step of separating the target blob from the enhanced image sequence described above includes:

[0077] The enhanced image sequence is binarized to generate a binarized image;

[0078] When the pixel value of a pixel in the binarized image is less than a preset threshold, the pixel in the binarized image is determined to be a target spot; when the pixel value of a pixel in the binarized image is greater than the preset threshold, the pixel in the binarized image is determined to be an asphalt background.

[0079] Binarization refers to setting the grayscale value of each pixel in an image to one of two fixed values, such as 0 or 255, according to a certain standard, thereby dividing the image content into foreground and background. This generates a binarized image containing only two pixel values, clearly representing the target and non-target areas. A pixel is the basic unit of an image, and its value represents the brightness or grayscale information of that point. A preset threshold is a key discrimination parameter; its selection can be based on experience or dynamically determined by an adaptive algorithm according to the statistical characteristics of the image to adapt to different lighting conditions and asphalt sample characteristics. Target spots in modified asphalt images typically appear as areas with significant grayscale differences from the surrounding asphalt background; these spots include solid quinoline insoluble particles and bubbles. The asphalt background refers to the liquid asphalt areas in the image other than the target spots.

[0080] The proposed solution simplifies complex grayscale images into binary images containing only the foreground (target spots) and background (asphalt background) by binarizing the enhanced image sequence. Specifically, since solid quinoline insoluble particles and bubbles in liquid asphalt images typically exhibit different grayscale or brightness characteristics from the asphalt background (e.g., they may be darker or brighter), they can be distinguished by setting an appropriate preset threshold. When a pixel's value is less than the preset threshold, the pixel is identified as a target spot; otherwise, it is identified as asphalt background. This mechanism, based on comparing pixel values ​​with a threshold, effectively and quickly separates target spots from the asphalt background, laying the foundation for subsequent feature extraction and classification.

[0081] The above technical solution enables effective separation of target spots in enhanced image sequences. This method leverages the simplicity and efficiency of binarization processing, clearly distinguishing solid quinoline insoluble particles and bubbles from the asphalt background by setting a preset threshold. This not only simplifies the complexity of subsequent image processing but also improves the accuracy and efficiency of target spot identification, providing a reliable image data foundation for subsequent feature extraction and quinoline insoluble content measurement.

[0082] In some embodiments, after the step of determining that a pixel in the binarized image is a target blob when the pixel value of the pixel in the binarized image is less than a preset threshold, the following processing procedure may be further included:

[0083] Eight-connected component analysis is performed on the pixels identified as target spots in the binarized image, and all connected pixels with the same label are marked as the same connected component;

[0084] Pixels in the same connected region are identified as independent target blobs.

[0085] Eight-connected region analysis is an image processing technique aimed at identifying interconnected pixel regions in an image. Specifically, for each pixel in a binarized image identified as a target blob, its eight neighboring pixels (horizontally, vertically, and diagonally) are examined. If these neighboring pixels also belong to the target blob, they are considered connected. By iteratively grouping all connected pixels into the same set, one or more connected regions are formed. All pixels within each connected region are assigned the same label to indicate that they belong to the same continuous region. Thus, all pixels within a connected region are collectively identified as a single, independent target blob, thereby recognizing discrete but physically continuous solid quinoline insoluble particles and bubbles in the image as single, independently processable objects.

[0086] The proposed solution effectively organizes discrete pixels identified as target spots in a binarized image into physically meaningful independent entities through octal connected component analysis. In asphalt images, solid quinoline insoluble particles and bubbles typically appear as regions composed of multiple adjacent pixels. Without connected component analysis, these regions might be treated as multiple unrelated pixels, leading to difficulties in subsequent feature extraction and classification. Octal connected component analysis ensures that all pixels constituting a complete physical entity are correctly identified as a whole, i.e., an independent target spot. This approach avoids incorrectly segmenting a complete particle or bubble into multiple parts and also prevents the incorrect merging of unrelated pixels.

[0087] The above technical solution significantly improves the accuracy and robustness of identifying independent target spots from enhanced image sequences. By marking all connected pixels of the same type as the same connected region and identifying them as independent target spots, it ensures that each identified independent target spot corresponds to a real physical entity, such as a complete solid quinoline insoluble particle or a complete bubble. This lays a solid foundation for subsequent extraction of surface texture features, edge complexity features, and internal grayscale distribution features for independent target spots, thereby improving the reliability of the classification results and ultimately making the measurement of quinoline insoluble content more accurate.

[0088] In some embodiments of this application described above, the step of extracting surface texture features for the independent target blob includes: converting the image within the region of the independent target blob into an eight-bit grayscale image; calculating a gray-level co-occurrence matrix based on the eight-bit grayscale image; and extracting surface texture features from the gray-level co-occurrence matrix, wherein the surface texture features include: energy, contrast, correlation, and entropy.

[0089] Specifically, the image within the independent target spot region is converted into an 8-bit grayscale image. This aims to map the pixel values ​​of the color image to grayscale levels from 0 to 255, thereby standardizing the image's brightness information and providing a unified data foundation for subsequent texture analysis. The 8-bit grayscale image retains sufficient grayscale levels to reflect the image's detailed information. Further, a Gray-Level Co-occurrence Matrix (GLCM) is calculated based on the 8-bit grayscale image. The GLCM is a statistical texture analysis method that describes the texture features of an image by calculating the frequency of occurrence of pixel pairs with specific grayscale values ​​and spatial relationships. For example, different displacement vectors (such as horizontal, vertical, and diagonal directions) can be set to capture texture information in different directions. The surface texture features extracted from the GLCM include energy, contrast, correlation, and entropy. Among them, energy reflects the uniformity of image gray-level distribution or the coarseness of texture; a high energy value usually indicates uniform texture. Contrast measures the magnitude of local gray-level differences in an image, reflecting the depth and clarity of texture; a high contrast value indicates large texture differences. Correlation describes the correlation of image gray-level space, reflecting the regularity or directionality of texture; a high correlation value indicates that the texture has strong directionality. Entropy, as a measure of image information, reflects the complexity or randomness of texture; a high entropy value indicates that the texture is complex and disordered.

[0090] This application's approach converts images of independent target spot regions into standardized 8-bit grayscale images, laying the foundation for subsequent texture analysis. Based on this, the gray-level co-occurrence matrix is ​​calculated, comprehensively capturing the spatial relationships between pixel grayscale values ​​in the image, thereby quantitatively describing the surface texture characteristics of independent target spots. By extracting four key texture features—energy, contrast, correlation, and entropy—the visual appearance of solid quinoline insoluble particles and bubbles can be finely characterized from multiple dimensions. These features can effectively distinguish the inherent differences in texture between solid particles and bubbles; for example, solid particles may exhibit a rougher, more irregular texture, while bubbles may have a smoother, more uniform surface. This provides highly discriminative input data for subsequent classification logic.

[0091] By employing the aforementioned technical solutions, the surface texture features of independent target spots are standardized and quantified in multiple dimensions, ensuring the accuracy and robustness of feature extraction. This detailed and targeted texture feature extraction method significantly enhances the classifier's ability to distinguish between solid quinoline insoluble particles and bubbles, thereby improving the accuracy and reliability of QI content detection in modified asphalt.

[0092] In some embodiments, the step of extracting edge complexity features for the aforementioned independent target spots includes:

[0093] For the individual target spots, an edge detection algorithm is used to extract the edge information of the individual target spot contours;

[0094] Edge complexity features are extracted from the edge information, including edge length, edge density, and edge roughness.

[0095] Edge detection algorithms are used to identify the boundary between individual target blobs and their surrounding background. Commonly used edge detection algorithms include, but are not limited to, the Canny operator, the Sobel operator, the Prewitt operator, or the Laplacian operator. These algorithms determine edge locations by calculating gradient changes in image pixels, thereby generating edge information of the individual target blobs' contours. This edge information is typically represented as a series of connected edge pixels that collectively delineate the outer contour of the individual target blobs.

[0096] Furthermore, the edge complexity features extracted from the edge information specifically include:

[0097] Edge length: can be understood as the perimeter of the outline of an independent target spot. It is usually calculated by counting the number of edge pixels that make up the outline, or by calculating and accumulating the distances between adjacent edge pixels.

[0098] Edge density: This refers to the ratio of the number of edge pixels of an individual target spot to the total number of pixels of that individual target spot. This feature reflects the density of the pixel distribution within the spot or on its surface edges.

[0099] Edge roughness: This can be used to quantify the degree of irregularity or complexity of the profile of an independent target blob. For example, its roughness can be characterized by calculating the local curvature variation of the profile, fractal dimension, or based on a specific mathematical model. These features can effectively capture the morphological differences between solid quinoline insoluble particles and bubbles; for example, solid particles typically have more irregular and rougher edges, while bubbles tend to have relatively smooth edges.

[0100] The proposed solution, through the detailed edge complexity feature extraction process described above, provides a more refined and comprehensive basis for subsequent classification. Edge detection algorithms can accurately locate the boundaries of independent target spots, thereby obtaining their precise contour information. Based on this, by quantifying edge length, edge density, and edge roughness, the essential morphological differences between solid quinoline insoluble particles and bubbles can be effectively captured. For example, due to their solid nature and formation process, solid quinoline insoluble particles often have irregular shapes and rough surface edges, resulting in relatively large edge lengths, high edge densities, and significant roughness. Bubbles, on the other hand, due to their fluid properties and surface tension, typically exhibit more regular circular or elliptical shapes, with relatively small edge lengths, low edge densities, and low edge roughness. Extracting these multi-dimensional edge features significantly enhances the classifier's ability to distinguish between the two types of target spots.

[0101] The above technical solution enables a more comprehensive and detailed description of the morphological characteristics of independent target spots. Edge length, edge density, and edge roughness, as complementary features, effectively capture subtle differences in edge morphology between solid quinoline insoluble particles and bubbles, improving the accuracy and comprehensiveness of feature extraction. This provides more reliable input for subsequent classification steps, thereby enhancing the overall accuracy and robustness of modified asphalt QI content detection. This meticulous feature extraction method helps reduce misjudgments and ensures more accurate measurement results for quinoline insoluble content.

[0102] In some embodiments of this application, the step of classifying independent target spots according to classification logic based on surface texture features, edge complexity features, and internal grayscale distribution features to obtain classification results, in order to distinguish between solid quinoline insoluble particles and bubbles, includes:

[0103] Track the independent target blobes to obtain their correlation relationships within the enhanced image sequence;

[0104] Based on the aforementioned correlation, the instantaneous deformation rate, deformation recovery rate, motion trajectory smoothness, acceleration, and rotational characteristics of the independent target spots are extracted.

[0105] Based on the surface texture features, edge complexity features, internal grayscale distribution features, instantaneous deformation rate, deformation recovery rate, motion trajectory smoothness, acceleration, and rotation characteristics, the independent target spots are classified according to classification logic to obtain classification results, so as to distinguish between solid quinoline insoluble particles and bubbles.

[0106] Specifically, tracking the independent target blobs and obtaining their associations in the enhanced image sequence refers to identifying and associating the same independent target blobs in consecutive image frames, thereby establishing their continuity in the temporal dimension. This can be achieved through target tracking algorithms, such as feature matching or motion prediction methods. Its purpose is to provide basic data for subsequent dynamic feature extraction. The instantaneous deformation rate can be understood as the rate at which the shape or size of an independent target blobs changes between adjacent image frames. Specifically, it can be obtained by calculating the rate of change of geometric parameters such as the blobs' area, perimeter, or principal axis length over time. Its purpose is to quantify the degree of instantaneous deformation of the blobs. The deformation recovery rate refers to the rate at which an independent target blobs recover their original or stable shape after undergoing deformation. This can be evaluated by comparing the degree of deformation of the blobs at different time points with a reference state. Its purpose is to reflect the elasticity or stability of the blobs. The motion trajectory smoothness refers to the smoothness of the motion path of an independent target blobs in the image sequence. Specifically, it can be evaluated by analyzing the positional changes of the blobs' center point in consecutive frames and applying smoothing algorithms (such as Kalman filtering). Its purpose is to distinguish between regular and irregular motion. Acceleration refers to the rate of change of the velocity of an independent target blob in an image sequence. Specifically, it can be obtained by calculating the velocity change of the blob's center point across consecutive frames. Its purpose is to reflect the intensity of the force or motion state of the blob. Rotation characteristics refer to the degree or direction of rotation of an independent target blob around its own center in an image sequence. Specifically, it can be obtained by analyzing the changes in the blob's principal axis direction across consecutive frames. In practical applications, the classification logic, based on the aforementioned surface texture features, edge complexity features, and internal grayscale distribution features, further integrates these dynamic features to form a more comprehensive feature vector, which is used to train classification models, such as Support Vector Machines (SVM), neural networks, or decision trees, thereby achieving more accurate classification.

[0107] This application's solution effectively addresses the misclassification problem that may occur when relying solely on static features for classification by introducing the dynamic characteristics of independent target spots. Solid quinoline insoluble particles typically have a relatively rigid structure, with their shape, size, and motion state changing little in a short period, resulting in relatively stable trajectories. In contrast, bubbles are highly deformable, with their shape and size changing significantly with environmental pressure, temperature, or their own motion, and their trajectories often exhibiting irregularities or specific upward trends due to buoyancy. By tracking independent target spots and extracting their instantaneous deformation rate, deformation recovery rate, trajectory smoothness, acceleration, and rotational characteristics, these essential differences can be captured. For example, bubbles typically exhibit a high instantaneous deformation rate and a low deformation recovery rate, with less smooth trajectories and potentially significant acceleration or rotation. Solid particles, on the other hand, exhibit a lower deformation rate and a higher deformation recovery rate (if deformation occurs), smoother trajectories, and more stable acceleration and rotational characteristics. Combining these dynamic features with existing surface texture features, edge complexity features, and internal grayscale distribution features can provide richer and more discriminative information for classification logic, thereby significantly improving the classification model's ability to distinguish between solid quinoline insoluble particles and bubbles.

[0108] In some preferred embodiments, it is assumed that a target spot exists in the image stream of liquid asphalt, whose static image features (such as surface texture, edge complexity, and internal grayscale distribution) are very similar to certain solid quinoline insoluble particles, making it difficult for classification methods based solely on static features to accurately determine whether it is an air bubble or a solid particle. In this case, the solution of this application tracks the target spot. Through analysis of consecutive image frames, it can be observed that the target spot undergoes significant shape changes in a short period of time. For example, its instantaneous deformation rate is high, and it fails to quickly recover to its initial shape after deformation, i.e., its deformation recovery rate is low. Simultaneously, its trajectory may exhibit an irregular upward trend, with low trajectory smoothness and accompanied by certain acceleration changes. These dynamic features are significantly different from the rigidity of solid quinoline insoluble particles, but highly consistent with the deformability and buoyancy characteristics of air bubbles. Therefore, even if the static features are misleading, combined with these dynamic features, the classification logic can accurately identify the target spot as an air bubble, avoiding misclassification as a solid quinoline insoluble particle, thereby ensuring the accuracy of quinoline insoluble content measurement.

[0109] In some embodiments, the steps of tracking independent target blobs and obtaining their association in the enhanced image sequence may include: for all independent target blobs identified in the current frame, calculating the similarity in position, size, and shape between them and the independent target blobs in the previous frame; and using the Hungarian algorithm to perform optimal matching based on the similarity, associating the independent target blobs in the current frame with the independent target blobs in the previous frame, and obtaining the association of the independent target blobs in the image sequence.

[0110] Specifically, when processing the enhanced image sequence, individual target blobs in each frame are identified. To establish the correspondence between these individual target blobs across consecutive frames, it is necessary to calculate the similarity between each individual target blob in the current frame and each individual target blobs in the previous frame. This similarity can consider multiple dimensions. For example, the positional similarity of individual target blobs can be obtained by calculating the Euclidean distance between their centroid coordinates; size similarity can be determined by comparing their area or perimeter; and shape similarity can be quantified by calculating the distance between shape descriptors (such as Hu moments, Zernike moments, or Fourier descriptors). The combined use of these similarity metrics can more comprehensively reflect the potential correspondence between individual target blobs across consecutive frames.

[0111] The Hungarian algorithm, a classic combinatorial optimization algorithm, is commonly used to solve assignment problems. In this application, it is used to find the optimal one-to-one correspondence between independent target blobs in the current frame and the previous frame. By using the similarity calculation result as the matching cost, the Hungarian algorithm can efficiently determine a globally optimal matching scheme, thereby accurately associating the independent target blobs in the current frame with those in the previous frame. This allows us to obtain the association relationship of independent target blobs throughout the entire image sequence, laying the foundation for subsequent dynamic feature extraction.

[0112] The proposed solution effectively addresses the tracking problem in complex situations where independent target spots may move, deform, split, or merge in consecutive image frames by calculating the similarity in position, size, and shape between the current frame and the previous frame, and combining this with the Hungarian algorithm for optimal matching. Specifically, the similarity calculation provides a quantitative basis for matching, ensuring its accuracy; while the Hungarian algorithm finds a unique match with the highest total similarity (or lowest total dissimilarity) among all possible matching combinations, thus avoiding erroneous associations between many-to-one or one-to-many pairs. It is precisely this precise and robust association mechanism that enables the accurate extraction of dynamic features such as instantaneous deformation rate, deformation recovery rate, motion trajectory smoothness, acceleration, and rotation characteristics of independent target spots. These dynamic features are crucial for distinguishing between solid quinoline insoluble particles and bubbles.

[0113] In some embodiments, the step of classifying independent target spots according to classification logic based on surface texture features, edge complexity features, internal grayscale distribution features, instantaneous deformation rate, deformation recovery rate, motion trajectory smoothness, acceleration, and rotational characteristics to obtain classification results and distinguish between solid quinoline insoluble particles and bubbles includes the following:

[0114] The classification results are continuously monitored to obtain the monitoring statistical distribution;

[0115] When the monitored statistical distribution deviates from the preset statistical distribution under stable operating conditions, an adaptive adjustment mechanism is triggered, the adaptive adjustment mechanism including:

[0116] Based on the current working condition data, the weights and discrimination thresholds of the surface texture features, edge complexity features, internal grayscale distribution features, instantaneous deformation rate, deformation recovery rate, motion trajectory smoothness, acceleration, and rotation characteristics in the classification logic are dynamically adjusted. The working condition data includes flow velocity, temperature, and asphalt composition.

[0117] The classification results are corrected based on the adjusted weights and discrimination thresholds.

[0118] Specifically, continuous monitoring of classification results to obtain statistical distribution refers to the real-time or near-real-time tracking and statistical analysis of the classification results of solid quinoline insoluble particles and bubbles output by the classifier. For example, it can be used to statistically analyze the proportion of particles and bubbles classified as solid quinoline insoluble particles and bubbles within a certain time window, as well as the distribution of average feature values, to form a statistical distribution reflecting the current classifier performance and sample characteristics. Its purpose is to promptly detect potential drift in classifier performance or changes in the external environment.

[0119] Specifically, an adaptive adjustment mechanism is triggered when the monitored statistical distribution deviates from the preset stable operating condition statistical distribution. The preset stable operating condition statistical distribution is a baseline distribution established based on historical data from system operation under ideal or standard operating conditions. When the current monitored statistical distribution differs significantly from this baseline distribution, it indicates that the system may be facing new operating conditions or that the classifier performance may be declining, at which point the adaptive adjustment mechanism needs to be activated.

[0120] In practical applications, the adaptive adjustment mechanism involves dynamically adjusting the feature weights and discrimination thresholds in the classification logic based on current operating condition data. Operating condition data can be understood as external environmental parameters that affect the features of asphalt images, such as flow velocity, temperature, and asphalt composition. Changes in these parameters directly affect the appearance of solid quinoline insoluble particles and bubbles in the image. By acquiring this operating condition data in real time, the system can intelligently adjust the relative importance (weights) of various features in the classification model (such as surface texture features, edge complexity features, internal grayscale distribution features, instantaneous deformation rate, deformation recovery rate, motion trajectory smoothness, acceleration, and rotational characteristics) as well as the discrimination boundary (threshold) for distinguishing between two types of targets. The aim is to enable the classification logic to better adapt to the current actual operating environment and improve classification accuracy.

[0121] Furthermore, the classification results are corrected based on the adjusted weights and discrimination thresholds. This means that after the classification logic parameters are dynamically adjusted, subsequent independent target spots will be discriminated using a new classification logic that is more adapted to the current working conditions, thereby obtaining more accurate classification results.

[0122] This application's solution effectively addresses the problem of image feature drift caused by changes in operating conditions (such as flow rate, temperature, and asphalt composition) in actual industrial production, which affects classification accuracy. Specifically, continuous monitoring of classification results allows for real-time monitoring of the classifier's operating status and output distribution. Once a deviation from the baseline distribution under stable operating conditions is detected, potential performance degradation or environmental changes in the classifier can be identified promptly. It is this identification of deviation that triggers the adaptive adjustment mechanism. This mechanism acquires current operating condition data, such as flow rate, temperature, and asphalt composition, to quantify the impact of the current environment on image features. Based on this, the system dynamically adjusts the weights and discrimination thresholds of various features in the classification logic, enabling the classifier to self-optimize based on the latest environmental information. For example, at high temperatures, bubbles may be more active and have a higher deformation rate; in this case, the system can increase the weight of the deformation rate feature. Conversely, under specific asphalt compositions, the texture of solid particles may be more pronounced; in this case, the weight of the texture feature can be adjusted. This dynamic adjustment ensures that the classification logic always matches the current actual working conditions, thereby guaranteeing the accurate differentiation between solid quinoline insoluble particles and bubbles.

[0123] In some preferred embodiments, it is assumed that on a modified asphalt production line, the system was initially trained and calibrated under stable operating conditions at a standard temperature of 200°C and a standard flow rate of 10 m / s. At this time, a preset statistical distribution under stable operating conditions is established, for example, the classification ratio of solid quinoline insoluble particles to bubbles is 1:5 in a minute-by-minute image sequence.

[0124] In actual operation, if the production line temperature rises to 220°C or the flow rate decreases to 8 m / s, these changes in operating conditions may cause changes in the size and deformation characteristics of bubbles in liquid asphalt, or the movement trajectory of solid particles may become irregular. At this time, the continuous monitoring module will collect the classification results in real time and calculate short-term statistics within a set time window (e.g., every 5 minutes). If it is found that the classification ratio of solid quinoline insoluble particles to bubbles becomes 1:3 within the current time window, which deviates significantly from the preset 1:5, it indicates that the classifier may no longer be suitable for the current operating conditions.

[0125] At this point, the adaptive adjustment mechanism will be triggered. The system will acquire the current operating conditions, namely a temperature of 220°C and a flow rate of 8 m / s. Based on this data, the adaptive adjustment mechanism will dynamically adjust the weights and thresholds of various features in the classification logic. For example, since increased temperature may lead to a higher bubble deformation rate, the system may increase the weights of the instantaneous deformation rate and deformation recovery rate features and adjust the corresponding thresholds to better capture the dynamic characteristics of the bubbles. Simultaneously, if a decrease in flow rate makes the particle trajectory smoother, the system may adjust the weights of the trajectory smoothness feature. Through this dynamic adjustment, the classification logic can readjust to the operating conditions of 220°C and 8 m / s, thereby correcting subsequent classification results to achieve accuracy consistent with reality, for example, readjusting the classification ratio to a reasonable range close to 1:5.

[0126] In some embodiments, the step of continuously monitoring the classification results to obtain the monitoring statistical distribution includes:

[0127] Continuously monitor the classification results;

[0128] Within a set time window, the classification results of each independent target spot are collected in real time, and the proportion of particles and bubbles classified as solid quinoline insoluble matter within the time window is calculated to obtain short-term statistics.

[0129] Based on the short-term statistics, the short-term moving average of the classification results is obtained as the monitoring statistical distribution.

[0130] Specifically, continuous monitoring of the classification results means that the system continuously receives and processes the classification judgment result of each independent target spot output by the classification module, i.e., whether the spot is identified as a solid quinoline insoluble particle or an air bubble. The set time window can be understood as a pre-configured duration, such as several seconds or tens of seconds, aimed at locally aggregating the classification results to smooth out instantaneous fluctuations. Within this time window, the system collects the classification results of all classified independent target spots in real time. Subsequently, based on these collected results, the system calculates the number of spots identified as solid quinoline insoluble particles and air bubbles within the time window, and further calculates their proportion in the total number, thus obtaining a short-term statistic. For example, if 100 independent target spots are identified within a certain time window, of which 80 are identified as solid quinoline insoluble particles and 20 as air bubbles, the short-term statistic can be expressed as 80% solid quinoline insoluble particles and 20% air bubbles. Finally, based on these continuously generated short-term statistics, the monitoring statistical distribution is obtained by calculating their short-term moving average. The moving average can further smooth out fluctuations in short-term statistics, providing a more representative and stable trend distribution, thus more accurately reflecting the actual changes in the ratio of solid quinoline insoluble particles to bubbles under current operating conditions.

[0131] This application's solution effectively addresses the instability issues that may arise from simple continuous monitoring by introducing short-term statistical calculations within a time window and processing with short-term moving averages. Specifically, collecting classification results and calculating quantity ratios within a set time window aggregates instantaneous, discrete classification judgments into statistical data with a certain time span, thereby filtering out random noise and instantaneous interference to some extent. Furthermore, calculating short-term moving averages based on these short-term statistics ensures that the monitoring statistical distribution no longer relies solely on the latest single-point-in-time data but comprehensively considers trends over a period of time. This approach makes the monitoring statistical distribution more resistant to short-term fluctuations and more accurately reflects the true trend of the proportion of solid quinoline insoluble particles and bubbles in modified asphalt, thus providing a more reliable input for subsequent adaptive adjustment mechanisms.

[0132] In some preferred embodiments, an example is illustrated below. Assume a time window of 5 seconds. The system continuously monitors the classification results, collecting the classification results of all independent target spots within that period every 5 seconds. For example, in the first 5-second window, the system might identify 100 independent target spots, of which 80 are classified as solid quinoline insoluble particles and 20 as bubbles, resulting in a short-term statistic of 80% for solid quinoline insoluble particles. In the second 5-second window, it might identify 110 independent target spots, of which 85 are classified as solid quinoline insoluble particles and 25 as bubbles, resulting in a short-term statistic of approximately 77.3% for solid quinoline insoluble particles. These short-term statistics are then used to calculate a short-term moving average. For example, a moving average can be calculated using a moving window containing the three most recent short-term statistics. When a new short-term statistic is generated, the oldest statistic is removed, the new statistic is added, and the average is recalculated. Therefore, the obtained monitoring statistical distribution is a smooth and continuously changing curve, which can effectively reflect the long-term trend of the content of solid quinoline insoluble particles, rather than discrete points affected by instantaneous fluctuations. When the monitoring statistical distribution represented by this moving average deviates significantly from the statistical distribution under the preset stable operating conditions, an adaptive adjustment mechanism will be triggered to dynamically adjust the classification logic.

[0133] This application also discloses a modified asphalt QI content detection system based on image analysis, the system comprising:

[0134] The image acquisition module is used to acquire the image stream of liquid asphalt to obtain an image sequence;

[0135] An image enhancement processing module is used to enhance the image sequence to obtain an enhanced image sequence;

[0136] The identification and separation module is used to separate target spots from the enhanced image sequence and identify independent target spots from the target spots, the target spots including solid quinoline insoluble particles and bubbles;

[0137] The feature extraction module is used to extract surface texture features, edge complexity features, and internal grayscale distribution features for the independent target spots.

[0138] The classification module is used to classify the independent target spots according to the classification logic based on the surface texture features, the edge complexity features, and the internal grayscale distribution features to obtain classification results, so as to distinguish between solid quinoline insoluble particles and bubbles;

[0139] The metering module is used to measure the independent target spots that are identified as solid quinoline insoluble particles according to the classification results, so as to obtain the quinoline insoluble content.

[0140] The image enhancement processing module can be a standalone image processing unit, such as an embedded processor or a graphics processing unit (GPU), programmed to execute image enhancement algorithms. Specifically, this module can be designed to receive raw image sequences from the image acquisition module and process them in real time, for example, by optimizing image quality through preset contrast adjustment, brightness correction, or noise filtering algorithms. As a preferred implementation, this module can employ a software-based image processing library, running on a general-purpose computing platform, where the algorithm parameters need to be preset and fixed.

[0141] The identification and separation module can be implemented as a software component running on the main processor's computing resources, or as a dedicated hardware accelerator, such as a programmable gate array (FPGA), to achieve efficient image segmentation and connected component analysis. Specifically, the module is configured to receive an enhanced image sequence and apply image segmentation techniques, such as fixed-threshold segmentation or region growing algorithms, to distinguish target spots from the asphalt background. Subsequently, the module further performs connected component analysis to ensure that each identified solid quinoline insoluble particle or bubble is treated as an independent analysis object.

[0142] The feature extraction module can be implemented as a set of algorithms deployed on a central processing unit (CPU) or a dedicated digital signal processor (DSP). This module is designed to receive independent target blob data output from the recognition and separation module and compute multiple features in parallel or serially. For example, this module may include sub-modules for calculating the gray-level co-occurrence matrix, sub-modules for performing edge detection algorithms, and sub-modules for statistically analyzing gray-level distribution parameters. These sub-modules work together to generate a multi-dimensional feature vector for each independent target blob, with pre-defined and fixed computational parameters and methods.

[0143] The classification module can be implemented as a machine learning model inference engine, loaded and running on a high-performance computing unit, such as a multi-core CPU, GPU, or dedicated neural network processor (NPU). This module is configured to receive multi-dimensional feature vectors output by the feature extraction module and classify each individual target blob in real time according to pre-trained classification logic, such as Support Vector Machine (SVM), decision tree, or neural network model. The classification results are output to clearly distinguish between solid quinoline insoluble particles and bubbles; the classification logic and discrimination threshold are fixed after training.

[0144] The metering module can be implemented as a data processing unit configured to receive the classification results output by the classification module. This module is designed to count all independent target spots identified as solid quinoline insoluble particles and determine the final quinoline insoluble content based on a preset metering algorithm, such as cumulative calculations based on spot number, area, or volume. Alternatively, this module can be a software program that displays the calculation results on a user interface and transmits data to the production control system, but its metering parameters and conversion factors remain fixed.

[0145] This application constructs a multi-dimensional, intelligent detection system by introducing the collaborative work of an image acquisition module, an image enhancement and processing module, a recognition and separation module, a feature extraction module, a classification module, and a metrology module. In particular, the feature extraction module can comprehensively capture various features such as surface texture, edge complexity, and internal grayscale distribution, while the classification module performs accurate discrimination based on these rich features. This multi-module, multi-feature fusion system architecture enables the system to effectively cope with interference from spherical solid quinoline insoluble particle aggregates and deformed bubbles, significantly improving the accuracy and reliability of QI content detection.

[0146] The advantages of this system are as follows: First, the image acquisition and image enhancement modules provide high-quality image input for subsequent analysis. Second, the recognition and separation modules ensure the independence of the analyzed objects. Most importantly, the combination of the feature extraction and classification modules enables the system to overcome the limitations of single feature judgment and provide more stable and reliable online detection data. Therefore, this system can provide solid technical support for the refined control of the production process and the improvement of product quality, effectively solving the drawbacks of unreliable data and lack of trust in operators in existing technologies.

[0147] The foregoing has provided a detailed description of the preferred embodiments of this application. However, this application is not limited to the above-described embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application. All such equivalent modifications or substitutions are included within the scope defined in this application.

Claims

1. A method for detecting the QI content of modified asphalt based on image analysis, characterized in that, Includes the following steps: The image stream of liquid asphalt is obtained to produce an image sequence; The image sequence is enhanced to obtain an enhanced image sequence; Target spots are separated from the enhanced image sequence, and individual target spots are identified from the target spots, the target spots including solid quinoline insoluble particles and bubbles; For the independent target spots, surface texture features, edge complexity features, and internal grayscale distribution features are extracted; Based on the surface texture features, the edge complexity features, and the internal grayscale distribution features, the independent target spots are classified according to the classification logic to obtain classification results, so as to distinguish between solid quinoline insoluble particles and bubbles; Based on the classification results, the individual target spots identified as solid quinoline insoluble particles are quantified to obtain the quinoline insoluble content.

2. The method according to claim 1, characterized in that, The step of separating the target blob from the enhanced image sequence includes: The enhanced image sequence is binarized to generate a binarized image; When the pixel value of a pixel in the binarized image is less than a preset threshold, the pixel in the binarized image is determined to be a target spot; when the pixel value of a pixel in the binarized image is greater than the preset threshold, the pixel in the binarized image is determined to be an asphalt background.

3. The method according to claim 2, characterized in that, The step of determining a pixel in the binarized image as a target blob when the pixel value of a pixel in the binarized image is less than a preset threshold includes the following: Eight-connected component analysis is performed on the pixels identified as target spots in the binarized image, and all connected pixels with the same label are marked as the same connected component; Pixels in the same connected region are identified as independent target blobs.

4. The method according to claim 1, characterized in that, For the individual target spots, the steps for extracting surface texture features include: For each individual target blob, the image within its region is converted into an eight-bit grayscale image; Calculate the gray-level co-occurrence matrix based on the eight-bit grayscale image; Surface texture features are extracted from the gray-level co-occurrence matrix, and the surface texture features include: energy, contrast, correlation and entropy.

5. The method according to claim 1, characterized in that, The steps for extracting surface texture features, edge complexity features, and internal grayscale distribution features for the independent target spots include: For the individual target spots, an edge detection algorithm is used to extract the edge information of the individual target spot contours; Edge complexity features are extracted from the edge information, including edge length, edge density, and edge roughness.

6. The method according to claim 1, characterized in that, The step of classifying the independent target spots according to classification logic based on the surface texture features, edge complexity features, and internal grayscale distribution features to obtain classification results, in order to distinguish between solid quinoline insoluble particles and bubbles, includes: Track the independent target blobes to obtain their correlation relationships within the enhanced image sequence; Based on the aforementioned correlation, the instantaneous deformation rate, deformation recovery rate, motion trajectory smoothness, acceleration, and rotational characteristics of the independent target spots are extracted. Based on the surface texture features, edge complexity features, internal grayscale distribution features, instantaneous deformation rate, deformation recovery rate, motion trajectory smoothness, acceleration, and rotation characteristics, the independent target spots are classified according to classification logic to obtain classification results, so as to distinguish between solid quinoline insoluble particles and bubbles.

7. The method according to claim 6, characterized in that, The step of tracking the independent target blob and obtaining the correlation relationship of the independent target blob in the enhanced image sequence includes: For all independent target blobs identified in the current frame, calculate the similarity in position, size, and shape between them and the independent target blobs in the previous frame; Based on the similarity, the Hungarian algorithm is used for optimal matching to associate the independent target blobs in the current frame with the independent target blobs in the previous frame, thereby obtaining the association relationship of the independent target blobs in the image sequence.

8. The method according to claim 6, characterized in that, The process of classifying the independent target spots according to classification logic based on the surface texture features, edge complexity features, internal grayscale distribution features, instantaneous deformation rate, deformation recovery rate, motion trajectory smoothness, acceleration, and rotational characteristics to obtain classification results, and then distinguishing between solid quinoline insoluble particles and bubbles, includes: The classification results are continuously monitored to obtain the monitoring statistical distribution; When the monitored statistical distribution deviates from the preset statistical distribution under stable operating conditions, an adaptive adjustment mechanism is triggered, the adaptive adjustment mechanism including: Based on the current working condition data, the weights and discrimination thresholds of the surface texture features, edge complexity features, internal grayscale distribution features, instantaneous deformation rate, deformation recovery rate, motion trajectory smoothness, acceleration, and rotation characteristics in the classification logic are dynamically adjusted. The working condition data includes flow velocity, temperature, and asphalt composition. The classification results are corrected based on the adjusted weights and discrimination thresholds.

9. The method according to claim 8, characterized in that, The steps for continuously monitoring the classification results to obtain the monitoring statistical distribution include: Continuously monitor the classification results; Within a set time window, the classification results of each independent target spot are collected in real time, and the proportion of particles and bubbles classified as solid quinoline insoluble matter within the time window is calculated to obtain short-term statistics. Based on the short-term statistics, the short-term moving average of the classification results is obtained as the monitoring statistical distribution.

10. A modified asphalt QI content detection system based on image analysis, characterized in that, The system includes: The image acquisition module is used to acquire the image stream of liquid asphalt to obtain an image sequence; An image enhancement processing module is used to enhance the image sequence to obtain an enhanced image sequence; The identification and separation module is used to separate target spots from the enhanced image sequence and identify independent target spots from the target spots, the target spots including solid quinoline insoluble particles and bubbles; The feature extraction module is used to extract surface texture features, edge complexity features, and internal grayscale distribution features for the independent target spots. The classification module is used to classify the independent target spots according to the classification logic based on the surface texture features, the edge complexity features, and the internal grayscale distribution features to obtain classification results, so as to distinguish between solid quinoline insoluble particles and bubbles; The metering module is used to measure the independent target spots that are identified as solid quinoline insoluble particles according to the classification results, so as to obtain the quinoline insoluble content.

Images