An intelligent evaluation method for rubber filler dispersion uniformity based on image recognition

Abstract

This invention relates to the field of rubber composite materials technology, and particularly to an intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition. The method includes: acquiring images of multiple regions on the sample surface; extracting image features and performing similarity matching with pre-stored standard feature data to obtain the preliminary dispersion level and matching confidence level of each image; determining the effective dispersion level and weight of each image based on the overall statistical distribution of all preliminary levels, combined with the confidence level and level of each image; calculating a weighted average dispersion level based on the effective level and weight, and determining the final dispersion level according to its mapping relationship with a preset level; simultaneously, calculating uniformity evaluation indicators based on the overall statistical distribution, and locating abnormal dispersion areas based on the effective level and weight; and finally outputting an evaluation report containing the above-mentioned level, indicator, and location information. This invention achieves an objective, comprehensive, and interference-resistant intelligent evaluation of the dispersion uniformity of rubber fillers.

Description

Technical Field

[0001] This invention relates to the field of rubber composite materials technology, and in particular to an intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition. Background Technology

[0002] In the rubber industry, the uniformity of filler dispersion (such as carbon black and silica) in the rubber matrix is ​​one of the key factors determining the physical and mechanical properties, dynamic heat generation characteristics, and durability of the final product. To accurately assess this indicator, a series of testing methods have been developed within the industry. Early evaluations relied primarily on experienced technicians directly observing the cut surfaces of samples using optical microscopes and visually comparing them with standard photographs to provide a qualitative judgment of the dispersion level. With advancements in computer technology, quantitative analysis methods based on digital image processing have gradually become mainstream. Furthermore, some more sophisticated laboratories employ scanning electron microscopy or transmission electron microscopy for nanoscale observation to obtain more detailed information on the filler distribution morphology. These technologies constitute the main technical spectrum for evaluating the dispersion of rubber fillers, from macroscopic qualitative to microscopic quantitative analysis, providing a basis for production process control and product quality assessment.

[0003] Despite the various evaluation methods offered by existing technologies, several pressing technical challenges remain in practical applications, particularly in quantitative analysis based on optical images. First, existing methods are generally susceptible to subjective judgment or the randomness of single observation areas. For example, in manual visual comparison or traditional image analysis, the observation area selected by the operator may not represent the true condition of the entire sample surface, especially when the dispersion itself is uneven. Evaluation results from a single point or a few points can lead to misjudgments of the overall dispersion level. Second, existing image analysis methods are highly sensitive to sample preparation quality (such as cutting marks, surface contamination, and uneven illumination). This noise introduced during sample preparation severely interferes with the accurate extraction of image features, resulting in distorted calculated dispersion parameters and compromising the reliability and repeatability of the evaluation results. Achieving objective, comprehensive, and interference-resistant intelligent dispersion evaluation is a current practical challenge in this technological field.

[0004] Chinese Patent Publication No. CN107505480A discloses a method for detecting the dispersibility of fillers in rubber composite materials. The method involves preparing rubber composite samples using a cryogenic ultramicrotome, then scanning the rubber composite using atomic force microscopy (AFM) in tapping mode, and observing the dispersibility of the fillers in the rubber through morphology and phase diagrams. However, this method for detecting the dispersibility of fillers in rubber composite materials has the following problems: its evaluation results are highly subjective, lack statistical representativeness, and are sensitive to sample preparation defects. Summary of the Invention

[0005] To address this, the present invention provides an intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition, which overcomes the problems of strong subjectivity, insufficient statistical representativeness, and sensitivity to sample preparation defects in the existing technology.

[0006] To achieve the above objectives, this invention provides an intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition, comprising:

[0007] Step S1: Obtain the first images of several different observation areas on the cut surface of the rubber sample under standard observation conditions;

[0008] Step S2: Extract image feature data of each first image respectively, and perform similarity matching calculation between the image feature data of each first image and the standard image feature data corresponding to several pre-stored standard dispersion levels. Determine the preliminary dispersion level and matching confidence level of each first image based on the matching calculation results.

[0009] Step S3: Calculate the overall statistical distribution of the initial dispersion levels based on the initial dispersion levels corresponding to all the first images; and determine the effective dispersion level of each first image and the weight value corresponding to each effective dispersion level based on the overall statistical distribution and the matching confidence and initial dispersion level corresponding to each first image.

[0010] Step S4: Calculate the weighted average dispersion level of the rubber sample based on the effective dispersion level and corresponding weight value of each of the first images; determine the final dispersion level of the rubber sample according to the mapping relationship between the weighted average dispersion level and the preset dispersion level.

[0011] Step S5: Calculate the dispersion uniformity evaluation index of the rubber sample based on the overall statistical distribution; determine the spatial location information of the dispersion anomaly region based on the effective dispersion level and corresponding weight value of each first image.

[0012] Step S6: Output an evaluation report containing the final dispersion level, the dispersion uniformity evaluation index, and the spatial location information of the dispersion anomaly area.

[0013] Furthermore, the image feature data includes frequency domain feature data, texture statistical feature data, and binarized morphological feature data.

[0014] Further, step S2 includes:

[0015] Step S21: Calculate the feature distance between the image feature data of each first image and the standard image feature data corresponding to each standard dispersion level;

[0016] Step S22: Based on the feature distances, the K-nearest neighbor algorithm is used to obtain the preliminary dispersion level corresponding to each of the first images;

[0017] Step S23: Calculate the matching confidence level corresponding to each of the first images based on the minimum feature distance corresponding to the initial dispersion level.

[0018] Further, in step S3, based on the overall statistical distribution and the matching confidence and preliminary dispersion level corresponding to each first image, determining the effective dispersion level of each first image and the weight value corresponding to each effective dispersion level includes:

[0019] Step S31: Calculate the mean of all initial dispersion levels;

[0020] Step S32: When the deviation between the initial dispersion level corresponding to any first image and the mean exceeds a first preset threshold, and the matching confidence level corresponding to the first image is lower than a second preset threshold, the weight value of the initial dispersion level corresponding to the first image is reduced.

[0021] Step S33: Determine the effective dispersion level of each of the first images based on the adjusted weight values ​​of each preliminary dispersion level.

[0022] Furthermore, the image feature data in step S2 includes: frequency domain feature data, texture statistical feature data, and binarized morphological feature data.

[0023] Further, in step S4, calculating the weighted average dispersion level of the rubber sample based on the effective dispersion level and the corresponding weight value of each first image includes: performing a weighted summation calculation based on each effective dispersion level and its corresponding normalized weight value to obtain the weighted average dispersion level.

[0024] Further, in step S4, determining the final dispersion grade of the rubber sample based on the weighted average dispersion grade and the preset dispersion grade mapping relationship includes: mapping the value of the weighted average dispersion grade to a preset discrete grade range to obtain the final dispersion grade in integer or half-integer form.

[0025] Further, in step S5, calculating the dispersion uniformity evaluation index of the rubber sample based on the overall statistical distribution includes: calculating the standard deviation of all preliminary dispersion levels and using the standard deviation as the dispersion uniformity evaluation index.

[0026] Further, in step S5, determining the spatial location information of the dispersion anomaly region based on the effective dispersion level and corresponding weight value of each of the first images includes:

[0027] Step S51: Calculate the mean of all valid dispersion levels;

[0028] Step S52: Identify observation areas where the deviation between the effective dispersion level and the mean exceeds a third preset threshold and the corresponding weight value is higher than a fourth preset threshold.

[0029] Step S53: Record the spatial location of the observation area identified in step S52 as the spatial location information of the dispersed anomaly area.

[0030] Furthermore, step S3 further includes: when there is a first image whose matching confidence is lower than a preset confidence threshold, re-acquiring a second image of the observation area corresponding to the first image, and updating the preliminary dispersion level and the matching confidence corresponding to the observation area based on the second image.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows: by acquiring images of multiple observation areas on the surface of the rubber sample, extracting multi-dimensional image features and matching them with standard grades for similarity, and combining the overall statistical distribution and confidence level to adjust the weights and correct the spatial consistency of the preliminary grades, the present invention finally outputs the weighted average dispersion grade, uniformity evaluation index and abnormal area location information, thereby realizing an objective, comprehensive and interference-resistant intelligent evaluation of the dispersion uniformity of rubber fillers, effectively overcoming the problems of strong subjectivity, insufficient statistical representativeness and sensitivity to sample preparation defects in traditional methods.

[0032] This invention improves the representativeness and reliability of the evaluation results by acquiring images of multiple independent observation areas on the sample surface and performing feature analysis, and by integrating multi-location information for overall statistics, thus avoiding misjudgment caused by the randomness of a single or a few observation areas.

[0033] Furthermore, this invention extracts multi-dimensional image feature data, including frequency domain features, texture statistical features, and binarized morphological features, to comprehensively characterize the dispersion state of fillers from different perspectives, thereby improving the completeness of feature expression and the accuracy of grade matching.

[0034] Furthermore, this invention introduces a matching confidence index and dynamically adjusts the weights based on the deviation between the initial dispersion level and the overall statistical distribution, thereby reducing the evaluation weights of low-confidence or significantly deviated areas and enhancing adaptability to local sample preparation defects and image noise.

[0035] Furthermore, this invention uses the standard deviation of the initial dispersion level as an evaluation index of dispersion uniformity, identifies abnormal areas that significantly deviate from the overall mean under high confidence levels, and records their spatial locations, thereby achieving quantitative evaluation of dispersion uniformity and location of defective areas.

[0036] Furthermore, by triggering image resampling and feature reanalysis of the corresponding observation area when the matching confidence level is found to be below a threshold, this invention achieves proactive verification and updating of low-reliability data, thereby further improving the quality of input data and the credibility of the final evaluation conclusion. Attached Figure Description

[0037] Figure 1 This is a flowchart of the intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition according to the present invention.

[0038] Figure 2 This is a flowchart of step S2 of the intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition in this invention.

[0039] Figure 3 This is a flowchart of step S3 of the intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition in this invention;

[0040] Figure 4 This is a flowchart of step S5 of the intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition in this invention. Detailed Implementation

[0041] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0042] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0043] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0044] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0045] Please see Figure 1 The diagram shows a flowchart of the intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition according to the present invention. The present invention provides an intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition, comprising:

[0046] Step S1: Obtain the first images of several different observation areas on the cut surface of the rubber sample under standard observation conditions;

[0047] In one specific embodiment, sample preparation strictly followed the provisions of Clause 6.3 (applicable to vulcanized rubber) and Clause 6.3.2 (applicable to unvulcanized rubber) of the national standard GB / T 6030-2006. First, the rubber sample was vertically cut using a single-edged razor blade mounted on a dedicated mechanical cutting stage to obtain a fresh, flat observation surface. During cutting, the blade should be ensured to be sharp and free of defects to minimize the interference of cutting marks on subsequent image analysis. For unvulcanized rubber samples, they were pre-heated and pressed between two plastic films at approximately 105°C and 1 kPa for about 5 minutes to remove pores. After appropriate cooling, a slow and steady cut was performed using a blade heated to approximately 100°C to prepare a surface suitable for observation.

[0048] The prepared sample is fixed on a sample support platform, with its freshly cut surface facing the observation optical path. A high-intensity oblique light source with an incident angle of 30° is used to illuminate the sample surface. The direction of the light source should be parallel to the cutting direction to further suppress the appearance of tool marks. An optical system with an effective magnification of 100x, in conjunction with a CCD camera, is used to capture images of the sample surface. To ensure the representativeness of the observation, at least five independent and non-overlapping observation areas are pre-planned on the sample surface or randomly selected. In this embodiment, each area corresponds to a field of view size of 1mm × 1mm. Digital images of each area are acquired sequentially, and these images constitute the "first image" set. Preferably, the image resolution is not less than 2 million pixels, and acquisition should be performed after the system has been preheated and stabilized, and after uniformity and light intensity calibration, to ensure that all images are obtained under consistent "standard observation conditions."

[0049] Step S2: Extract image feature data of each first image respectively, and perform similarity matching calculation between the image feature data of each first image and the standard image feature data corresponding to several pre-stored standard dispersion levels. Determine the preliminary dispersion level and matching confidence level of each first image based on the matching calculation results.

[0050] Please continue reading. Figure 2 As shown, it is a flowchart of step S2 of the intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition of the present invention;

[0051] Specifically, the image feature data in step S2 includes: frequency domain feature data, texture statistical feature data, and binarized morphological feature data.

[0052] In a specific embodiment, the extraction of image feature data in step S2 is implemented as follows:

[0053] First, for each first image obtained in step S1, three sets of image feature data are extracted, including:

[0054] Frequency domain feature data is used to perform a Fast Fourier Transform (FFT) on the image to calculate the energy distribution of its power spectrum in a specific radial frequency range, which is used to characterize the periodicity of the distribution of filler agglomerates or the texture roughness.

[0055] Specifically, the radial frequency range calibration method involves analyzing a series of standard sample images with known dispersion levels to establish the typical size of aggregates. In this embodiment, this ranges from the smallest observable aggregate to the largest common aggregate size, corresponding to the radial frequency in the power spectrum from 5 μm to 50 μm. This correspondence is directly related to the magnification (100×) of the optical system and the pixel size of the image sensor. Ultimately, the specific radial frequency range used for feature extraction is defined as the energy concentration band that most effectively distinguishes different standard dispersion levels. In this embodiment, this corresponds to texture components with a spatial period of 10 to 100 pixels.

[0056] Specifically, the energy integral (or mean) of the power spectrum within the pre-calibrated radial frequency range is calculated, and this value is used as a dimension of the frequency domain characteristics. This value reflects the strength of the periodic structure corresponding to this size range in the image. The better the dispersion, the more uniform and lower the energy in this frequency band; the worse the dispersion, the more likely there are large agglomerates or their aggregation patterns, and energy peaks may appear in this frequency band.

[0057] Texture statistical feature data is used to calculate the gray-level co-occurrence matrix (GLCM) of the image and extract four statistical measures, namely contrast, correlation, energy and homogeneity, to describe the uniformity of filler distribution and texture structure.

[0058] Binarization of morphological feature data: An adaptive thresholding method is used to binarize the image, distinguishing filler aggregates (foreground) from the rubber matrix (background). Then, morphological parameters of the foreground region are calculated, including area, perimeter, roundness, and Euler number, to quantify the size, shape, and number of aggregates. Specifically:

[0059] Image binarization is used to perform an adaptive threshold segmentation algorithm. The preprocessed grayscale image is processed to automatically determine the optimal threshold and segment the image into foreground (filler agglomerates) and background (rubber matrix).

[0060] Morphological post-processing involves performing necessary morphological operations on the initially binarized image to optimize the result, including:

[0061] Denoising: Morphological opening operations (erosion followed by dilation) are applied to eliminate isolated noise points that are too small, which may be image acquisition noise or tiny scratches rather than true clusters.

[0062] Filling holes: Filling the small holes inside the foreground area to ensure that each aggregate area is solid, which facilitates accurate calculation of its morphological parameters.

[0063] Separating Adhesions: When necessary, morphological watershed algorithms or distance transformation combined with thresholding methods can be used to separate slightly adhered aggregates for more accurate counting and measurement.

[0064] Feature parameter calculation: For each independent connected region (i.e., a cluster) in the processed binary image, the following morphological parameters are calculated and extracted. The statistical features of the entire image are then summarized as part of the final feature vector:

[0065] Area, the total number of pixels within each connected region, directly reflects the size of the aggregate;

[0066] Perimeter, the pixel length of the boundary of each connected region, can be used to calculate shape complexity when combined with area;

[0067] Circularity measures how close an aggregate is to a circle; the closer the value is to 1, the more regular the shape.

[0068] The Euler number is the number of all connected regions in an image minus the total number of holes within them. This parameter can simply characterize the "fragmented" or "aggregated" state of a cluster. Within a given region, a large number of isolated small clusters results in a large Euler number, while a small number of large clusters or complex-shaped clusters (which may have holes) results in a small Euler number.

[0069] The extracted three sets of image feature data are combined into a multidimensional feature vector X=[f1, f2, ..., f m ]^T, where m is the total number of features. A standard image feature database corresponding to dispersion levels 1 to 10 in the national standard GB / T 6030-2006 is pre-stored, and each level j (j=1,2,...,10) corresponds to a standard feature vector Sj.

[0070] Specifically, step S2 includes:

[0071] Step S21: Calculate the feature distance between the image feature data of each first image and the standard image feature data corresponding to each standard dispersion level;

[0072] In one specific implementation, the feature vector X of the first image and the feature vector S of each standard level are calculated. j The weighted Euclidean distance D between them j As the feature distance, the calculation formula is as follows:

[0073] ;

[0074] Among them, D j x is the feature distance between the j-th level standard and the current first image, which is dimensionless; the smaller the value, the higher the similarity. i s is the i-th eigenvalue of the current first image feature vector X; ji S is the j-th level standard eigenvector. j The i-th eigenvalue; w i is the weight coefficient of the i-th feature, dimensionless, with a value range of 0.1 to 1.0, preferably 0.5. Its value is determined by feature importance analysis (based on the Relief-F algorithm) based on the discriminative power of this feature in distinguishing different dispersion levels. The purpose is to balance the contributions of features with different dimensions and varying importance to the total distance.

[0075] Step S22: Based on the feature distances, the K-nearest neighbor algorithm is used to obtain the preliminary dispersion level corresponding to each of the first images;

[0076] In a specific implementation, the K nearest neighbor levels with the smallest feature distance Dj in the current first image are obtained through iteration. The level that appears most frequently among these K nearest neighbor levels is then determined as the preliminary dispersion level L corresponding to the current first image. p Where K is the number of nearest neighbors, which is a positive integer ranging from 3 to 7, preferably 5. The purpose is to smooth out potential misjudgments in a single nearest neighbor match due to noise or minor fluctuations in features through a local majority voting mechanism.

[0077] Step S23: Calculate the matching confidence level corresponding to each of the first images based on the minimum feature distance corresponding to the initial dispersion level.

[0078] In a specific implementation, based on the initial dispersion level L obtained in step S21... p The corresponding minimum feature distance D min Calculate the matching confidence C of the current first image. The calculation formula is as follows:

[0079] C=e −α×Dmin ;

[0080] Where C is the match confidence score, which is dimensionless and ranges from (0, 1]. The closer the value is to 1, the higher the match confidence score; D minα is the minimum feature distance; α is the scale attenuation coefficient, with units of D. min The reciprocal of the unit, ranging from 0.5 to 2.0, preferably 1.0, is calibrated by testing on multiple sets of standard images of known levels, with the goal of making the confidence level C reasonably distinguish between "high matching degree" and "low matching degree".

[0081] Understandably, the weighted Euclidean distance formula introduces a weight coefficient w. i This allows features with strong discriminative power to play a larger role in distance calculation, thereby improving matching accuracy. The parameter K in the K-nearest neighbors algorithm is chosen to be a small odd number, which utilizes local similarity information while avoiding overfitting due to an excessively small K value or excessive smoothing due to an excessively large K value. The confidence calculation formula uses a negative exponential form, the principle of which lies in the feature distance D. min The smaller the value, the closer the image is to a certain standard level of "essential features", and the higher the certainty of the matching result should be. The exponential function can map this degree of proximity to a confidence measure between 0 and 1 that is sensitive to change.

[0082] Understandably, different dispersion states of fillers will inevitably manifest in images as measurable patterns such as texture coarseness, the presence or absence of periodicity, and differences in the geometric morphology of aggregates. Frequency domain features capture the periodicity of the distribution and overall roughness, texture statistical features describe the regularity of local gray-level changes, and binarized morphological features directly quantify the geometric properties of the aggregates themselves. These three features, from different dimensions, jointly construct a complete digital characterization of the dispersion state. By calculating the "distance" between the features of the image under test and a series of standard grade features in multidimensional space, the similarity between the image and each standard grade can be quantitatively assessed. The K-nearest neighbor algorithm mimics the decision-making process of "finding the most similar images and seeing which category the majority belongs to" during manual comparison, obtaining a preliminary grade determination in a data-driven manner. Confidence level provides an inherent quality assessment indicator for this preliminary determination. It reflects the "probability" of the match and provides a basis for subsequent steps of weight adjustment and data reliability screening based on confidence level, thus upgrading a simple grade determination into an intelligent analysis process with uncertainty assessment.

[0083] Step S3: Calculate the overall statistical distribution of the initial dispersion levels based on the initial dispersion levels corresponding to all the first images; and determine the effective dispersion level of each first image and the weight value corresponding to each effective dispersion level based on the overall statistical distribution and the matching confidence and initial dispersion level corresponding to each first image.

[0084] Please continue reading. Figure 3 The diagram shows a flowchart of step S3 of the intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition according to the present invention.

[0085] Specifically, in step S3, based on the overall statistical distribution and the matching confidence and preliminary dispersion level corresponding to each first image, determining the effective dispersion level of each first image and the weight value corresponding to each effective dispersion level includes:

[0086] Step S31: Calculate the mean of all initial dispersion levels;

[0087] In a specific implementation, the mean of the initial dispersion level is the arithmetic mean of the initial dispersion levels. Specifically, the arithmetic mean μ_L of the n initial dispersion levels is calculated to characterize the central tendency of the overall dispersion level of the sample.

[0088] ;

[0089] Where μ_L is the mean of the initial dispersion level, and its dimension is the same as that of the dispersion level (usually the dimensionless level number); L_pk is the initial dispersion level of the k-th observation area; and n is the total number of observation areas.

[0090] Step S32: When the deviation between the initial dispersion level corresponding to any first image and the mean exceeds a first preset threshold, and the matching confidence level corresponding to the first image is lower than a second preset threshold, the weight value of the initial dispersion level corresponding to the first image is reduced.

[0091] In one specific implementation, an initial weight W_k' is assigned to the preliminary dispersion level L_pk of each observation region, with an initial value of 1. Subsequently, the weight is dynamically adjusted based on the deviation of the region's level from the overall mean and its matching confidence. The adjustment rule is as follows: if the absolute deviation of the preliminary dispersion level L_pk of an observation region from the overall mean μ_L exceeds a preset first threshold T1, and its matching confidence C_k is lower than a preset second threshold T2, then the rating result of that region is considered unreliable due to local sample preparation defects, image noise, or matching errors, and its weight needs to be reduced.

[0092] The weight adjustment formula is as follows:

[0093] ;

[0094] Wherein, W_k is the adjusted weight of the k-th observation region, dimensionless; W_k' is the initial weight of the region, dimensionless, here set to 1; T1 is the grade deviation threshold, dimensionless, ranging from 1.5 to 2.5 grades, preferably 2.0 grades, based on the empirical assumption that differences spanning two complete grades (e.g., level 5 and level 7) are usually significant; T2 is the confidence threshold, dimensionless, ranging from 0.5 to 0.7, preferably 0.6, the calibration principle being that matching results below this value are considered insufficiently certain; β is the weight decay factor, dimensionless, ranging from 0.1 to 0.4, preferably 0.2, its function is to significantly reduce the influence of a region when it is judged as an anomaly rather than completely eliminating it, thereby achieving a balance between suppressing noise and preserving potential true information.

[0095] Step S33: Based on the adjusted weight values ​​and the corresponding preliminary dispersion levels, calculate the effective dispersion level for each observation area;

[0096] In one specific embodiment, to obtain the dispersion level (effective dispersion level Lek) of each observation area after reliability correction, it is necessary to comprehensively utilize the area's initial dispersion level Lpk, the adjusted weight Wk, and the overall statistical information of all areas. The core principle is to ensure that the weight value Wk is not only used for the subsequent overall weighted average but also has a positive impact on the "robust" estimation of each area's own dispersion level.

[0097] One specific implementation employs a locally weighted average-based strategy to calculate Lek. For the k-th observation region, its effective dispersion level Lek is calculated by considering the preliminary levels of the region and its spatially or sequentially adjacent regions, using the adjusted weights Wi of each region as weighting coefficients. The calculation formula can be expressed as follows:

[0098] ;

[0099] Where, Lek is the effective dispersion level of the k-th observation area, with the dimension of level number; Ωk is the set of observation areas participating in the calculation of the effective level of the k-th area, which usually includes the k-th area itself and several of its neighboring areas; Lpi is the preliminary dispersion level of the i-th area belonging to the set Ωk; and Wi is the weight corresponding to Lpi, adjusted by step S32.

[0100] Understandably, the denominator in this formula normalizes the weights involved in the weighting, ensuring that the calculated Lek is within a reasonable range of levels. The setting of the set Ωk is crucial. If it only includes itself (i.e., Ωk = {k}), the formula degenerates into Lek = Lpk, which fails to reflect the corrective effect of the weight Wk on the "effective level" and fails to utilize spatial context information, making it a less than optimal solution. Preferably, Ωk includes the k-th region and several of its nearest neighbors (e.g., using spatial 4-neighborhoods or 8-neighborhoods, or 1-2 regions before and after the sampled sequence). This is equivalent to smoothing the initial level using a local filter with weights as the confidence index. The effect is that for a region whose weight is reduced (i.e., identified as a low-reliability outlier), the calculation of its Lek will rely more on the levels of its surrounding high-weight regions, thus correcting the "effective level" of that region to better reflect its true neighborhood and effectively suppressing the influence of isolated outliers.

[0101] Understandably, the purpose of this step is to obtain a set of effective grade data that has been screened for reliability and smoothed for spatial consistency, better representing the true dispersion state of each observation area. The principle is that while the dispersion of filler in rubber may exhibit macroscopic inhomogeneity, it usually possesses a certain continuity or gradual change within microscopic local regions. Completely isolated anomalous grade points, distinctly different from their surrounding areas, are more likely caused by non-essential factors such as sample preparation damage, surface contamination, or image analysis errors, rather than representing the true distribution of the filler. Step S32 identifies these suspicious "low-weighted points" using both grade deviation and confidence level criteria. Step S33 goes further, not simply removing or retaining these low-weighted points, but introducing a local weighted average model to reasonably estimate and correct the grade values ​​of these suspicious points using information from high-weighted neighboring points. The advantage of this approach is that it not only reduces the influence of outliers in the "weight" dimension but also repairs the data at the "grade value" level itself, ensuring that any subsequent overall statistical calculations based on these grades (such as mean and standard deviation) are grounded in cleaner and more reliable data. This method combines the ideas of statistical outlier detection and spatial data filtering, which enhances the overall evaluation method against local interference and makes the final evaluation results more accurately reflect the essential characteristics of filler dispersion, rather than accidental observational flaws.

[0102] Specifically, step S3 further includes: when there is a first image whose matching confidence is lower than a preset confidence threshold, re-acquiring a second image of the observation area corresponding to the first image, and updating the preliminary dispersion level and the matching confidence corresponding to the observation area based on the second image.

[0103] In one specific embodiment, step S3 of the present invention further includes a sub-step of reviewing and updating low-confidence observations. This sub-step is executed before or after the main process of steps S31, S32, and S33, and aims to improve the overall reliability of the data basis input to subsequent statistical and weighted analysis through proactive data quality intervention.

[0104] In practice, the system presets a confidence threshold Cth to trigger the review mechanism. After step S2 is completed, the system iterates through the matching confidence Ck corresponding to all first images. If it finds that one or more first images have a matching confidence Ck lower than the preset confidence threshold Cth, it determines that the reliability of the preliminary analysis results of the observation areas corresponding to these images is questionable, and the review process needs to be initiated.

[0105] The confidence threshold Cth is a preset, dimensionless constant ranging from 0.4 to 0.6. Preferably, Cth is set to 0.5. Its determination is based on the fact that the matching confidence Ck is calculated based on the feature matching distance; theoretically, Ck = 1 for a perfect match and approaches 0 for a complete mismatch. Setting the threshold around 0.5 means that when the certainty of the match is below "half," the result is considered unreliable and requires verification. This balances the need for verification costs with the need for data quality assurance.

[0106] For each observation area where the matching confidence level Ck < Cth, the system automatically triggers a re-acquisition command. The image acquisition device (such as a microscope stage movement device) is repositioned to the original coordinate position of that observation area on the sample surface, and a new image, called the second image, is acquired under the same standard observation conditions (including illumination angle, intensity, magnification, etc.). Subsequently, the processing flow of step S2 is completely repeated for this second image: that is, its image feature data is extracted, and similarity matching calculation is performed with the pre-stored standard image feature data to obtain a new preliminary dispersion level Lpknew and a new matching confidence level Cknew based on the second image.

[0107] After obtaining the new results, the system performs an update operation: directly replacing the old Lpk and Ck values ​​corresponding to the observation area in the original database with the newly calculated Lpknew and Cknew. Thereafter, the main process of step S3 (S31, S32, S33) will continue based on this updated complete dataset, which includes both the original high-confidence data and the verified updated data.

[0108] Understandably, the matching confidence level calculated in step S2 is essentially a self-assessment of the inherent reliability of a single image acquisition and analysis process. When this assessment value is too low, it indicates that the observation may have been subject to severe transient interference that is difficult to completely eliminate through subsequent mathematical weighting (e.g., a tiny contaminant on the cut surface happens to be in the field of view, a flicker of the light source at the moment of shooting, or a slight vibration of the stage). In this case, relying solely on subsequent statistical weight adjustments can only reduce its impact, but cannot correct its potentially completely erroneous level value. By setting a clear confidence threshold as an "alarm line," the system can automatically identify these high-risk data points. Triggering re-acquisition and re-analysis is equivalent to performing a "retest" at the physical level, using a new, independent observation to verify or correct the previously questionable results.

[0109] Step S4: Calculate the weighted average dispersion level of the rubber sample based on the effective dispersion level and corresponding weight value of each of the first images; determine the final dispersion level of the rubber sample according to the mapping relationship between the weighted average dispersion level and the preset dispersion level.

[0110] Specifically, in step S4, calculating the weighted average dispersion level of the rubber sample based on the effective dispersion level and the corresponding weight value of each first image includes: performing a weighted summation calculation based on each effective dispersion level and its corresponding normalized weight value to obtain the weighted average dispersion level.

[0111] Specifically, in step S4, determining the final dispersion grade of the rubber sample based on the weighted average dispersion grade and the preset dispersion grade mapping relationship includes mapping the value of the weighted average dispersion grade to a preset discrete grade range to obtain the final dispersion grade in integer or half-integer form.

[0112] In a specific embodiment, firstly, the weight values ​​are normalized and denoted as ω1,ω2,...,ωn, satisfying ω1+ω2+,...,+ωn=1.

[0113] Calculate the weighted average dispersion grade of the rubber sample. Collect the effective dispersion grades Le1, Le2, ..., Len of all n observation areas obtained in step S3, along with their corresponding normalized weight values ​​ω1, ω2, ..., ωn. The weighted average dispersion grade Lˉ of the rubber sample is calculated through weighted summation. Specifically, multiply the effective dispersion grade Lek of each observation area by its corresponding normalized weight ωk, and then sum all n such products. The sum is Lˉ. Here, the weight ωk represents the proportion of the evaluation result of the k-th observation area in the overall evaluation; the larger the value, the more reliable and representative the evaluation result of that area is. Through this weighted averaging operation, areas determined to be highly reliable (i.e., with large weight values) in step S3 will have a greater impact on the overall grade, while the influence of potentially disturbed areas with low weights is weakened, thus allowing Lˉ to more robustly represent the central tendency of the overall dispersion level of the sample.

[0114] Next, the final dispersion grade of the rubber sample is determined. The calculated weighted average dispersion grade Lˉ is a continuous value. To make it compatible with the industry-standard, discrete dispersion grade evaluation system, this implementation adopts grades 1 to 10 as specified in the national standard GB / T 6030-2006, which requires mapping it to a final dispersion grade Lfinal in integer or half-integer form. This is achieved through a preset dispersion grade mapping relationship. This mapping relationship defines the conversion rules from the continuous value Lˉ to the discrete grade Lfinal. In one specific implementation, this mapping relationship can be a preset discrete grade interval lookup table. For example, the continuous grade axis (e.g., from 1 to 10) is divided into several sequentially adjacent intervals, each interval corresponding to a unique final discrete grade value (integer or half-integer). The grade value corresponding to the interval in which the value of Lˉ falls is determined as Lfinal. The interval division can be set according to the actual evaluation accuracy requirements. For example, dividing the interval with a step size of 0.5 levels may produce final grades such as 1.0, 1.5, 2.0, ..., 10.0; if the step size is 1.0 levels, the final grades will be integers from 1 to 10. The principle of mapping is to ensure that Lfinal can most appropriately represent the dispersion quality level indicated by Lˉ, which will not be elaborated further.

[0115] Understandably, the principle behind this implementation lies in the synthesis and transformation of complex analytical data from multiple regions and indicators into a single, clear, and practically consistent conclusive grade. The weighted average calculation is not a simple arithmetic average, but rather an information fusion process based on data reliability. It assigns higher decision weights to local observations with high confidence and consistent with the overall trend, effectively suppressing the negative impact of random errors and local interference on the overall judgment, resulting in a more statistically robust overall grade estimate. Subsequently, mapping this continuous estimate to a discrete grade system is a crucial step in connecting technical evaluation with practical application. The discretized grade output not only conforms to existing standards and industry professionals' cognitive habits, facilitating result communication and comparison, but also, through reasonable interval mapping rules, avoids over-interpreting minor fluctuations in continuous calculated values ​​while maintaining sufficient evaluation accuracy, making the final evaluation result more stable and practical. The entire process embodies the quantitative integration logic from microscopic image features to macroscopic quality judgment, ultimately transforming the refined numerical results obtained by intelligent algorithms into quality grade conclusions that can guide production practice.

[0116] Step S5: Calculate the dispersion uniformity evaluation index of the rubber sample based on the overall statistical distribution; determine the spatial location information of the dispersion anomaly region based on the effective dispersion level and corresponding weight value of each first image.

[0117] Please continue reading. Figure 4 The flowchart shown is a step S5 of the intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition of the present invention.

[0118] Specifically, in step S5, calculating the dispersion uniformity evaluation index of the rubber sample based on the overall statistical distribution includes: calculating the standard deviation of all preliminary dispersion levels and using the standard deviation as the dispersion uniformity evaluation index.

[0119] Specifically, in step S5, determining the spatial location information of the dispersion anomaly region based on the effective dispersion level and corresponding weight value of each of the first images includes:

[0120] Step S51: Calculate the mean of all valid dispersion levels;

[0121] In one specific embodiment, the arithmetic mean of the effective dispersion levels of each observation area obtained in step S3 is calculated. This mean represents the global center of the sample dispersion level after reliability correction.

[0122] Step S52: Identify observation areas where the deviation between the effective dispersion level and the mean exceeds a third preset threshold and the corresponding weight value is higher than a fourth preset threshold.

[0123] In a specific embodiment, two thresholds are set: a third preset threshold (used to determine the significance of grade deviation) and a fourth preset threshold (used to determine the reliability of the results). All observation areas are traversed to identify areas that simultaneously meet the following two conditions: First, the absolute value of the difference between the effective dispersion grade of the area and the overall mean calculated in step S51 exceeds the third preset threshold; second, the normalized weight value corresponding to the area (derived from step S3) is higher than the fourth preset threshold. The third preset threshold is typically set based on the sensitivity of actual quality control to grade differences, for example, it can be 1.0 to 1.5 grades. The fourth preset threshold is used to filter results with high confidence, for example, it can be the median of the weight values ​​of all areas or a high quantile (e.g., 0.7). The logic behind setting these two conditions is that only when the dispersion grade of a certain area significantly deviates from the overall level (condition one), and the evaluation result of that area itself has a high reliability weight (condition two), can we reasonably be certain that the anomaly in that area is caused by genuine uneven packing dispersion, rather than measurement noise or interference, thus identifying it as a "dispersion anomaly area" requiring attention.

[0124] Step S53: Record the spatial location of the observation area identified in step S52 as the spatial location information of the dispersed anomaly area.

[0125] In one specific embodiment, for each dispersed anomalous region identified in step S52, its spatial coordinates on the original sample surface are recorded. These coordinates correspond to the pre-defined observation area grid or sampling point sequence in step S1. The recorded information may include the region number, its relative position in the sample coordinate system (such as the distance from the edge), etc., thus constituting complete spatial location information of the dispersed anomalous region.

[0126] Understandably, the principle behind this implementation lies in conducting a more comprehensive spatial statistical analysis of the packing dispersion state, going beyond a single "average grade" evaluation. Calculating the standard deviation of the initial dispersion grade helps to grasp the uniformity of dispersion from a macroscopic statistical perspective. Poorly dispersed samples not only have a low overall grade but are often accompanied by large fluctuations in grades across different regions (high standard deviation). This provides a key quality dimension that complements the overall average grade. Identifying abnormal regions, on the other hand, concretizes the macroscopic inhomogeneity indicators into microscopic spatial locations. Its core technical idea employs a dual criterion of "significant deviation under high confidence." This is related to, yet distinct from, the logic of reducing the weight of "significant deviation points with low confidence" in step S3. The purpose of step S3 is to suppress the influence of unreliable data on the overall average, a kind of "defensive" data cleaning; while the purpose of step S5 is to proactively discover those real local defects supported by sufficient evidence. A region's effective grade significantly deviates from the overall mean, indicating an abnormal state; at the same time, its high weight indicates a high confidence level in this anomaly judgment, which is very likely to reflect real process or material problems (such as uneven mixing, local overheating, etc.). By recording the location of these areas, quality evaluation can be advanced from "what grade" and "whether it is uniform" to "where it is uneven," providing direct and objective data support for tracing the source of problems in the production process and making targeted process adjustments. This extends the evaluation results from grade determination to problem diagnosis.

[0127] Step S6: Output an evaluation report containing the final dispersion level, the dispersion uniformity evaluation index, and the spatial location information of the dispersion anomaly area.

[0128] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for intelligently evaluating the dispersion uniformity of rubber fillers based on image recognition, characterized in that, include: Step S1: Obtain the first images of several different observation areas on the cut surface of the rubber sample under standard observation conditions; Step S2: Extract image feature data of each first image respectively, and perform similarity matching calculation between the image feature data of each first image and the standard image feature data corresponding to several pre-stored standard dispersion levels. Determine the preliminary dispersion level and matching confidence level of each first image based on the matching calculation results. Step S3: Calculate the overall statistical distribution of the initial dispersion levels based on the initial dispersion levels corresponding to all the first images; and determine the effective dispersion level of each first image and the weight value corresponding to each effective dispersion level based on the overall statistical distribution and the matching confidence and initial dispersion level corresponding to each first image. Step S4: Calculate the weighted average dispersion level of the rubber sample based on the effective dispersion level and corresponding weight value of each of the first images; determine the final dispersion level of the rubber sample according to the mapping relationship between the weighted average dispersion level and the preset dispersion level. Step S5: Calculate the dispersion uniformity evaluation index of the rubber sample based on the overall statistical distribution; determine the spatial location information of the dispersion anomaly region based on the effective dispersion level and corresponding weight value of each first image. Step S6: Output an evaluation report containing the final dispersion level, the dispersion uniformity evaluation index, and the spatial location information of the dispersion anomaly area.

2. The intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition according to claim 1, characterized in that, The image feature data includes frequency domain feature data, texture statistical feature data, and binarized morphological feature data.

3. The intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition according to claim 2, characterized in that, Step S2 includes: Step S21: Calculate the feature distance between the image feature data of each first image and the standard image feature data corresponding to each standard dispersion level; Step S22: Based on the feature distances, the K-nearest neighbor algorithm is used to obtain the preliminary dispersion level corresponding to each of the first images; Step S23: Calculate the matching confidence level corresponding to each of the first images based on the minimum feature distance corresponding to the initial dispersion level.

4. The intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition according to claim 3, characterized in that, In step S3, based on the overall statistical distribution and the matching confidence and preliminary dispersion level corresponding to each first image, determining the effective dispersion level of each first image and the weight value corresponding to each effective dispersion level includes: Step S31: Calculate the mean of all initial dispersion levels; Step S32: When the deviation between the initial dispersion level corresponding to any first image and the mean exceeds a first preset threshold, and the matching confidence level corresponding to the first image is lower than a second preset threshold, the weight value of the initial dispersion level corresponding to the first image is reduced. Step S33: Determine the effective dispersion level of each of the first images based on the adjusted weight values ​​of each preliminary dispersion level.

5. The intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition according to claim 4, characterized in that, The image feature data in step S2 includes: frequency domain feature data, texture statistical feature data, and binarized morphological feature data.

6. The intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition according to claim 5, characterized in that, In step S4, calculating the weighted average dispersion level of the rubber sample based on the effective dispersion level and corresponding weight value of each of the first images includes: performing a weighted summation calculation based on each effective dispersion level and its corresponding normalized weight value to obtain the weighted average dispersion level.

7. The intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition according to claim 6, characterized in that, In step S4, determining the final dispersion grade of the rubber sample based on the weighted average dispersion grade and the preset dispersion grade mapping relationship includes mapping the value of the weighted average dispersion grade to a preset discrete grade range to obtain the final dispersion grade in integer or half-integer form.

8. The intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition according to claim 7, characterized in that, In step S5, calculating the dispersion uniformity evaluation index of the rubber sample based on the overall statistical distribution includes: calculating the standard deviation of all preliminary dispersion levels and using the standard deviation as the dispersion uniformity evaluation index.

9. The intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition according to claim 8, characterized in that, In step S5, determining the spatial location information of the dispersion anomaly region based on the effective dispersion level and corresponding weight value of each of the first images includes: Step S51: Calculate the mean of all valid dispersion levels; Step S52: Identify observation areas where the deviation between the effective dispersion level and the mean exceeds a third preset threshold and the corresponding weight value is higher than a fourth preset threshold. Step S53: Record the spatial location of the observation area identified in step S52 as the spatial location information of the dispersed anomaly area.

10. The intelligent evaluation method for the dispersion uniformity of rubber fillers based on image recognition according to claim 9, characterized in that, Step S3 further includes: when there is a first image whose matching confidence is lower than a preset confidence threshold, re-acquiring a second image of the observation area corresponding to the first image, and updating the preliminary dispersion level and the matching confidence corresponding to the observation area based on the second image.

Images