Sand grading detection method based on sand two-dimensional image features

Abstract

The invention discloses a sand grading detection method based on sand two-dimensional image features, and relates to the technical field of sand grading detection. According to the method, the particle size classification based on the neural network model is applied and combined with threshold division, so that the sand particle size division accuracy can be effectively improved, and the classification accuracy can be greatly improved by combining the particle size classification based on the neural network model and the threshold division as the interval emphasis of the particle size classification is different; according to the method, in the classification process of applying the neural network model, the data set is constructed through the shot single-stage sand grain image and the extracted sand grain feature information of the single-stage sand grain image, the problem that the data set is difficult to mark due to the fact that the sand grains are small and large in number is solved, and the training cost of the network model is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of sand gradation detection, and specifically to a sand gradation detection method based on two-dimensional image features of sand. Background Art

[0002] Currently, the sieving method is the main method for detecting sand gradation. Sand grains in a sand sample are separated through a series of sieves with different pore sizes, and the gradation of the sand is determined according to the retained amounts on each sieve. However, the traditional manual sieving method has a large workload and is greatly affected by human subjective factors. Although the automated vibrating sieving method reduces manual operation, it is difficult to accurately judge the stop condition of sieving during the vibration process, which may damage the original particle size distribution of the sand. In addition, the sieving accuracy is also easily affected significantly by the shape of sand grains. At the same time, the sieving method has limitations in that it cannot obtain particle size and particle shape parameters simultaneously and is difficult to achieve automatic detection. In recent years, with the development of digital image processing technology, new ways have been opened up for comprehensively studying and characterizing aggregate particle shape. Using two-dimensional images to characterize and evaluate aggregate particle shape characteristics is the trend of future research, and the image-based method has become an important way to improve the gradation detection problem.

[0003] To achieve the automated detection of sand gradation through image recognition technology, the key lies in the characterization of sand grain shape and particle size and the gradation analysis. Scholars at home and abroad have conducted a large number of studies in this field. Ding reconstructed the three-dimensional model of asphalt concrete through two-dimensional slice images, demonstrating the similarity relationship between the gradation distribution of aggregates in a two-dimensional cross-section and the true gradation distribution in a three-dimensional state. Leonardo realized the image-based gradation detection of asphalt concrete based on the thresholding and watershed algorithms. Chao Lu obtained sand grain images of sandy soil through a scanning electron microscope and achieved gradation recognition below the millimeter level based on the thresholding algorithm. Mu Yu established a rapid detection system for the gradation of earth and stone materials based on the thresholding and edge detection algorithms and achieved rapid industrial-grade gradation detection.

[0004] Regarding the characterization of the particle shape and particle size of sand grains, Kumara used digital image processing technology to measure the grain size of gravel, and selected the minor axis of the equivalent ellipse as the particle size. This measurement result is larger than that of the sieving method (SM). By using the Feret minor axis of the equivalent ellipse to describe the sand grains in the two-dimensional image, the research shows that this method can make the results of the image method closer to those of the sieving method. Liu used an image analysis system to detect the shape of manufactured sand grains and found that the aspect ratio of the equivalent ellipse is the best index for evaluating the shape of sand grains. Jianhua Zhou explored the method of aggregate particle size characterization parameters, and the results showed that the combination of the projected area and the Feret minor axis of the equivalent ellipse has the highest correlation between the equivalent volume and the actual volume. Weijun Fan used principal component analysis and probabilistic neural network to classify aggregate sand grains, and then calculated the aggregate gradation, pointing out that volume parameters of sand grains need to be obtained when using the image method for gradation detection. However, since volume is a three-dimensional feature, it is difficult to accurately obtain this information solely relying on two-dimensional images. Therefore, Li Liang used three-dimensional scanning technology to study the appearance geometric characteristics of aggregates and concrete. In addition, Su used microfocus X-ray CT scanning technology to obtain high-resolution three-dimensional aggregate morphology and used spherical harmonic functions to describe its morphological characteristics. The research of the aforementioned scholars shows that there are certain limitations in the results obtained by estimating the three-dimensional volume only by multiplying the equivalent area and the equivalent Feret minor axis in two-dimensional features.

[0005] With the wide application of artificial intelligence, in order to quickly characterize sand grains and detect gradation, Siyao addressed the problems existing in traditional grain size measurement methods based on an improved U-net architecture and designed a special loss function. mu Yu proposed an intelligent detection of soil and stone material gradation based on a depth threshold convolution model, and realized fast detection of soil and stone material gradation based on images through edge detection and convolution network models. Yuxuan Li studied the prediction method of aggregate gradation based on machine learning. For the classification problem of aggregate grade categories, an improved ResNet50 classification algorithm was used for model fine-tuning and training. Feizhi Huang proposed an online gradation detection of manufactured sand based on deep learning. The Mask R-CNN instance segmentation model was applied to effectively segment complete sand grains in the stacked manufactured sand scenario, realizing automatic detection of the conveyor belt sand gradation.

[0006] In the above-mentioned methods, Leonardo, Chao Lu, Mu Yu and others generally rely on traditional thresholding-based techniques in image recognition methods. Although this method is computationally efficient and has a fast recognition speed, due to the limitations of its fixed threshold, it is vulnerable to interference such as image noise and illumination changes, resulting in insufficient accuracy in complex situations. Especially in applications with high-precision requirements, there is still a large room for improvement in its accuracy and stability. In terms of the characterization of the grain shape and particle size of sand grains, although many scholars have proved that the application of three-dimensional scanning of sand grain characteristics can improve the detection accuracy, the acquisition cost of three-dimensional image data is relatively high, and in practical applications, it is difficult to process a large number of samples and small particle size sand grains. In the application of deep learning models, although the detection accuracy and efficiency of sand gradation have been greatly improved, training a high-quality deep learning model requires a large amount of data, and the data annotation process of manufactured sand images is complex, and currently no publicly available and suitable high-quality related datasets have been found. Therefore, the lack of training data has become the biggest challenge faced by deep learning methods in this field.

[0007] Therefore, a new solution needs to be proposed for the above problems. Summary of the Invention

[0008] The purpose of the present invention is to provide a method for detecting sand gradation based on two-dimensional image features of sand to solve the technical problems proposed in the background art.

[0009] To achieve the above purpose, the present invention provides the following technical solution: A method for detecting sand gradation based on two-dimensional image features of sand, at least including the following steps:

[0010] S1: Set up experimental equipment and take images of sand grains for preliminary screening of sand grains;

[0011] S2: Adopt a threshold division method and a classification method based on a neural network model to enhance the stability of classification through sand grain classification;

[0012] S3: Design a method for characterizing the volume of sand grains to finally obtain the optimal sand gradation detection result.

[0013] Further, the experimental equipment in S1 at least includes a feeding bin, a servo motor, a pressure sensor, a flexible vibrating disk and an industrial camera. The industrial camera is located 23.8 cm directly above the flexible vibrating disk, and the light source is located directly below the flexible vibrating disk, and backlight is used to take images of sand grains. The main size range of the sand grains used for screening by the experimental equipment is 0.075 - 4.75 mm.

[0014] Further, the threshold division method in S2 sets six division thresholds according to the sieve hole size of the sieve used in manual screening to classify sand grains into seven intervals.

[0015] Furthermore,

[0016] The classification method based on the neural network model in S2 at least includes the following steps:

[0017] Construct two four-class neural network models to divide and fuse the four intervals of sand grains respectively;

[0018] The four-class neural network model consists of two four-class network structures, the upper and lower layers. At the same time, each network structure includes an input layer, a hidden layer, and an output layer;

[0019] The two four-class neural network models are used to divide the seven particle size intervals into two four-class intervals;

[0020] When the two-dimensional sand grain feature information is input into the two four-class neural network models, the two four-class neural network models simultaneously output the sand grain categories in the 0.6 - 1.18 mm interval for comparison and verification, thereby improving the classification accuracy;

[0021] At the same time, the two four-class neural network models can also output the sand grain categories in the other six particle size intervals.

[0022] Furthermore, the sand grain sizes in the four-class intervals include 0.075 - 0.15 mm, 0.15 - 0.3 mm, 0.3 - 0.6 mm, 0.6 - 1.18 mm, as well as sand grains of 0.6 - 1.18 mm, 1.18 - 2.36 mm, 2.36 - 4.75 mm, and above 4.75 mm.

[0023] Furthermore, the input layer is responsible for receiving the normalized training sample data and transmitting it to the hidden layer. The number of neurons in the input layer is equal to the number of sand grain features extracted;

[0024] The hidden layer consists of five fully connected layers;

[0025] The number of neurons in each network output layer is 4, representing the probability of the sand grain belonging to the particle size interval, and the two networks simultaneously predict the category of the fourth particle size interval.

[0026] Furthermore, S3 at least includes the following steps:

[0027] The grading detection of sand requires measuring the mass of sand grains in each particle size interval. However, the detection method based on two-dimensional images cannot directly obtain the mass of sand grains. In materials science, the density of sand grains in the same batch of materials is usually approximately the same. Therefore, its volume ratio is basically the same as the mass ratio. Therefore, the grading calculation of sand can be transformed into calculating the volume ratio of sand grains of different particle sizes, that is, converting the mass ratio into the volume ratio, as shown in Equation (1);

[0028]

[0029] where a i represents the percentage of the mass of the i-th particle size range in the total mass, m i represents the mass of the i-th particle size range, v i represents the volume of the i-th particle size range. Therefore, the three-dimensional sand grain volume can be accurately characterized by two-dimensional sand grain characteristics, effectively improving the detection accuracy of sand gradation;

[0030] The parameters of the equivalent ellipse Feret minor axis are optimized to ensure more accurate characterization of the equivalent volume of three-dimensional sand grains;

[0031] For the threshold division and the classification method based on the neural network model, the optimized expressions of the equivalent ellipse Feret minor axis for calculating the sand grain volume characterization are obtained through experiments, as shown in formulas (2) and (3);

[0032] h1 = a1d 2 + b1d + c1 (10)

[0033] h2 = a2d 2 + b2d + c2 (11)

[0034] where h1 is the optimized value of the equivalent ellipse Feret minor axis for sand grains in threshold classification, and a1, b1, c1 are the optimized parameters of this optimized expression; h2 is the optimized value of the equivalent ellipse Feret minor axis for sand grains in neural network classification, and a2, b2, c2 are the optimized parameters of this optimized expression. d is the equivalent ellipse Feret minor axis;

[0035] The calculation of the total volume of sand grains in each interval is shown in formulas (4) and (5);

[0036]

[0037] represents the total volume of the i-th interval to which the sand grains applied with threshold division belong, represents the total volume of the i-th interval to which the sand grains applied with neural network classification belong, s j refers to the projected area corresponding to the j-th sand grain;

[0038] For the volume of each obtained interval, the percentage of the volume of each interval in the total volume can be obtained by applying formula (3);

[0039] Furthermore, the AMPL optimization algorithm is further used to optimize the parameters of the equivalent ellipse Feret minor axis. Since the sand grains are relatively fine, it is challenging to measure the volume of each sand grain individually. Therefore, by comparing the true value of the grading of the standard 500g manufactured sand with the cumulative volume summation result calculated using the optimized expression of the equivalent ellipse Feret minor axis, the volume proportion of each sand grain interval is analyzed, and thus the parameters of the optimized expression are improved, as specifically shown in formulas (6) and (7);

[0040]

[0041] Wherein y in the formula i represents the true grading value of the i-th interval of the standard 500g manufactured sand;

[0042] Through these two formulas, the optimized parameter expressions of the equivalent ellipse Feret minor axis for threshold division and neural network classification can be finally obtained respectively.

[0043] After experimental verification, for the final detection result of the sand grading, when applying the threshold division method, the proportion of sand grains in the interval of 0.15 - 0.3mm is relatively high, while the proportion in the interval of 1.18 - 2.36mm is relatively low. On the contrary, the detection result of the neural network model shows the opposite trend;

[0044] Therefore, by combining the detection results of the two methods and taking the sum average, the optimal sand grading detection result is finally obtained, as specifically shown in formula (8);

[0045]

[0046] Wherein refers to the volume proportion of the i-th interval of the sand grains obtained by threshold division, refers to the volume proportion of the i-th interval of the sand grains obtained by neural network classification, refers to the volume proportion of the i-th interval after taking the average of the two.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0048] 1. By applying particle size classification based on a neural network model and combining it with threshold division, the present invention can effectively improve the accuracy of particle size division of sand grains. Since the particle size classification of the neural network model and threshold division focus on different intervals of particle size classification, the combination of the two can greatly improve the classification accuracy;

[0049] 2. During the process of applying the neural network model for classification, by taking pictures of the single-sized sand grain images and extracting the sand grain feature information of the single-sized sand images to construct a data set, the present invention solves the problem that it is difficult to label the data set due to the small and numerous sand grains, and reduces the training cost of the network model;

[0050] 3. The present invention optimizes the equivalent elliptical Feret minor axis through the AMPL optimization algorithm to enhance the calculation of volume characterization. That is, in the calculation of the equivalent sand grain volume, the optimization algorithm is used to improve the equivalent elliptical Feret minor axis, and the equivalent elliptical Feret minor axis is optimized into a quadratic function expression, significantly improving the characterization accuracy of the volume of fine sand grains. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0052] Figure 1 It is a diagram of the experimental equipment of the present invention;

[0053] Figure 2 It is a neural network structure diagram of the present invention;

[0054] Figure 3 It is an analysis diagram of the cumulative gradation error and fineness modulus error of 10 groups of manufactured sand of the present invention;

[0055] Figure 4 It is an analysis diagram of the gradation detection error of 10 groups of manufactured sand of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0056] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments.

[0057] A method for detecting the gradation of sand based on the two-dimensional image features of sand at least includes the following steps:

[0058] S1: Set up the experimental equipment and take pictures of sand grains for preliminary screening of sand grains;

[0059] S2: Adopt the threshold division method and the classification method based on the neural network model to enhance the stability of classification through sand grain classification;

[0060] S3: Design a method for characterizing the volume of sand grains, and finally obtain the optimal sand gradation detection result.

[0061] Please refer to Figure 1, the experimental equipment in S1 includes at least a feeding bin, a servo motor, a pressure sensor, a flexible vibrating disk, and an industrial camera. The industrial camera is located 23.8 cm directly above the flexible vibrating disk, and the light source is located directly below the flexible vibrating disk, and backlighting is used to capture the images of sand grains. The main size range of the sand grains used for screening by the experimental equipment is 0.075 - 4.75 mm.

[0062] The threshold division method in S2 sets six division thresholds according to the sieve hole sizes of the sieves used in manual screening, so as to classify the sand grains into seven intervals.

[0063] When classifying sand grains based on the neural network model, since the sand grain size intervals increase in a multiple form and there are too many classification types, after using the maximum-minimum normalization, there are still differences in the proportion ranges of different grain size intervals. Using a seven-classification network model will greatly increase the problem of uneven distribution of feature proportions, resulting in low classification efficiency, and thus significantly affecting the training accuracy of the neural network model.

[0064] The classification method based on the neural network model in S2 includes at least the following steps:

[0065] Construct two four-classification neural network models, which are used to divide and fuse four intervals of sand grains respectively. The specific structure is as Figure 2 shown.

[0066] The four-classification neural network model consists of two four-classification network structures, the upper and lower layers. At the same time, each network structure includes an input layer, a hidden layer, and an output layer;

[0067] The two four-classification neural network models are used to divide the seven particle size intervals into two four-classification intervals;

[0068] When inputting the two-dimensional sand grain feature information into the two four-classification neural network models, the two four-classification neural network models simultaneously output the sand grain categories in the interval of 0.6 - 1.18 mm for comparison and verification, so as to improve the classification accuracy;

[0069] At the same time, the two four-classification neural network models can also output the sand grain categories of the other six particle size intervals. The experimental results show that compared with applying a seven-classification network model, the two four-classification networks can effectively reduce the influence brought by the multiple increase of particle size intervals, and by classifying the same gradation multiple times, the classification accuracy and stability of sand grains can be improved.

[0070] The sand grain sizes in the four-classification intervals include 0.075 - 0.15 mm, 0.15 - 0.3 mm, 0.3 - 0.6 mm, 0.6 - 1.18 mm, and 0.6 - 1.18 mm, 1.18 - 2.36 mm, 2.36 - 4.75 mm, and sand grains above 4.75 mm.

[0071] The input layer is responsible for receiving the normalized training sample data and transmitting it to the hidden layer. The number of neurons in the input layer is equal to the number of sand grain features extracted.

[0072] The hidden layer consists of five fully connected layers.

[0073] The number of neurons in each network output layer is 4, representing the probability of the sand grain belonging to a particle size range, and the two networks simultaneously predict the fourth particle size range category.

[0074] S3 includes at least the following steps:

[0075] The grading detection of sand requires measuring the mass of sand grains in each particle size range. However, the detection method based on two-dimensional images cannot directly obtain the mass of sand grains. In materials science, the density of sand grains in the same batch of materials is usually approximately the same. Therefore, its volume ratio is basically the same as the mass ratio. Therefore, the grading calculation of sand can be transformed into calculating the volume ratio of sand grains of different particle sizes, that is, converting the mass ratio into the volume ratio, as shown in Equation (1);

[0076]

[0077] where a i represents the percentage of the mass of the i-th particle size range in the total mass, m i represents the mass of the i-th particle size range, and v i represents the volume of the i-th particle size range. Therefore, accurately characterizing the three-dimensional sand grain volume with two-dimensional sand grain features can effectively improve the accuracy of sand grading detection;

[0078] Existing equivalent volume characterization methods include equivalent ellipsoids, equivalent cylinders, and equivalent spheres, etc. Research shows that in many cases, the equivalent ellipsoid model can more accurately reflect the true volume of particles compared to equivalent spheres and equivalent cylinders, especially when the particle shape is irregular. Because the ellipsoid describes the particle shape through three principal axes and has higher adaptability than a single sphere or cylinder model. However, recent research also shows that by accurately calculating the equivalent projected area of particles and combining the equivalent cylinder model to estimate the volume, more accurate volume estimates can be obtained in some cases. The equivalent Feret minor axis, as an important parameter describing the particle contour, can effectively represent the size characteristics of particles in some cases, but it cannot fully and accurately reflect the change in height in the cylinder model. Therefore, simply estimating the height of the equivalent cylinder through the Feret minor axis may lead to deviations in volume calculation.

[0079] The parameters of the equivalent elliptical Feret minor axis are optimized to ensure more accurate characterization of the equivalent volume of three-dimensional sand grains;

[0080] For the threshold division and the classification method based on the neural network model, the optimized expressions of the equivalent ellipse Feret minor axis for the calculation of sand grain volume characterization are obtained through experiments, as shown in formulas (2) and (3);

[0081] h1 = a1d 2 + b1d + c1 (18)

[0082] h2 = a2d 2 + b2d + c2 (19)

[0083] Where h1 is the optimized value of the equivalent ellipse Feret minor axis of sand grains for threshold classification, and a1, b1, c1 are the optimized parameters of this optimized expression; h2 is the optimized value of the equivalent ellipse Feret minor axis of sand grains for neural network classification, and a2, b2, c2 are the optimized parameters of this optimized expression. d is the equivalent ellipse Feret minor axis;

[0084] The calculation of the total volume of sand grains in each interval is as shown in formulas (4) and (5);

[0085]

[0086] represents the total volume of the i-th interval to which the sand grains applying threshold division belong, represents the total volume of the i-th interval to which the sand grains applying neural network classification belong, and sj refers to the projected area corresponding to the j-th sand grain;

[0087] For the volume of each obtained interval, applying formula (3) can obtain the percentage of the volume of each interval in the total volume;

[0088] Furthermore, the AMPL optimization algorithm is used to optimize the parameters of the equivalent ellipse Feret minor axis. Since the sand grains are relatively small, there are great challenges in measuring the volume of each sand grain individually. Therefore, by comparing the true value of the standard 500g machine-made sand gradation with the cumulative sum result of the volume calculated by applying the optimized expression of the equivalent ellipse Feret minor axis, the volume proportion of each sand grain interval is analyzed, so as to improve the parameters of the optimized expression, as shown in formulas (6) and (7) specifically;

[0089]

[0090] Where in the formula y i represents the true value of the i-th interval gradation of the standard 500g machine-made sand;

[0091] Through these two formulas, the optimized parameter expressions of the equivalent ellipse Feret minor axis for threshold division and neural network classification can be finally obtained respectively.

[0092] After experimental verification, for the final detection result of sand gradation, when applying the threshold division method, the proportion of sand grains in the range of 0.15 - 0.3 mm is relatively high, while the proportion in the range of 1.18 - 2.36 mm is relatively low. On the contrary, the detection result of the neural network model shows the opposite trend;

[0093] Therefore, by combining the detection results of the two methods and through summation averaging, the optimal sand gradation detection result is finally obtained, as shown in Equation (8);

[0094]

[0095] where refers to the volume proportion of the i-th interval of sand grains divided by the threshold, refers to the volume proportion of the i-th interval of sand grains classified by the neural network, refers to the volume proportion of the i-th interval after taking the average of the two.

[0096] Specifically, the following experimental comparisons are proposed:

[0097] According to the standard of construction sand, the present invention collected image data of machine-made sand with 3 specifications and a particle size range of 0.075 - 4.75 mm for experimental research. A total of 10 samples were collected, with each sample weighing about 500 g, to verify the detection result of sand gradation. At the same time, 400 pictures were collected for each single gradation sample to extract the characteristic information of sand grains in different particle size intervals for network model training. To verify the feasibility of the sand grain gradation calculation method, the following three experiments were carried out: (1) Comparative experiment on sand grain particle size classification based on the combination of neural network and threshold division, (2) Optimization experimental analysis of three-dimensional volume calculation according to two-dimensional sand grain characteristics, (3) Analysis of the detection result of sand gradation. The machine-made sand used in the experiment is limestone. First, an industrial camera was used to photograph the sand grains scattered by the vibrating disk. Subsequently, in the sand images after image processing, the characteristic information of sand grains was extracted. The experiment was carried out based on the characteristic information of sand grains.

[0098] 1) Sand grain particle size classification experiment

[0099] In the sand particle classification experiment, for 7 particle size ranges, the present invention constructed two four-classification network models for training. For sand particles in the range of 0.075 - 0.6 mm, 10,000 pieces of sand particle feature information were extracted for each particle size range; while for sand particles in the range of 0.6 - 4.75 mm, 4,300 pieces of feature information were extracted for each particle size range. In the constructed feature information dataset, 80% was used for network model training and 20% was used for testing. The trained neural network classified the sand particle sizes of the test samples. At the same time, the present invention set the screening threshold according to the aperture size of the artificial screening tool's sieve holes, and for sand particles in the range of 0.075 - 4.75 mm, 4,300 pieces of sand particle feature information were extracted for each range for screening test. The relevant results are shown in Table 2. Among them, the results show that the network training error is mainly concentrated in sand particles with a size of 0.15 - 0.6 mm, and the overall test accuracy reaches 97.4%. For the threshold division, the error in particle size classification is mainly reflected in misclassifying sand particles in this particle size range into the previous particle size range. The overall test accuracy of the threshold division is 89.6%.

[0100] Although in terms of sand particle classification accuracy, the classification effect of the network model is much higher than that of the threshold division, it can be seen from the results in Table 1 that for the particle size classification test based on the neural network model, the probability of misclassifying 0.15 - 0.3 mm into 0.3 - 0.6 mm is 3.97%, and the probability of misassigning 0.3 - 0.6 mm to 0.15 - 0.3 mm is 3.70%. The sand particles classified in the 0.15 - 0.3 mm particle size range are relatively reduced. The probability of misclassifying 1.18 - 2.36 mm into 2.36 - 4.75 mm is 0.05%, but the sand particles misclassified from 2.36 - 4.75 mm into 1.18 - 2.36 mm are 1.50%. The sand particles classified in the 1.18 - 2.36 mm particle size range are relatively increased.

[0101] For the particle size classification by threshold division, the probability of misclassifying 0.15 - 0.3 mm into 0.3 - 0.6 mm is 2.36%, and the probability of misassigning 0.3 - 0.6 mm to 0.15 - 0.3 mm is 12%. The sand particles classified in the 0.15 - 0.3 mm particle size range are relatively increased. The probability of misclassifying 1.18 - 2.36 mm into 2.36 - 4.75 mm is 4.1%, but the sand particles misclassified from 2.36 - 4.75 mm into 1.18 - 2.36 mm are 2.07%. The sand particles classified in the 1.18 - 2.36 mm particle size range are relatively increased. Therefore, the present invention believes that the two classification methods have different focuses in particle size classification categories, and through experiments, it is proved that applying the combination of the two classification methods will make the grading detection results more stable. The present invention selected three representative samples of coarse sand, medium sand, and fine sand respectively according to the building sand standard for experimental data display, and the specific results are shown in Table 2.

[0102] Table 1 Test Results Table of Sand Grain Classification

[0103]

[0104] Among them, Result represents the probability of correct classification of test samples. Left_P represents the probability that sand grains in this particle size range are classified into the adjacent left particle size range, and Right_P represents the probability that sand grains are classified into the adjacent right particle size range.

[0105] Table 2 Results of Comparative Experiment on Grading Detection of Manufactured Sand

[0106]

[0107] Among them, Δ net represents the calculation error result of sand grading by only applying the particle size division based on the neural network model, and Δ th represents the calculation error result of sand grading by only applying the threshold to divide the particle size. Δ tnet represents the calculation result of sand grading error by combining the two. Among them, "+" means that the detection ratio is less than the actual ratio, and "-" means that the detection ratio is more than the actual ratio. For example, +2.43% means that the detection ratio of this grading range is 2.43% less than the actual ratio. Error represents the cumulative error value of the detection ratio and the true ratio in all grading ranges.

[0108] Equivalent Volume Calculation Optimization Experiment

[0109] Since the sand grains between 0.075 and 4.75 mm are relatively fine, it is not easy to detect the volume of each sand grain. Therefore, in the equivalent ellipse Feret minor axis optimization experiment, we optimized the equivalent ellipse Feret minor axis by comparing the volume ratio of sand grains in each detected particle size range with the true volume ratio and applying the Ampl optimization algorithm. In order to select the best optimization formula, the present invention compared the first-order, second-order, and third-order equivalent ellipse Feret minor axis optimization functions with the unoptimized equivalent ellipse Feret minor axis, and conducted experiments on ten standard 500g manufactured sand samples by combining the method of dividing particle size by threshold and the particle size classified by the neural network model. The results are as Figure 3 shown in Table 3. It can be seen from the chart information that the cumulative error of the manufactured sand grading detection obtained by using the d2 characterization method is more excellent. The obtained average absolute cumulative error and fineness modulus error are the smallest, which are 8.82% and 0.076 respectively.

[0110] Figure 3The sand gradation detection results were obtained by applying four different equivalent ellipse Feret minor axis volume characterization methods to 10 groups of samples. Here, d represents the equivalent ellipse Feret minor axis, and d1, d2, and d3 represent the Feret minor axis values calculated from the optimized linear function, quadratic function, and cubic function expressions. (a) represents the error between the fineness modulus obtained by the four volume characterization methods and the true fineness modulus, and (b) represents the cumulative error diagram of the gradation detected by the four volume characterization methods and the true gradation. From the information in the figure, it can be seen that the cumulative error of the manufactured sand gradation detection obtained by applying the d2 characterization method is more excellent.

[0111] Table 3 shows the average gradation detection errors of the four equivalent volume characterization methods. According to the data, when applying the optimized equivalent ellipse Feret minor axis quadratic function formula, the gradation results of the 10 groups of samples are closer to the sieving method, and the average absolute cumulative error and fineness modulus error are the smallest, which are 8.82% and 0.076 respectively.

[0112] Table 3 Detection results table of 10 groups of manufactured sand gradations

[0113]

[0114] 3) Experiment for validating the effectiveness of parameter optimization

[0115] In order to verify the effectiveness of the optimized parameters of the equivalent ellipse Feret minor axis, the present invention conducted a large number of parameter optimization experiments. In the quadratic function parameter optimization experiment, the present invention randomly selected one sample from ten standard 500g sand samples for parameter optimization, and the other samples were used for testing. Each sample contains approximately 1300 pictures, so the data volume is relatively sufficient. Although only one sample was used for optimization each time, the differences in the experimental results were still small, indicating that the optimization method has good stability and effectiveness.

[0116] The present invention shows the example results of the optimized quadratic function parameters of three groups of threshold divisions and classifications based on the neural network model. As Figure 4 shown, three groups of quadratic function parameter optimization examples are presented, which are the gradation detection results diagrams of the ten 500g standard sand samples of the present invention. Among them, the maximum error in the fineness modulus detection of a single example is 0.07, and the maximum error in the cumulative gradation detection is 0.0276. From Figure 4 the trend diagram, it can be seen that although only one example was used for the quadratic function parameter optimization, due to the sufficient sample data volume in each example, the parameter values changed little, further verifying the good robustness of the quadratic function parameter optimization method.

[0117] In summary, the present invention trains a network model by extracting these sand grain features and combines a threshold division strategy to perform sand gradation detection. By capturing single-graded sand grain images, five sand grain feature information are extracted to train a mature network model. Then, the equivalent volume of the sand grains is calculated by multiplying the equivalent projected area by the short Feret diameter of the equivalent ellipse, and the short Feret diameter of the equivalent ellipse is optimized through an optimization algorithm to enhance the volume characterization calculation. The experimental results show that the average cumulative error between the sand gradation calculated by this method and the manual screening result is 8.82%. This method significantly improves the detection efficiency, reduces the labor input, and has important application prospects in the construction industry.

[0118] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be encompassed within the present invention. Any reference signs in the claims should not be construed as limiting the claims involved.

Claims

1. A sand gradation detection method based on sand two-dimensional image features, characterized in that: At least the following steps are included: S1: Build experimental equipment and take images of sand particles for preliminary screening of sand particles; S2: Using the threshold division method and the classification method based on the neural network model to enhance the stability of classification through sand grain classification; S3: Design a sand volume characterization method to ultimately obtain the optimal sand grading test results.

2. A sand gradation detection method based on sand two-dimensional image features according to claim 1, characterized in that: The experimental equipment in S1 at least includes a feeding bin, a servo, a pressure sensor, a flexible vibration plate and an industrial camera. The industrial camera is located 23.8 cm above the flexible vibration plate. The light source is located below the flexible vibration plate and uses backlight to capture images of sand particles. The main size range of sand particles used for screening by the experimental equipment is 0.075 to 4.75 mm.

3. The sand gradation detection method based on sand two-dimensional image features according to claim 1 is characterized in that: The threshold division method in S2 is to set six division thresholds according to the mesh size of the sieve used in manual screening, so as to classify the sand particles into seven intervals.

4. The sand gradation detection method based on sand two-dimensional image features according to claim 3 is characterized in that: The neural network model classification method in S2 at least comprises the following steps: Two four-classification neural network models were constructed to divide the four intervals of sand particles and fuse them; The four-classification neural network model is composed of two four-classification network structures in the upper and lower layers, and each network structure includes an input layer, a hidden layer and an output layer; The two four-classification neural network models are used to divide the seven particle size intervals into two four-classification intervals; When the two-dimensional sand feature information is input into two four-classification neural network models, the two four-classification neural network models simultaneously output the sand categories in the range of 0.6 to 1.18 mm for comparison and verification, thereby improving the classification accuracy; At the same time, the two four-classification neural network models can also output sand particle categories in the other six particle size ranges.

5. The sand gradation detection method based on sand two-dimensional image features according to claim 4 is characterized in that: The sand grain sizes in the four classification ranges include 0.075-0.15 mm, 0.15-0.3 mm, 0.3-0.6 mm, 0.6-1.18 mm, 0.6-1.18 mm, 1.18-2.36 mm, 2.36-4.75 mm and sand grains above 4.75 mm.

6. The sand gradation detection method based on sand two-dimensional image features according to claim 4 is characterized in that: The input layer is responsible for receiving the normalized training sample data and transmitting it to the hidden layer, and the number of neurons in the input layer is equal to the number of extracted sand grain features; The hidden layer consists of five fully connected layers; The number of neurons in the output layer of each network is 4, which represents the probability of the particle size interval to which the sand grain belongs, and both networks simultaneously predict the fourth particle size interval category.

7. The sand gradation detection method based on sand two-dimensional image features according to claim 1 is characterized in that: The S3 at least includes the following steps: The sand gradation detection needs to measure the mass of sand particles in each particle size range, but the detection method based on two-dimensional images cannot directly obtain the mass of sand particles. In materials science, the density of sand particles in the same batch of materials is usually approximately the same, so their volume ratio and mass ratio are basically the same. Therefore, the sand gradation calculation can be transformed into calculating the volume ratio of sand particles of different particle sizes, that is, converting the mass ratio into the volume ratio, as shown in formula (1); where a i Indicates the percentage of the mass of the i-th particle size interval to the total mass, m i represents the mass of the i-th particle size interval, v i Represents the volume of the i-th particle size interval. Therefore, the two-dimensional sand characteristics are used to accurately characterize the three-dimensional sand volume, effectively improving the accuracy of sand grading detection; The parameters of the equivalent ellipse Feret short diameter were optimized to ensure more accurate representation of the equivalent volume of three-dimensional sand particles; For the threshold division and classification method based on the neural network model, the equivalent ellipse Feret short-path optimization expressions for sand grain volume characterization calculation were obtained through experiments, as shown in formulas (2) and (3); h1=a1d 2 +b1d+c1 (2) h2=a2d 2 +b2d+c2 (3) Among them, h1 is the optimization value of the Feret short diameter of the equivalent ellipse of sand grains for threshold classification, and a1, b1, and c1 are the optimization parameters of the optimization expression; h2 is the optimization value of the Feret short diameter of the equivalent ellipse of sand grains for neural network classification, and a2, b2, and c2 are the optimization parameters of the optimization expression. d is the equivalent Feret short diameter; The total volume of sand particles in each interval is calculated as shown in formula (4) and formula (5); represents the total volume of the i-th interval of the sand grains divided by the threshold, represents the total volume of the i-th interval of the sand grains classified by the neural network, s j It refers to the projection area corresponding to the jth sand grain; For the volume of each interval obtained, formula (3) can be applied to obtain the percentage of the volume of each interval to the total volume; The AMPL optimization algorithm is further used to optimize the parameters of the equivalent ellipse Feret short diameter. Since the sand particles are relatively small, it is very challenging to measure the volume of each sand particle individually. Therefore, by comparing the actual value of the standard 500g machine-made sand grading with the volume cumulative summation result calculated by applying the equivalent ellipse Feret short diameter optimization expression, the volume proportion of each sand particle interval is analyzed, so as to improve the parameters of the optimization expression, as shown in formula (6) and formula (7). Where y i It represents the true value of the gradation of the i-th interval of standard 500g machine-made sand; Through these two formulas, we can finally get the equivalent elliptical Feret short-path optimization parameter expressions for threshold division and neural network classification respectively. After experimental verification, for the final test results of sand grading, when the threshold division method is applied, the proportion of sand particles in the range of 0.15-0.3 mm is relatively high, while the proportion of sand particles in the range of 1.18-2.36 mm is relatively low. On the contrary, the test results of the neural network model show the opposite trend; To this end, the test results of the two methods are combined and averaged to finally obtain the optimal sand gradation test result, as shown in formula (8); in It refers to the volume proportion of the i-th interval of sand particles divided by the threshold, refers to the volume proportion of the i-th interval of sand particles classified by the neural network, It refers to the volume proportion of the ith interval after taking the average of the two.

Images