Aggregate particle recognition and gradation automatic analysis method based on deep learning

Abstract

The application discloses a method for automatic analysis of aggregate particle identification and grading based on deep learning, which comprises the following steps: preparing a standard data set of aggregates, training a deep learning model Mask R-CNN in a training set to obtain an optimal deep learning model of the deep learning model of the material aggregate; determining a test platform area meeting accuracy requirements according to the recognition accuracy and area relationship of the smallest particle size group in the aggregate; realizing automatic identification and segmentation of aggregate particles based on segmentation and splicing technology of large images; calculating the equivalent particle diameter of the aggregate particles, determining the coefficient of each particle in the aggregate through the particle shape, and calculating the equivalent particle diameter of each particle in the aggregate; calculating the volume of the aggregate particles, and calculating the mass proportion of all particles in the particle size range according to the equivalent particle diameter division. The method can output the grading image and grading data of the aggregate in real time, greatly improves the accuracy and efficiency, and can effectively replace the traditional screening method.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aggregate gradation analysis, and particularly relates to an aggregate particle identification and gradation automatic analysis method based on deep learning. BACKGROUND

[0002] In the field of civil engineering, concrete and asphalt are commonly used raw materials, and aggregate, as an important component of concrete and asphalt, directly determines its various road performances, and is also an important index of engineering construction. Therefore, the precision and efficiency of aggregate gradation analysis have important theoretical significance and application value. Gradation analysis is a method for reflecting the percentage of a series of particles in different particle size intervals in the total amount of particles. Common methods include laser method, screening method, microscope method, ultrasonic particle size analysis, and particle image method. At present, the gradation analysis of aggregate mainly adopts the traditional screening method, which is mechanical screening or manual screening. However, the screening method is time-consuming and laborious, and the precision is unstable. Moreover, it cannot obtain gradation data in real time. Therefore, establishing an efficient and accurate aggregate gradation automatic identification method will provide necessary technical support for engineering construction. SUMMARY

[0003] The present application provides an aggregate particle identification and gradation automatic analysis method based on a deep learning model Mask R-CNN. Compared with the traditional mechanical or manual screening method, in actual application, only the aggregate image needs to be shot, and the aggregate image is input into the optimal model obtained by training, so that the qualitative and quantitative analysis results of the gradation can be obtained, and the gradation image and gradation data of the aggregate can be output in real time. The precision and efficiency are greatly improved, and the traditional screening method can be effectively replaced. At the same time, it is verified that the model has strong robustness for aggregate with different clay contents. Through the transfer learning technology, the generalization of the model on other materials of aggregate is verified.

[0004] The technical scheme adopted by the present application is as follows:

[0005] The aggregate particle identification and gradation automatic analysis method based on deep learning comprises the following steps:

[0006] Step 1: Prepare a standard data set of aggregate of a corresponding material, divide the standard data set into a training set, a verification set and a test set, train a deep learning model Mask R-CNN in the training set, and obtain the optimal deep learning model of the deep learning model of the aggregate of the material;

[0007] Step 2: Determine the test platform area meeting the accuracy requirement according to the recognition accuracy and area relationship of the smallest particle group in the aggregate;

[0008] Step 3: Realize the automatic identification and segmentation of aggregate particles based on the segmentation and splicing technology of large images;

[0009] Step four: Calculate the equivalent particle size of the aggregate particles, determine the coefficient by the particle shape, and calculate the equivalent particle size of each particle in the aggregate;

[0010] Step five: Calculate the volume of the aggregate particles, and calculate the mass proportion of all particles in the particle size range according to the equivalent particle size of the particles, that is, the gradation of the aggregate;

[0011] Step six: Through the aggregate particle test of different clay content, evaluate the influence of clay content on the precision of the model and the gradation analysis.

[0012] Step seven: Through the production of standard image data set in other material aggregates, and using the transfer learning method, increase the model training based on the model, and get the optimal model of other materials.

[0013] The step one includes the following steps:

[0014] S1.1: First, collect the mixed aggregate images of different particle sizes of the material, get the original data set and complete the preprocessing, then label the original data set in the labeling software, and expand the data set through data augmentation technology, get the standard data set of the material;

[0015] S1.2: Divide the standard data set into training set, validation set and test set according to the proportion of 8:1:1, train the deep learning model Mask R-CNN in the training set, and verify in the validation set, test in the test set, finally get the optimal hyperparameters and optimal deep learning model of the deep learning model Mask R-CNN of the material aggregate, and download and save the optimal hyperparameters and optimal deep learning model.

[0016] The step two includes the following steps:

[0017] S2.1: First, randomly select the smallest particle size aggregate and place it in the square area test platform with side length a0;

[0018] S2.2: The same batch of aggregate is unchanged, and the area is sequentially reduced by 90%, 80%, 70%, 60%, 50%, 40%, and 30%. Since the actual application does not require less than 2% accuracy, through the test, the minimum area S0 that meets the accuracy requirement of 2% is found to be 70% of the original area.

[0019] The step three includes the following steps:

[0020] S3.1: According to the size of the test platform area, design and realize the 1kg aggregate feeding platform device with vibration function, which can also take pictures and obtain images through vertical sound control method. Each time, three groups of parallel tests are carried out, and 10 high-definition images are obtained by vibrating 10 times for each group;

[0021] S3.2: Adopt image segmentation technology to divide one super large image into 16 independent local images, identify and segment respectively, then adopt image splicing technology to merge into the original image size, and finally output the result image.

[0022] In the fourth step, the equivalent area diameter is used as the particle diameter for the elliptical shape, and the equivalent elliptical Feret short diameter is used as the particle diameter for the needle-like shape. The specific steps are as follows:

[0023] First, the optimal deep learning model is used to identify and segment the projected area of the aggregate particles. Then, the minimum circumscribed rectangle of the projected area is calculated. Finally, the ratio of the long and short axes of the minimum circumscribed rectangle is obtained. Through experiments, the ratio of the long and short axes of the class ellipse and the class needle is determined as the particle shape judgment coefficient c. Finally, the equivalent particle diameter D of each particle in the aggregate is calculated, as shown in formula (1).

[0024] In the fifth step, through comparative experiments, it is found that the equivalent volume obtained by multiplying the projected area A of the particle by the Feret short diameter R of the projected area is closest to the actual volume of the particle. Therefore, this method is used to calculate the volume of the aggregate particles, as shown in formula (2).

[0025] The aggregate gradation is shown in formula (3).

[0026] The effect of vibration on gradation is studied through aggregate vibration test, and the vibration frequency that does not affect the accuracy of gradation is obtained:

[0027] A 1kg aggregate feeding platform is designed and implemented. The mixed aggregate diabase 1kg is fed on the feeding platform and vibrated 10 times. One image is obtained by sound control shooting for each vibration. A total of 10 images are obtained. The aggregate particle gradation GIR in each image is analyzed in turn, and the average value is accumulated. The accumulated gradation average GIR after several vibrations is completely stable, close to the sieving gradation, and has high accuracy.

[0028] In the sixth step, in actual application, the clay content of aggregate is generally not higher than 4%. Different aggregate clay content tests 1%, 2%, 3%, 4%, 5%, 6% are designed. After the images obtained by the test platform are preprocessed, the optimal deep learning model is used for identification and segmentation to automatically obtain the gradation of the aggregate.

[0029] In the seventh step, through transfer learning, standard data sets are made in other common materials: diabase, limestone, basalt, and pebble aggregate images, and the deep learning model Mask R-CNN is trained. The gradation analysis can meet the accuracy requirements, verifying the transferability and generality of the optimal deep learning model Mask R-CNN.

[0030] The technical effects of the aggregate particle recognition and gradation automatic analysis method based on deep learning are as follows:

[0031] 1) The aggregate particle recognition and gradation automatic analysis method based on deep learning is proposed, which first recognizes the aggregate particles and then analyzes the aggregate gradation. This method can be applied to the recognition of aggregate particles of different materials and the automatic analysis of aggregate gradation. Compared with the traditional screening method, the automatic gradation recognition model established by the deep learning framework of the application can obtain efficient and accurate results in real time.

[0032] 2) The aggregate particle recognition and gradation automatic analysis method based on deep learning provided by the application can obtain qualitative and quantitative analysis results of the gradation by only shooting the aggregate image and inputting the aggregate image into the optimal model obtained by training, and can output the gradation image and gradation data of the aggregate in real time, which greatly improves the accuracy and efficiency and can effectively replace the traditional screening method. At the same time, it is verified that the model has strong robustness to aggregate with different clay contents.

[0033] 3) The application verifies that the model meets the accuracy requirements for aggregate with different clay contents and verifies the robustness of the model by designing different clay content tests. The optimal model of other materials is obtained by increasing model training based on the diabase model through standard image dataset of other common materials such as limestone, basalt and pebble aggregate and using transfer learning. The model automatically obtains the gradation of the aggregate, and verifies the transferability and generalization of the model.

[0034] 4) The evaluation indexes used in the application include 6 indexes: absolute error AE (Absolute Error), mean absolute percentage error MAPE (Mean Absolute Percentage Error), mean intersection over union mIoU (mean Intersection over Union), mean average precision mAP (mean Average Precision), Euclidean distance ED (Euclidean Distance), and one-way distance OWD (One Way Distance). Among them: the first four evaluation indexes can objectively and accurately evaluate the performance of the deep learning model, and the last two evaluation indexes evaluate the accuracy of the aggregate gradation. BRIEF DESCRIPTION OF DRAWINGS

[0035] The application will be further described below in combination with the drawings and examples:

[0036] Figure 1 It is a flow chart for the recognition and segmentation of mixed aggregate particles based on the deep learning Mask R-CNN framework.

[0037] Figure 2(a) is a qualitative analysis diagram of the original mixed aggregate 1 kg;

[0038] Figure 2(b) is a qualitative analysis diagram of the grading identification result of Figure 2(a).

[0039] Different colors represent different particle size ranges.

[0040] Figure 3 It is a cumulative mass curve diagram of the mixed aggregate 1 kg.

[0041] Figure 4 It is a particle diagram of the aggregate with five different size ranges.

[0042] Figure 5 It is a schematic diagram of the vibration shooting system structure of the mixed aggregate 1 kg.

[0043] Figure 6(a) is a true value diagram of polygon labeling of mixed aggregate particles with five different particle sizes in Labelme software;

[0044] Figure 6(b) is a mask diagram corresponding to the true value of Figure 6(a).

[0045] Figure 7 It is a schematic diagram of the process of using data set enhancement technology for images.

[0046] Figure 8 It is a training loss and validation loss curve diagram in the training process of the deep learning model MaskR-CNN.

[0047] Figure 9(a) is a diabase aggregate 50g original image (minimum particle size 2.36-4.75mm);

[0048] Figure 9(b) is a diabase aggregate 50g model identification and segmentation result image.

[0049] Figure 10(a) is a large image processing flowchart;

[0050] Figure 10(b) is a segmentation identification and instance merging diagram of a large image.

[0051] Figure 11 It is a cumulative distribution curve diagram of the first batch of aggregate particles.

[0052] Figure 12 It is an image of the identification result of diabase aggregate with different clay contents.

[0053] Figure 13(a) is a limestone original image Figure 1 Model untrained result image Model trained result image;

[0054] Figure 13(b) is a limestone original image 2 model untrained result image Model trained result image.

[0055] Figure 14(a) is a limestone aggregate particle 100g original drawing;

[0056] Figure 14(b) is a model identification result drawing of Figure 14(a).

[0057] Figure 15 For limestone aggregate particle grading curve drawing. DETAILED DESCRIPTION

[0058] The aggregate particle identification and grading automatic analysis method based on deep learning includes the following steps:

[0059] First step: according to the actual application requirement, self-made standard data set of corresponding material aggregate;

[0060] First, collect the mixed aggregate images of different particle sizes of the material to obtain the original data set and complete the preprocessing, then label the data set in the labeling software, and expand the data set through data enhancement technology to obtain the standard data set of the material.

[0061] Then, the standard data set is divided into training set, validation set and test set according to the proportion 8:1:1. Train the deep learning model Mask R-CNN in the training set, verify it in the validation set, and test it in the test set. Finally, the optimal hyperparameters and optimal deep learning model of the material aggregate are obtained, and the model is downloaded and saved.

[0062] Second step: according to the accuracy requirement of actual application, determine the area of the mixed aggregate 1kg pouring test platform whose model accuracy is not lower than the accuracy requirement;

[0063] Taking the commonly used material diabase as an example, since the image resolution accuracy is high (the resolution is 2048 pixels*2048 pixels), in order to unify the model unit conversion and ensure the accuracy, the aggregate pouring area is designed as 20.48cm*20.48cm, that is, 1 pixel corresponds to 0.1mm. Since the smaller the particle size is, the more difficult the model identification is, in order to effectively evaluate the model accuracy, first, randomly select 50g of the smallest particle size 2.36-4.75mm aggregate, and pour it in the test platform with a square area (a0*a0=20.48cm*20.48cm) with a side length a0 of 20.48cm, and try to disperse and not overlap; Next, the same batch of aggregate is unchanged, and the area is sequentially reduced by 90%, 80%, 70%, 60%, 50%, 40%, and 30%. Since not less than 2% accuracy can meet the demand in actual application, through the test, the minimum area S0 that meets the accuracy requirement (2%) is found to be 70% of the original area.

[0064] According to the minimum area required by the aggregate 50g, the required area size S of 1kg and the corresponding side length a are calculated, and the calculation process is as follows:

[0065] The minimum area S0 required by the aggregate 50g and the side length a0 are as follows:

[0066] S0=a0*a0*0.7=20.48cm*20.48cm*0.7=293.60128cm 2

[0067] The minimum area required by 1kg of crushed stone aggregate is derived as follows:

[0068]

[0069]

[0070] From the above calculation, it is known that the standard size of the feeding platform required by 1kg of aggregate under the requirement of not less than 2% accuracy is 76.63cm*76.63cm.

[0071] Step 3: According to the area size of the test platform, a 1kg aggregate feeding platform device with vibration function is designed and implemented, which also has vertical sound control method for image acquisition. Three groups of parallel tests are performed each time, and 10 high-definition images are obtained by vibrating 10 times each group.

[0072] Step 4: A solution to the problem of limited software and hardware resource computing power of deep learning model is proposed by using image segmentation and splicing technology. The number of 1kg mixed aggregate particles is about 3000, and the image obtained by shooting has the characteristics of high resolution, high precision and large capacity. Due to the limitation of computing power, the deep learning model cannot process 1 super large image at a time, so the image segmentation technology is used to divide 1 super large image into 16 independent local images, which are identified and segmented respectively, and then spliced into the original image size, and finally the result image is output. If the original image is directly segmented, identified and merged, the particles on the segmentation line will be divided into two, which reduces the particle recognition accuracy of the model. Therefore, the present application designs a detachable aggregate tray with a 16-grid metal frame at the top and a horizontal steel plate at the bottom. During the test, 1kg of aggregate is evenly placed in the 16 grids of the tray, the metal frame is removed, and the aggregate particles are photographed, then the large image obtained by photographing is segmented, the model is identified and merged, effectively solving the problem of incorrect segmentation of particles on the segmentation line.

[0073] Step 5: A new method for calculating the equivalent particle size of aggregate particles is proposed:

[0074] Due to the different shapes of aggregate particles, they can be roughly divided into ellipse-like and needle-like. The present application proposes a new mixed mode to calculate the equivalent particle diameter of aggregate particles. The equivalent area diameter can be used as the particle diameter for ellipse-like particles, and the equivalent ellipse Ferret short diameter can be used as the particle diameter for needle-like particles. The specific steps are as follows: first, the projection area of the aggregate particles is identified and segmented through a deep learning model, then the minimum circumscribed rectangle of the projection area is calculated, and finally the ratio of the long and short axes of the circumscribed rectangle is obtained. The ratio of the long and short axes of the ellipse-like and needle-like is determined by experiment as the particle shape judgment coefficient c, and the equivalent particle diameter D of each particle in the aggregate is finally calculated, as shown in formula (1).

[0075]

[0076] Where D is the equivalent particle diameter, a is the minimum circumscribed rectangle short axis, b is the minimum circumscribed rectangle long axis, and c is the judgment coefficient determined by experiment.

[0077] Step 6: Calculate the volume of the aggregate particles, and calculate the mass proportion of all particles in the particle size range (divided by the equivalent particle diameter of the particles), that is, the aggregate gradation. Through comparison test, it is found that the equivalent volume obtained by multiplying the projection area A of the particles by the Feret short diameter R of the projection area is closest to the actual volume of the particles, so this method is used to calculate the volume of the aggregate particles, as shown in formula (2).

[0078] V≈A · R(2)

[0079] Where: A represents the projection area of the aggregate particles, and R represents the Feret short diameter corresponding to the projection area. The total volume of the aggregate particles in each particle size range is calculated, and the mass proportion GIR (Gradation in Range) is obtained, that is, the aggregate gradation, as shown in formula (3):

[0080]

[0081] Where Q represents the mass of the aggregate particles, A represents the projection area of the aggregate particles, D represents the Feret short diameter corresponding to the projection area, m and n represent the total number of aggregate particles and the number of particles in a certain particle size range, respectively, ρ represents the particle density, v represents the particle volume, and i and j represent the particle serial number in a certain particle size range and the total particle serial number, respectively.

[0082] Step 7: Study the effect of vibration on gradation through aggregate vibration test to obtain the vibration frequency that does not affect the accuracy of gradation. The gradation curve of the mixed aggregate is obtained by the traditional screening method, and the deep learning model is used to automatically obtain the gradation curve for comparison and analysis, and the evaluation index Euclidean distance ED is used to represent the closeness of the sample gradation curve and the screening curve.

[0083] The specific process is as follows: a 1kg aggregate feeding platform is designed and implemented, 1kg of mixed aggregate diabase is fed on the feeding platform, and vibration is completed for 10 times, 1 image is obtained by sound control shooting for each vibration, 10 images are obtained in total, the aggregate particle gradation GIR in each image is analyzed in turn, and the average value is accumulated, and the accumulated gradation average GIR obtained after several vibrations is completely stable, close to the screening gradation, and high in accuracy.

[0084] Step 8: Through the aggregate particle test of different clay contents, the influence of clay content on model accuracy and gradation analysis is evaluated. In actual application, the clay content of aggregate is generally not higher than 4%, and different aggregate clay contents 1%, 2%, 3%, 4%, 5%, and 6% are designed for testing. After the images obtained by the test platform are preprocessed, they are input into the deep learning model for recognition and segmentation, and the gradation of the aggregate is automatically obtained.

[0085] Table 1 is the qualitative analysis result of the aggregate particle gradation, and different colors represent different particle size ranges. Table 2 is the quantitative analysis result of the aggregate particle gradation.

[0086] Table 1 corresponding color of different particle size range

[0087]

[0088] Table 2 quantitative analysis data of 1kg gradation of mixed aggregate

[0089]

[0090]

[0091] Step 9: Through the transfer learning method, standard data sets are made and models are trained in other common materials: diabase, limestone, basalt, and aggregate images of pebbles, and the gradation analysis meeting the accuracy requirements can be obtained, verifying the transferability and generalization of the original model.

[0092] Step 10: The evaluation indexes used in the present application are absolute error AE, mean absolute percentage error MAPE, mean intersection over union mIoU, mean pixel accuracy mPA, and Euclidean distance ED.

[0093] The specific calculation method is as follows:

[0094] 1) Absolute error AE (Absolute Error)

[0095] The difference between the true value and the number of recognized numbers is that the true value of the aggregate in the image is used as a standard, and the number of aggregate particles recognized by the model is compared, and the difference is used as a key evaluation index. The smaller the difference, the better the performance.

[0096]

[0097] wherein, y represents the actual value.

[0098] 2) Mean Absolute Percentage Error (MAPE)

[0099] The ratio of the absolute value of all sample errors to the actual value, the closer the index is to 0, the more accurate the model is, and the better the model performance

[0100]

[0101] wherein, y represents the actual value.

[0102] 3) Intersection over Union (IoU)

[0103] Pixel evaluation index, the ratio of the intersection pixel value to the union pixel value of the aggregate particle image identified by deep learning and the aggregate true value image, the closer to 1, the higher the accuracy. That is, the intersection of the predicted region and the actual region divided by the union of the predicted region and the actual region, simply referred to as the intersection over union. Calculate the proportion of the intersection and union between two sets, in image segmentation, it is the proportion of the ground truth and the predicted value of the two sets.

[0104]

[0105] Mean Intersection over Union (mIoU)

[0106] First calculate the intersection over union of each class, then average to get the average intersection over union of all classes.

[0107]

[0108] 4) Average Precision (AP)

[0109] The average precision AP is used to evaluate the overall performance of the model. AP is the average precision of predicting a single target class, which is calculated from precision and recall. The P-R (Precision-recall) curve is drawn with precision as the vertical coordinate and recall as the horizontal coordinate, and the AP is the integral of the P-R curve. Precision refers to the ratio of positive samples in the test set that are labeled as particles; Recall is the proportion of all positive samples in the test set that are correctly identified as particles.

[0110] The calculation of Precision and Recall is as follows:

[0111]

[0112]

[0113] wherein TP represents the number of particles that are actually particles and are predicted as particles by the model, i.e. positive samples are detected as positive samples; FP represents the number of particles that are actually background but are predicted as particles by the model, i.e. negative samples are detected as positive samples; FN represents the number of particles that are actually particles but are not identified as particles by the model, i.e. positive samples are not detected as positive samples.

[0114] AP 50 , AP 75 , AP 80 respectively represent the values of AP when IoU = 0.5, 0.75, 0.8.

[0115] The average precision mean mAP is the arithmetic mean of the average precision AP of each class, i.e. the average precision AP of each class is first calculated, and then the arithmetic mean of the AP of these classes is calculated, and the classes are divided into two categories: particles and background.

[0116]

[0117] wherein K is the number of classes.

[0118] 5) Euclidean distance ED (Euclidean Distance)

[0119] ED is a distance definition index, which refers to the true distance between two points in n-dimensional space, or the natural length of a vector (i.e. the distance of the point to the origin). By calculating the ED distance between points on different curves, the closeness of the curves is represented, and the smaller the ED value, the closer the curves, and vice versa, the farther the curves deviate.

[0120] Given two points A = (a1, a2,..., a n ), B = (b1, b2,..., b n ), the distance between A and B is:

[0121]

[0122] 6) One-way distance OWD (One Way Distance)

[0123] The OWD refers to the area surrounded by the two trajectories. When the area is large, the distance between the trajectories is far, and the similarity is low. Conversely, if the surrounding area is 0, the two trajectories coincide, and the similarity is the highest. The similarity can be calculated according to the formula.

[0124]

[0125]

[0126] wherein T1 and T2 are two trajectories, and p is a discrete point on T1.

[0127] Verification example:

[0128] The present application provides a new method for automatic identification of aggregate particle gradation based on deep learning, and a flow chart is shown in Figure 1 The method comprises the following steps:

[0129] S1: preparing aggregates.

[0130] The raw material used in this embodiment is diabase. According to the actual industry application requirements, five kinds of diabase aggregate particles with different particle sizes are prepared, and the particle sizes are 2.36mm-4.75mm, 4.75mm-7mm, 7mm-9.5mm, 9.5mm-13.2mm, and 13.2mm-16mm, as shown in Figure 4 .

[0131] S2: obtaining an original data set.

[0132] Five kinds of mixed aggregates with different particle sizes are prepared, a high-definition digital camera is used to fix the position and focus vertically to shoot, 50 original high-definition images are obtained, and a schematic diagram of the shooting system is shown in Figure 5 . The image resolution is 2048*2048, and after shooting, image preprocessing such as trapezoidal correction is carried out in Photoshop software to form an original data set.

[0133] S3: data set labeling.

[0134] The Labelme image labeling software is used to manually label the 50 original data sets to form the true value data of the aggregate particle images. Since the deep learning model Mask R-CNN can realize image pixel-level recognition and segmentation, the labeling accuracy will directly determine the accuracy of the model in identifying aggregate particles. In order to improve the labeling accuracy as much as possible, first, the image is enlarged by 3 times in the labeling software Labelme, and then the edge of the aggregate particle is marked with a polygon point. After labeling, the original image size is restored to save the true value data of the aggregate particles, and the corresponding mask image is automatically formed, as shown in Figs. 6(a) and 6(b).

[0135] S4: Data set augmentation forms a standard data set.

[0136] Expanding the original data set by using data set augmentation technology can effectively increase the diversity of data, thereby improving the precision and generalization performance of the model. In the present application, the augmentation techniques used include color variation, adding different proportions of Gaussian noise, various mirror transformations, scaling images by different proportions, rotating images by different angles, shear transformation, etc., as shown in Figure 7 The original 50 labeled image data sets are augmented 20 times by data set augmentation, totaling 1000, as the standard data set for model training.

[0137] S5: Constructing a software and hardware experimental platform for deep learning model.

[0138] The running environment of the deep learning model Mask R-CNN is deployed on an Ubuntu server, and the software and hardware configurations are as follows. Operating system: Ubuntu 18.04, independent graphics card: NVIDIA T4 GPU, graphics card driver version: NVIDIA-Linux-x86_64-450.80.02, Cuda version: cuda_9.0.176_384.81_linux, Cudnn version: cudnn-9.0-linux-x64-v7.0.5.15, Tensorflow version: TensorFlow 1.14.0+Keras 2.2.5.

[0139] S6: Training a deep learning model.

[0140] The standard data set is divided into training set, validation set and test set according to the proportion of 8:1:1. The model is trained on the training set, and the optimal hyperparameters of the model are obtained by continuously adjusting the model parameters based on the performance results on the validation set. Finally, the model running results are given on the test set. The training process is as follows: the training of Mask R-CNN model is carried out under the TensorFlow deep learning framework, and GPU acceleration is adopted. First, the Mask R-CNN model selects the network ResNet-101 as the feature extraction backbone network, and the COCO data set pre-training network is used to initialize the network parameters of the model. Then, the model is trained, validated, and tested on the standard data set of diabase aggregate. The whole training of the model is divided into two stages:

[0141] S6_1: First, freeze the backbone network, and train the randomly initialized layers (all layers except Backbone) that do not use COCO pre-training weights;

[0142] S6_2: Second, train the entire Mask R-CNN model.

[0143] The whole training process of the deep learning model Mask R-CNN is carried out under the deep learning framework of Tensorflow and is accelerated by GPU. ResNet-101 network is selected as the backbone network for feature extraction. Then, the network parameters are initialized by the pre-trained weights on the COCO dataset. Then the model training is completed on the self-made diabase aggregate particle dataset. The various parameter settings in the model training are as follows: the category parameter is set to 2 (including diabase particles and background); the anchor box size is set to 8, 16, 32, 64, 128 in turn; the anchor box corresponds to three ratios (0.5, 1.0, 2.0) for each size, and the anchor box number of each image is set to 256; the alpha weight coefficient is set to 1; the weight decay coefficient is set to 0.0005. In addition, the activation function of the model adopts the rectified linear unit (ReLU), and the optimization function adopts the stochastic gradient descent (SGD) based; the momentum factor is set to 0.9; the initial learning rate of the network weight parameter is set to 0.005; a total of 400 iterations (Epoch) are set to complete the training, and when the Epoch reaches 20, the learning rate is reduced to 0.0001. The training loss curve and the validation loss curve of the model are output in real time during the training process. During this training process, it is found that the loss curve of the original model tends to be flat after the 50th Epoch, and tends to converge at the 300th iteration. During the whole training process, the training loss curve on the training set and the validation loss curve on the validation set are drawn in real time by the program, as shown in FIG. 2. Figure 8

[0144] S7: Determine the area of the mixed aggregate diabase 1 kg drop test platform.

[0145] S7_1: First determine the identification result of the minimum particle size group aggregate 50 g in the diabase aggregate.

[0146] ​According to the recognition accuracy requirement of diabase aggregate particles in practical application, the area of the 1 kg diabase aggregate mixing aggregate test platform is calculated. Since the smaller the particle size of the aggregate with equal mass, the more the quantity, and the higher the model accuracy requirement, the model is first used to identify and segment the minimum particle size (2.36-4.75 mm) aggregate with equal mass. The specific process is as follows: weigh 50 g of diabase aggregate with a particle size of 2.36-4.75 mm, and the resolution of the image is 2048 pixels*2048 pixels. In order to unify the model unit conversion and ensure the accuracy, the design aggregate placement area is 20.48 cm*20.48 cm, that is, 1 pixel corresponds to 0.1 mm. Place 50 g of the minimum particle size group (2.36-4.75 mm) of the aggregate on the test platform with an area of 20.48 cm*20.48 cm, and try not to stack, conduct 3 groups of tests, and take 3 groups of images. One of the original images and the deep learning model identification and segmentation result images are shown in FIG. 9(a) and FIG. 9(b), and the identification and segmentation evaluation indexes of the 3 groups of images are shown in Table 3.

[0147] Identification and segmentation evaluation indexes of 3 groups of images in Table 3

[0148]

[0149] The experimental results show that the effect of placing 50 g of the minimum particle size aggregate on the platform with an area of 20.48 cm*20.48 cm for particle identification and segmentation is ideal, the recognition accuracy is high, the number error percentage MAPE is less than 0.5%, and the segmentation index mAP 75 is also higher than 90%, which is much higher than the accuracy requirement.

[0150] S7_2: Determine the minimum placement platform area of 50 g of the minimum particle size group of diabase aggregate.

[0151] The same batch of 50 g of the minimum particle size group of diabase aggregate is placed in a smaller and smaller area, then identified and segmented, and the model accuracy is calculated to find the relationship between the area and the accuracy, and finally determine the area size with a model accuracy not less than 2%. The test process is as follows: (1) randomly select 50 g of diabase aggregate with the minimum particle size (2.36-4.75 mm), and place it on the test platform with a square area (a0*a0=20.48 cm*20.48 cm) with a side length a0 of 20.48 cm, and try to disperse and not stack; then, the same batch of aggregate is not changed, and the area is placed in proportion, and the proportion used is 90%, 80%, 70%, 60%, 50%, 40%, and 30% in turn. The proportion setting and the corresponding area and side length are shown in Table 4.

[0152] Corresponding relationship between area after scaling by proportion and side length

[0153]

[0154]

[0155] (2) Then, randomly select 50g diabase aggregates of 5 different particle sizes, and apply and scale them in the same way. Compare and analyze the recognition effect of a single minimum particle size group and the recognition effect and accuracy of the mixed 5 particle size groups. Since an accuracy of no less than 2% is sufficient in practical applications, the minimum area S that meets the accuracy requirement (2%) is found through experiments.

[0156] The identification results of the experiment are shown in Table 5(1) and Table 5(2), and the error analysis results are shown in Table 5(3).

[0157] Table 5(1) Original aggregate drawing and recognition result drawing after area scaling, and number of recognized aggregates (minimum particle size)

[0158]

[0159]

[0160] Table 5(2) Original aggregate drawing and identification result drawing after area scaling, and number of identified aggregates (mixed particle size)

[0161]

[0162]

[0163]

[0164] Table 5(3) Number of particles with a minimum particle size of 50g and the error of 50g particles of five different area ratios identified.

[0165]

[0166] The experimental results show that the identification and segmentation of the smallest particle size aggregate is more difficult than that of the mixed particle size, and the identification accuracy is lower than that of the mixed particle size under the same conditions. 70% of the area meets the accuracy requirement of less than 2% error. Therefore, the area corresponding to the error requirement of the smallest particle size is used as the minimum design area of ​​the feeding platform for 50g aggregate.

[0167] S7_3: Determine the minimum area of ​​the feeding platform for 1 kg of diabase mixed aggregate.

[0168] According to the minimum sampling mass for sieve analysis specified in the "Specifications for Testing Aggregates in Highway Engineering" (JTG E42-2005 T0302), 1 kg of diabase mixed aggregate needs to be weighed for gradation analysis to obtain the gradation. Based on the minimum feeding area S0 required for 50g of diabase aggregate, the area S required for 1kg and the corresponding side length a are calculated as follows:

[0169] Minimum area required for 50g of diabase aggregate:

[0170] S0 = a0 * a0 * 0.7 = 20.48cm * 20.48cm * 0.7 = 293.60128cm 2

[0171] The minimum area required for 1kg of diabase aggregate is derived:

[0172]

[0173]

[0174] From the above calculation, the standard size of the feeding platform required for 1kg of aggregate is 76.63cm * 76.63cm under the requirement of pattern recognition accuracy not less than 2%.

[0175] S8: Propose segmentation and stitching technology for large images, solve the problem of limited computing power of deep learning model, and apply it to 1kg diabase mixed aggregate.

[0176] First, divide the diabase aggregate large image into 16 local images, use the deep learning model to identify and segment the 16 local images respectively, and then use the image stitching technology to merge the 16 local image results into one original size image for output. The logic flow chart is shown in Figure 10(a).

[0177] The specific description is as follows: Since the number of 1kg aggregate particles is about 3000, the hardware and software resource computing power of the deep learning model is limited, the maximum number of target objects identified by the model at a time is 300, and it cannot directly complete the identification and segmentation task of all particles at a time. It is necessary to crop and segment the original image into multiple local images for identification and segmentation task and then stitch them into the original image size. Since the cropping operation will mistakenly split one particle on the segmentation line into two particles, resulting in a decrease in the overall accuracy of the model. Therefore, the present application designs a detachable aggregate tray with a 16-grid metal frame at the top and a horizontal steel plate at the bottom. During the test, first put 1kg of aggregate evenly into the 16 grids of the tray, take out the metal frame, take a picture of the aggregate particles, then segment the large image obtained by photographing, model identification and merging, effectively solving the problem of mistaken segmentation of particles on the segmentation line.

[0178] To ensure the operation accuracy and efficiency of the model, the feeding platform is equally divided into 16 areas, and the areas are separated by a 0.5mm thick organic plastic film. Before feeding the aggregate, divide it into 16 equal parts, and put them into different areas in turn and perform reasonable vibration to make the aggregate as little overlapping as possible and exist in a stable posture. The test platform design diagram is as follows: Figure 5The model recognition and segmentation effect is shown in Fig. 10(b).

[0179] S9: Calculate the equivalent particle diameter of diabase aggregate particles.

[0180] The present application proposes a new mixed mode to calculate the equivalent particle diameter of aggregate particles, which is described as follows: First, the determination coefficient c of the equivalent particle diameter is established through experiments (the ratio of the major and minor axes of the circumscribed ellipse of the two-dimensional projection image of the aggregate particle is taken as the determination coefficient). If the ratio of the major and minor axes of the minimum circumscribed rectangle of the projection area of the aggregate particle is greater than or equal to c, the equivalent Feret short diameter is taken as the equivalent particle diameter; if the ratio of the major and minor axes is less than c, the equivalent area diameter is taken as the equivalent particle diameter, as shown in formula (14).

[0181]

[0182] where D is the equivalent particle diameter, a is the minimum circumscribed rectangle short axis, b is the minimum circumscribed rectangle long axis, and c is the determination coefficient determined through experiments.

[0183] In this example, the equivalent particle diameter of diabase aggregate particles is calculated. In order to determine the ratio coefficient of the major and minor axes of the diabase aggregate particles, three groups of parallel experiments are designed. Each group of experiments respectively takes approximately equal amount of sampling from 5 different particle size aggregates, and the sampling forms are different, with 120 particles in each group. In order to reduce the error caused by sampling difference, three groups of parallel experiments are performed, i.e. three batches of aggregate sampling in the same way, a total of 360, to form the true value data; in order to reduce the error caused by different attitudes of the same particle, the same batch of aggregate is vibrated for 10 times, a total of 3600 particles, and 30 images are taken, and the average value of the 30 images is taken as the sampling result.

[0184] Through the deep learning model, the diabase aggregate particles in the image are recognized and segmented to obtain the minimum circumscribed rectangle long and short axis, equivalent area diameter, and equivalent Ferret short diameter of each particle. Then different minimum circumscribed rectangle long and short axis ratio coefficients (1.3, 1.4, 1.5, 1.6, 1.7, 1.8 respectively) are set to calculate the equivalent particle diameter of the mixed mode, and compared with the true value. The coefficient closest to the true value is taken as the determination coefficient of the equivalent particle diameter, and the test data is shown in Table 6.

[0185] Table 6 Test data of determination coefficient of equivalent particle diameter

[0186]

[0187] From table 6, the number of sieving of 5 different particle sizes and the number of corresponding different coefficients, when the equivalent particle diameter takes the equivalent area diameter, the value fluctuates greatly; when the equivalent particle diameter takes the equivalent ellipse Feret short diameter, the value is all smaller; when the equivalent particle diameter takes the ratio of the minimum circumscribed rectangle long and short diameter as 1.6, the difference value of each grade number and the true value is at most 1, which is closest to the sieving number.

[0188] In summary, in the examples of the present application, for diabase aggregate particles, the equivalent particle diameter determination coefficient 1.6 is proposed as a division standard to calculate the equivalent particle diameter of the particles, which has higher precision than other methods such as directly using the equivalent area diameter and the equivalent ellipse Feret short diameter, that is, c=1.6 is determined in formula (1), and formula (15) is obtained.

[0189]

[0190] S10: Calculate the volume of diabase aggregate particles, and calculate the mass proportion GIR of all particles in the particle size range (divided by the equivalent particle diameter of the particles), and study the influence of the change of the posture of the aggregate particles brought by vibration on GIR.

[0191] The specific process is as follows:

[0192] Randomly sample and weigh 1kg of diabase mixed aggregate (5 particle sizes from small to large are 2.36mm-4.75mm, 4.75mm-7mm, 7mm-9.5mm, 9.5mm-13.2mm, and 13.2mm-16mm). The grading curve of the mixed aggregate is obtained by the traditional sieving method, and the grading curve is obtained by the deep learning model for comparative analysis. Due to the different vertical projection areas of particles in different postures, the grading result of the deep learning model is affected, the particle posture transformation is completed by the vibration of the test platform, and the image is photographed by the deep learning model for grading analysis to study the influence of vibration frequency on grading. According to the requirements of the deep learning model training, the particles are laid on the designed feeding platform (as much as possible to avoid particle stacking), the area size of the feeding platform is S=76.63cm*76.63cm, and vertical shooting is performed at a fixed position. The same batch of aggregate is vibrated by the rolling of the pulley at the bottom of the feeding platform, and the original aggregate image is obtained by shooting once for each vibration. In order to ensure that the shooting image is not at an angle with the ground, the angle correction of the original image is completed in the Photoshop software, and then input to the deep learning model for identification and segmentation.

[0193] To ensure the accuracy and reliability of the test, six batches of aggregate samples were designed for six groups of tests, with 1 kg of aggregate sample taken from each batch. The third and fourth groups and the fifth and sixth groups were parallel tests with the same design gradation. For each group, the sieve test was first completed to obtain the sieve gradation data, and then 10 vibration operations were completed to obtain 10 images for preprocessing before inputting into the deep learning model for identification and segmentation to obtain the aggregate gradation data GIR. The gradation data of one test in the six groups and the segmentation results are shown in Table 7.

[0194] Table 7 Aggregate particle sieve gradation data and deep learning automatic identification gradation data of the first batch

[0195]

[0196] The corresponding gradation curve diagram is shown in Figure 11 From the data in Table 7, it can be seen that as the number of vibrations increases, the difference between the data generated by each vibration is small (it is also found in practice that the particle posture changes little after each vibration, and most particles exist in the same stable state), and the gradation analysis results do not become closer to the sieve test results. At the same time, the values of OWD and ED do not show a gradual decreasing trend with the increase of the number of vibrations, but show irregular fluctuations. This may be due to the fact that when the mass of the aggregate used in the analysis of the gradation is greater than a certain value, different aggregate particles exhibit different representative stacking states. Thus, the error in the gradation analysis caused by the stacking state of the aggregate is reduced. Therefore, it can be known that when the image analysis of the aggregate gradation based on deep learning proposed in the present application is used, if the mass of the aggregate used is not less than 1 kg, only one image of the stacking state is needed (without multiple images of different stacking states).

[0197] S11: Calculate the model identification accuracy of diabase aggregate particles with different clay contents.

[0198] According to the “Standard for Construction Pebbles and Crushed Stones”, the clay content of sand is not greater than 5.0%, and the clay content of coarse aggregate (crushed stone or pebble) is not greater than 2.0%. The above tests are all aimed at different types of washed aggregate, while the original aggregate is natural diabase aggregate crushed stone, and the measured clay content is 1.9% (100g-98.1g) / 100g=1.9%. Therefore, this test adds different proportions of clay content after washing the original aggregate, compares the deep learning identification and segmentation effects of aggregates with different clay contents, and studies the influence of clay content on the accuracy of the deep learning model.

[0199] In actual engineering, the clay content is generally not higher than 5%, so four groups of parallel tests were designed for model identification and segmentation of aggregates with clay contents of 1%, 2%, 3%, 4%, 5%, and 6%, respectively. The results are shown in Figure 12 , and the data are shown in Table 8.

[0200] Table 8: Number of diabase aggregate particles identified at different clay contents

[0201]

[0202] Figure 12 Neutron image a is an image of the upper left corner of the particle in the aggregate image at different clay contents, magnified 10 times, Figure 12 Neutron image b is the original image of the aggregate, Figure 12 Neutron image c is the identification result of the aggregate particles. By Figure 12 It can be seen that the model shows strong robustness to aggregate with different clay contents, with low misidentification and missed identification probability. As shown in Table 8, under different clay content conditions, the number of original and identified particles is less than or equal to 2, except that the number of identified particles of the second group of aggregate with 3% clay content is 3 more than the original number. The IoU 75 is stable and greater than 0.98, and most of the values are 1. In summary, the clay content has little effect on the recognition and segmentation accuracy of the model, and the effect of clay content on the recognition and segmentation accuracy of the model can be ignored in practice.

[0203] S12: Through transfer learning, standard data sets are prepared and models are trained in aggregate images of other common materials (limestone, basalt, pebble), and gradation analysis that meets the accuracy requirements can be obtained, verifying the transferability and generalizability of the original model.

[0204] S12_1: Limestone aggregate particles

[0205] (1) Limestone aggregate particle identification

[0206] First, the deep learning model of diabase is used through transfer learning to identify and segment limestone, and the identification and segmentation results are obtained. Second, 10 limestone aggregate images (about 3000 particles) are collected for expert annotation to form an initial limestone data set, and the data set is expanded through data augmentation technology to form a standard limestone data set. Train the model on the standard data set to obtain the optimal limestone model for identification and segmentation. Randomly select 3 limestone images, compare the identification results before and after training the model, and evaluate the indicators AP 50 , AP 75 , AP 80 as shown in Table 9:

[0207] Table 9: Limestone aggregate image data before and after training based on transfer learning

[0208]

[0209] The identification results are shown in Figures 13(a) and 13(b). As shown in Table 9, the untrained result is basically around 0.98, and the AP50 ,AP 75 ,AP 80 (Except for the AP 80 of 0.98) are all improved to 1. As can be seen from FIG. 13(a) and FIG. 13(b), before training, by using the deep learning model trained for diabase for direct identification and segmentation of limestone through transfer learning, most of the limestone aggregate particles can be identified and segmented, and occasionally some particles are misidentified and missed, and the edge segmentation accuracy is not high (misidentification at a in FIG. 13, missed identification at b, and low edge segmentation accuracy at c); after using the trained limestone optimal model, misidentification and missed identification are reduced, and the edge segmentation effect is obviously improved. In summary, training the model on the limestone standard dataset through transfer learning can improve the accuracy of the model in identifying limestone aggregate.

[0210] (2) Limestone aggregate gradation identification

[0211] The sampled limestone aggregate 100g (containing different particle sizes) is shown in FIG. 14(a) and FIG. 14(b). The model identification and segmentation are completed by taking 10 images after 10 times of vibration, and the gradation analysis is performed. The gradation data of the 10 images is shown in Table 10,

[0212] Table 10 Mass percentage of limestone aggregate in different particle size ranges

[0213]

[0214]

[0215] The gradation curve and the sieving curve of the 10 images are shown in FIG. 15. Figure 15 As can be seen from Table 9, among the mass percentages of the five different particle sizes identified by the deep learning model, except for the smallest particle size 2.36-4.75, the mass percentage (the maximum value is 3.89% after 10 times of vibration) is slightly smaller than the mass percentage 4.00% in the sieving method, and the other four mass percentages are slightly fluctuating around the sieving mass percentage and close to it. As can be seen from Table 9, the particle gradation curve formed by 10 times of vibration fluctuates around the sieving curve, but is within a reasonable error range, indicating that the limestone exists in a relatively stable state, and the vibration has little effect on the gradation of the limestone, and the effect of vibration on the model accuracy can be ignored in practice.

[0216] S12_2: Pebble aggregate particles:

[0217] (1) Pebble aggregate particle identification

[0218] Pebble is a particle with a particle size greater than 5mm formed by rock through natural conditions in nature, and the size, shape, color and texture characteristics are different, and it is difficult to consider various characteristics by using traditional image processing algorithm for particle recognition, resulting in poor recognition effect, but the deep learning model has migration, and shows strong robustness to texture, light, color and other characteristics, and the application of the model to the pebble aggregate is applied.

[0219] Similarly, the deep learning model of diabase is directly used for identification and segmentation of pebble aggregate by using the transfer learning technology, then 10 pebble aggregate images (about 3000 particles) are expertly labeled to form an initial pebble data set, and the data set is expanded by data enhancement technology to form a standard pebble data set. The optimal pebble model is obtained by training the model on the standard pebble data set, and the identification and segmentation results of the untrained model and the trained model are given by randomly selecting images, and the training is improved to 1.

[0220] (2) Pebble aggregate grading identification

[0221] The experiment also shows that the vibration has little effect on the pebble grading, and the grading curve formed after each vibration is below the screening curve, but is within a reasonable error range.

[0222] In view of the fact that the traditional screening detection method is time-consuming and labor-consuming and cannot obtain the result in real time, and the analysis accuracy of the traditional image analysis method is limited, the application establishes an efficient and accurate aggregate particle identification method and aggregate grading analysis method based on deep learning, the grading analysis result of the method is close to the screening result, and the traditional screening method can be effectively replaced to obtain the grading data of the aggregate in real time. At the same time, by using the transfer learning technology and increasing a certain model training, the application can identify the aggregate particles of other common materials and obtain the qualitative and quantitative results of the grading analysis.

Claims

1. A method for automatic analysis of aggregate particle identification and grading based on deep learning, characterized in that The method comprises the following steps: Step 1: making a standard data set of the corresponding material aggregate, dividing the standard data set into a training set, a verification set and a test set, training the deep learning model Mask R-CNN in the training set, and obtaining the optimal deep learning model of the deep learning model of the material aggregate; Step 2: determining the test platform area meeting the accuracy requirement according to the recognition accuracy and area relationship of the smallest particle size group in the aggregate; Step 3: realizing automatic recognition and segmentation of aggregate particles based on segmentation and splicing technology of large images; Step 4: calculating the equivalent particle diameter of the aggregate particles, determining the judgment coefficient through the particle shape, and calculating the equivalent particle diameter of each particle in the aggregate; Step 5: calculating the volume of the aggregate particles, dividing according to the equivalent particle diameter of the particles, and calculating the mass proportion of all particles in the particle size range, i.e. the gradation of the aggregate; Step 6: evaluating the influence of the clay content on the model accuracy and gradation analysis through aggregate particle tests of different clay contents; Step 7: making a standard image data set of other material aggregates, and increasing model training on the basis of the model by using transfer learning to obtain the optimal model of other materials; In step 4, the equivalent area diameter is used as the particle diameter for the class-elliptical shape, and the equivalent elliptical Feret short diameter is used as the particle diameter for the class-needle shape. The specific steps are as follows: First, the projection area of the aggregate particles is recognized and segmented by the optimal deep learning model, then the minimum circumscribed rectangle of the projection area is calculated, and finally the ratio of the long and short axes of the minimum circumscribed rectangle is obtained. The ratio of the long and short axes of the class-elliptical shape and the class-needle shape is determined through experiments as the judgment coefficient c of the particle shape, and the equivalent particle diameter D of each particle in the aggregate is finally calculated, as shown in formula (1); （1）； In formula (1): D Dv50 is the equivalent particle diameter, a is the short axis of the minimum circumscribed rectangle, b is the long axis of the minimum circumscribed rectangle, and c is a determination coefficient determined by experiment. The projected area of the particle in step five A The Feret short diameter multiplied by the projected area R， The equivalent volume obtained is closest to the actual volume of the particle, so this method is used to calculate the volume of the aggregate particle, as shown in equation (2); （2）； In formula (2), denotes the projected area of the aggregate particle, R denotes the Feret short diameter corresponding to the projected area; the total volume of the aggregate particles in each particle size range is calculated, and the mass proportion GIR (Gradation in Range), i.e. the aggregate gradation, is obtained, as shown in formula (3): （3）； In formula (3), denotes the mass of aggregate particles, denotes the projected area of aggregate particles, denotes the Feret short diameter corresponding to the projected area, m and n respectively denote the total number of aggregate particles and the number of particles in a certain particle size range, denotes the particle density, denotes the particle volume, i and j respectively denote the particle serial number in a certain particle size range and the particle serial number of the total particles.

2. The aggregate particle recognition and gradation automatic analysis method based on deep learning according to claim 1, characterized in that: Step 1 comprises the following steps: S1.1: First, collect the aggregate images of the material with different particle sizes, obtain the original data set and complete the preprocessing, then label the original data set in the labeling software, and expand the data set by data enhancement technology to obtain the standard data set of the material; S1.2: Divide the standard data set into a training set, a verification set and a test set, train the deep learning model Mask R-CNN in the training set, verify in the verification set, and test in the test set, finally obtain the optimal hyperparameters and the optimal deep learning model of the deep learning model Mask R-CNN of the material aggregate, and download and save the optimal hyperparameters and the optimal deep learning model.

3. The aggregate particle recognition and gradation automatic analysis method based on deep learning according to claim 1, characterized in that: Step 2 comprises the following steps: S2.1: randomly select the smallest particle size aggregate and place it in a square area test platform with side length a0; S2.2: without changing the same batch of aggregate, sequentially place it in a smaller area by a certain proportion, and find the minimum area S0 that meets the accuracy requirement through experiments, which is 70% of the original area.

4. The aggregate particle recognition and gradation automatic analysis method based on deep learning according to claim 1, characterized in that: The step three comprises the following steps: S3.1: According to the size of the test platform area, design and implement the aggregate feeding platform device with vibration function, and shoot to obtain images; three groups of parallel tests are performed each time, and N high-definition images are obtained by vibrating N times for each group; S3.2: The image segmentation technology is used to divide one super-large image into independent multiple local images, which are respectively identified and segmented, and then the image splicing technology is used to combine into the original image size, and finally the result image is output.

5. The automatic aggregate particle identification and grading analysis method based on deep learning according to claim 1, characterized in that: Through aggregate vibration test, the influence of vibration on grading is studied to obtain the vibration frequency which does not affect the accuracy of grading: The feeding platform of the aggregate is designed, the mixed aggregate diabase is put on the feeding platform, and vibration is completed for 10 times, 1 image is obtained by sound control shooting for each vibration, 10 images are obtained in total, the aggregate particle gradation in each image is analyzed in sequence GIR , and the average value is accumulated GIR The accumulated gradation average value obtained after several vibrations is completely stable, close to the screening gradation, and high in accuracy.

6. The automatic aggregate particle identification and grading analysis method based on deep learning according to claim 1, characterized in that: In step seven, through the transfer learning method, standard data sets are made in the aggregate images of other common materials: diabase, limestone, basalt, and pebbles, and the deep learning model Mask R-CNN is trained, and the grading analysis meeting the accuracy requirements can be obtained, verifying the transferability and generalizability of the optimal deep learning model Mask R-CNN.

7. The automatic aggregate particle identification and grading analysis method based on deep learning according to any one of claims 1-6, characterized in that: In the method, the evaluation indexes include: absolute error , mean absolute percentage error , mean intersection over union , mean pixel accuracy , Euclidean distance , one-way distance OWD , and the specific calculation methods are as follows: 1) Absolute Error (Absolute Error): The difference between the true value number and the identified number is used as a key evaluation index, and the smaller the difference is, the better the performance is; （4）； In formula (4), denotes the true value, denotes the actual value; 2) Mean Absolute Percentage Error (MAPE): The ratio of the absolute value of all sample errors to the actual value, the closer the index is to 0, the more accurate the model is, and the better the model performance is （5）； In formula (5), denotes a true value, denotes an actual value; 3) Intersection over Union (Intersection over Union): Pixel evaluation index, the ratio of the intersection pixel value of the aggregate particle image identified by deep learning to the union pixel value of the aggregate true value image, the closer to 1 the accuracy is; that is, the intersection of the predicted area and the actual area divided by the union of the predicted area and the actual area, simply referred to as intersection-union ratio; calculate the proportion of intersection and union between two sets, which is the proportion of ground truth (Ground Truth) and prediction value in image segmentation; （6）； mean intersection over union (mean Intersection over Union) First, calculate the intersection-union ratio of each class, and then average to get the average intersection-union ratio of all classes; （7）； 4) Average Precision (Average Precision): Adopting the precision mean The overall performance of the model is evaluated; is the average precision of predicting a single target class, calculated from precision and recall, with precision Precision as the ordinate and recall Recall as the abscissa, a curve P-R (Precision-recall) is drawn, that is, P-R the integral of the curve; Precision is the ratio of the number of positive samples of particles in all samples labeled as particles in the test set; Recall is the proportion of all positive samples of particles in the test set that are correctly identified as particles; Precision and Recall are calculated as follows: （8）； （9）； wherein, TP the number of particles that are actually particles and are predicted by the model to be particles, i.e. true positives detected as true positives; FP the number of particles that are actually background but are predicted by the model to be particles, i.e. false positives detected as true positives; FN the number of particles that are actually particles but are not identified by the model as particles, i.e. true positives not detected as true positives; mean average precision the mean average precision for each class the arithmetic mean of the mean average precision for each class the arithmetic mean of the mean average precision for each class, the classes being divided into particles and background （10）； wherein, K is the number of classes; 5) Euclidean distance (Euclidean Distance): is the distance definition index, which refers to the real distance between two points in n-dimensional space, or the natural length of a vector; by calculating the distance between points on different curves , the closeness of the curve is represented, ED The smaller the value, the closer the curve, on the contrary, the farther the curve deviates; The distance between two points is ​ （11）； 6) One-way distance OWD (One Way Distance): OWD refers to the area surrounded by two trajectories; when the area is large, the distance between the trajectories is far, and the similarity is low; on the contrary, if the enclosed area is 0, the two trajectories coincide, and the similarity is the highest, which can be calculated according to the formula; （12）； （13）； Where T1 and T2 are two trajectories, and p is a discrete point on T1.

Images