CN111523616A - Coarse-grained soil filler grading identification method based on convolutional neural network and application system - Google Patents
Coarse-grained soil filler grading identification method based on convolutional neural network and application system Download PDFInfo
- Publication number
- CN111523616A CN111523616A CN202010404996.8A CN202010404996A CN111523616A CN 111523616 A CN111523616 A CN 111523616A CN 202010404996 A CN202010404996 A CN 202010404996A CN 111523616 A CN111523616 A CN 111523616A
- Authority
- CN
- China
- Prior art keywords
- neural network
- particle size
- coarse
- filler
- image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to a coarse-grained soil filler grading identification method based on a convolutional neural network and an application system thereof, wherein a first neural network and a plurality of second neural networks are constructed; collecting a field filler image, and outputting a classification result of a particle size range by adopting a first neural network; inputting the filler image and the classification result into a second neural network with corresponding particle size range to obtain the mass of the coarse-grained soil in each single particle size range; and (4) counting the mass distribution of the coarse-grained soil in various particle size ranges to obtain gradation and a gradation curve. According to the invention, the particle size range is obtained through the first neural network, the total mass of the single particle size group is obtained through the corresponding second neural network, the accuracy of mass calculation is improved, and the accuracy of grading calculation is further ensured. And the plurality of second neural networks are processed in parallel, so that the training efficiency and the field processing efficiency of the second neural networks are ensured. The automatic degree is high, complex image processing algorithms are not needed, manual intervention is not needed, experience of operators is not relied on, the environmental suitability is strong, and the precision is high.
Description
Technical Field
The invention relates to the technical field of coarse-grained soil roadbed filling and rolling construction and quality control thereof, in particular to a coarse-grained soil filler grading identification method based on a convolutional neural network and an application system.
Background
The coarse-grained soil filler is used as an important component of a railway ballast bed and a roadbed and consists of coarse and fine particles with different sizes, the particle size range of the particles is large, the characteristic difference of the coarse and fine particles is large, the coarse and fine particles are generally in a single-particle structure, small particles fill gaps formed by large particles, the coarse-grained soil with good gradation has higher compactness than coarse-grained soil with poor gradation, and the coarse-grained soil is easier to compact during filling and rolling construction. The grading of coarse-grained soil greatly influences the service performances such as strength, stability, deformability, durability and the like of the railway track bed and the roadbed. Along with the rapid development of railway transportation industry, the aspects of construction, operation management, maintenance and repair of railway infrastructure tend to be more and more intelligent, grading is used as an important physical index of coarse-grained soil filler and is mainly determined by a manual or mechanical screening method for a long time, and along with the gradual introduction of an intelligent rolling technology and equipment in the filling process of a railway roadbed, how to identify the grading of the coarse-grained soil filler before and after rolling in real time, rapidly, intelligently, efficiently and accurately and predict the modulus, damping and other mechanical parameters of the coarse-grained soil filler, so that the real-time, intelligent and accurate evaluation of the rolling quality during filling construction is the key of the intelligent rolling and filling technology.
Currently, the grading of coarse-grained soil is mainly determined by a screening method, the coarse-grained soil needing to be graded is screened according to the size and the combination of sieve pores given by specifications, and the grading curve of the coarse-grained soil is determined according to the mass ratio of particles left on each sieve pore. Although the method is mature in technology, wide in application and simple in operation, the method is time-consuming and low in efficiency, coarse soil needs to be sampled and sieved, and the abrasion and crushing of coarse particles and the deformation of sieve pores in the sieving process can change the original grading of the coarse soil. During manual or mechanical screening, the filler is screened one by one according to the size sequence of the screen holes, the filler is required to continuously vibrate on the screen surface in the horizontal and vertical directions simultaneously, and filler particles smaller than the size of the screen holes pass through the screen holes until the mass of the filler particles passing through the screen holes is less than 0.1 percent of the residual mass on the screen surface within 1 min. When the sieve shaker is used for sieving, after the sieve shaker is used for sieving, the sieve is supplemented manually one by one, the sieved particles are merged into the next sieve and sieved together with the sample in the next sieve, and the sieving is sequentially carried out until all the sieves are sieved. It should be confirmed that the mass passing through the sieve holes within 1min is actually less than 0.1% of the residue on the sieve, the sieve residue on each sieve is weighed to the nearest 0.1% of the total mass, the percent of the net residue on the basis of the total mass is nearest 0.1%, the cumulative percent of the net residue is nearest 0.1%, and the percent passing through each sieve is nearest 0.1%. Grading is plotted against the cumulative percent pass of each mesh to determine the grading of the coarse soil. The screening method mainly has the following disadvantages:
(1) the conventional screening test needs to collect soil samples at different positions on the site for testing, can destroy the integrity and working state of the soil body on the site, is time-consuming and labor-consuming, and therefore cannot realize real-time, quick, efficient and intelligent identification of coarse-grained soil gradation during intelligent rolling of the railway roadbed.
(2) The screening test has long time consumption and low efficiency, and can not realize intelligent, automatic and quick identification and analysis.
(3) During the screening process, due to the vibration of the vibrating screen, part of coarse particles are broken, and the accuracy of the result is affected.
(4) The sieve mesh can take place certain degree deformation in long-time use, causes the experimental error.
With the continuous development of the technology, some filler gradation determination methods based on digital image recognition also appear. An image analysis method represented by Wipfrag software is generally used for measuring the particle size of the crushed stone, and finally, a grading curve is directly output. The main reason for the low precision is that the existence of more fine-grained soil causes certain interference to the analysis of coarse-grained soil. The ImageJ software can also perform image segmentation, and then calculate the particle size of each particle by using Excel, but the accuracy of the result is low, mainly because the image segmentation has large errors. In addition, the grading curve describes the relationship between the particle size and the particle mass, and the image analysis obtains the direct relationship between the particle size or the two-dimensional area of the particle and the number of the particles, and how to realize the conversion between the two relationships still needs to be solved.
Disclosure of Invention
Aiming at the problems of long time consumption, field state disturbance, need of manual intervention and the like in the existing grading determination, a coarse-grained soil filler grading identification method and an application system based on a deep learning convolutional neural network are provided.
In order to achieve the aim, the invention provides a coarse-grained soil filler grading automatic identification method based on a convolutional neural network, which comprises the following steps:
collecting filler images containing coarse soil particles with various particle sizes, and marking particle size ranges to form a first sample library; constructing a second sample library containing samples in various particle size ranges by using a filler image containing coarse soil particles in a single particle size range and marking the quality as a sample;
constructing a first neural network and training by adopting samples in a first sample library;
constructing a plurality of second neural networks, and respectively training by taking the filler images of the coarse-grained soil particles with different grain size ranges in a second sample library as samples;
counting the mass distribution of coarse-grained soil in various particle size ranges by a probability statistical algorithm of a statistical module, and outputting grading and grading curves;
packaging the first neural network, the second neural network and the statistical module into a program capable of operating independently;
acquiring a filler image of a roadbed filling and rolling construction site, outputting a particle size range classification result of particles after the filler image is processed by a first neural network after being packaged, and inputting the filler image of the construction site and the classification result of the particle size range into a second neural network after being packaged in a corresponding particle size range to obtain different particle size range information and corresponding particle quality of coarse-grained soil in the filler image of the construction site; and the packaged statistic module outputs the grading and the grading curve.
Further, the second sample library is obtained by artificially preparing a plurality of groups of coarse-grained soil fillers only containing particles within a particle size range and shooting at different positions and different angles.
Further, the first neural network comprises three convolutional layers, three full-connection layers and a classifier; deleting negative values from the first two convolution layers through the corrected linear units or the ReLU, and performing down-sampling through the maximum pooling layer; and the first two full-connection layers adopt ReLU to delete negative values, and finally, a classification result of the particle size range is output through a classifier.
Further, the training of the first neural network comprises:
(1) selecting a training image, a verification image and a test image from a first sample library; inputting a training image;
(2) initializing parameters in a first neural network;
(3) performing a round of training;
(4) judging whether the error grade is accepted, and if so, outputting a training model; if not, entering the step (5);
(5) setting the iteration times to be 0;
(6) judging whether the iteration times are the maximum iteration times, if not, entering the step (7); if the verification image arrives, the verification image is used for verification, and the step (4) is returned;
(7) and (4) inputting the test image into the first neural network, outputting a result by the first neural network, calculating an output error, improving the definition of the test image or denoising the test image, adding 1 to the iteration number, and returning to the step (6).
Further, the second neural network comprises three convolutional layers and three fully-connected layers; deleting negative values from the first two convolution layers through the corrected linear units or the ReLU, and performing down-sampling through the maximum pooling layer; and the first two full connection layers adopt ReLU to delete negative values, and the output outputs the quality through a classifier.
Further, the training of the second neural network comprises:
(1) selecting images in the same grain size range from a second sample library as training images, verification images and test images; inputting a training image;
(2) initializing parameters in a second neural network;
(3) performing a round of training;
(4) judging whether the error grade is accepted, namely whether the output particle size quality error is accepted, and if so, outputting a training model to finish training; if not, entering the step (5);
(5) setting the iteration times to be 0;
(6) judging whether the iteration times are the maximum iteration times, if not, entering the step (7); if the verification image arrives, the verification image is used for verification, and the step (4) is returned;
(7) and (4) inputting the test image into a second neural network, outputting a result through the second neural network, calculating an output error, improving the definition of the test image or denoising the test image, adding 1 to the iteration number, and returning to the step (6).
The invention provides a coarse-grained soil filler grading identification system based on a convolutional neural network, which comprises an acquisition module, a first neural network module, a plurality of second neural network modules and a statistical module, wherein the acquisition module is used for acquiring a coarse-grained soil filler grading identification signal;
the acquisition module acquires a filler image of a construction site;
the first neural network module outputs classification results of different particle size ranges based on the filler image collected on the construction site, and the calculated particle size ranges and the filler image on the construction site are input to the second neural network module in the corresponding particle size range;
each second neural network module outputs the particle size information and the quality of coarse soil particles in the input filler image of the construction site;
and the statistical module is used for counting the particle size information and the quality of the particles output by each second neural network module to obtain the mass distribution of the coarse soil particles in various particle size ranges and outputting grading and grading curves.
Further, the first neural network module is internally provided with a first neural network; collecting filler images containing coarse soil particles in various particle size ranges, and marking the particle size ranges to form a first sample library; the first neural network is obtained by adopting sample training in a first sample library.
Further, a second neural network is built in the second neural network module; constructing a second sample library containing samples in various particle size ranges by taking a filler image containing coarse soil particles in a single particle size range as a sample; and each second neural network is obtained by training by respectively adopting the filler images of the coarse soil particles with different particle size ranges in the second sample library as samples.
Further, the first neural network comprises three convolutional layers, three full-connection layers and a classifier; deleting negative values from the first two convolution layers through the corrected linear units or the ReLU, and performing down-sampling through the maximum pooling layer; the first two full-connection layers adopt ReLU to delete negative values, and finally, a classifier is used for outputting a classification result of a particle size range;
the first two convolutional layers are subjected to linear unit or ReLU correction to delete negative values, and then are subjected to down-sampling through the maximum pooling layer; and the first two full connection layers adopt ReLU to delete negative values, and the output outputs the quality through a classifier.
The technical scheme of the invention has the following beneficial technical effects:
(1) according to the method, the particle size range of the particles in the coarse-grained soil filler image is obtained through the first neural network, the mass of the particles in each particle size range is obtained through the corresponding second neural network, the accuracy of mass calculation is improved, and the accuracy of grading calculation is further ensured. And the plurality of second neural networks are processed in parallel, so that the training efficiency and the field processing efficiency of the second neural networks are ensured.
(2) The coarse-grained soil filler grading identification method based on the deep learning convolutional neural network and the application system have strong adaptability, the initial modeling process may need certain time and energy, and once the initial training is completed, the migration learning can be used for perfecting the model so as to adapt to the condition which may have different characteristics from the original training scene.
(3) The coarse-grained soil filler grading identification method based on the deep learning convolutional neural network and the application system thereof have high identification efficiency and high image processing speed on site. According to the preliminary research results, the analysis and classification of the field images should not exceed a few seconds, and the real-time and rapid identification of the gradation can be realized. The method may be deployed in a localized or distributed mode (or both). The well-trained model can be applied to the filler image collected in the roadbed filling and rolling construction site only by proper computing power. Thus, the grading identification can be calculated immediately by a system installed on the field rolling equipment, and the image can be sent to a centralized or cloud system for processing.
(4) High automation, other classical image processing methods, such as object detection, etc., usually involve human-human or human-machine interactions. When the program is implemented, the system proposed by the invention will be fully automatic, without human intervention.
(5) The method is used for establishing and training the neural network based on the shot image and the deep learning convolutional neural network architecture and parameters thereof, finally obtaining the grading recognition model, further expanding the sample library in field application and further realizing continuous verification and perfection of the model. The range of grading categories can be further increased to improve the accuracy of the model, and the calculation speed can be further improved along with the increase of the training times.
Drawings
FIG. 1 is a first neural network architecture;
FIG. 2 is a second neural network architecture;
FIG. 3 is a schematic diagram of a first neural network training process;
FIG. 4 is a schematic diagram of the operation of the first and second neural networks;
FIG. 5 is a flow of training and processing of the first and second neural networks in the embodiment.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail with reference to the accompanying drawings in conjunction with the following detailed description. It should be understood that the description is intended to be exemplary only, and is not intended to limit the scope of the present invention. Moreover, in the following description, descriptions of well-known structures and techniques are omitted so as to not unnecessarily obscure the concepts of the present invention.
The invention provides a coarse-grained soil roadbed filler grading identification method based on a deep learning convolutional neural network, which is used for evaluating the size distribution (grading) of coarse-grained soil fillers.
Firstly, constructing a training sample library
Shooting the coarse-grained soil filler of the same gradation from different angles, positions and the like, and collecting images of the coarse-grained soil filler with different grain sizes; images of coarse-grained soil filler combinations under various grading are shot according to different angles, and coarse-grained soil filler image libraries covering various different sizes, different angles and different grading combinations are formed. Different parts of the same image may also be divided into a plurality of sub-image samples; and marking the filling images of the coarse soil particles with various particle sizes according to the particle size range to form a first sample library. Manually preparing coarse-grained soil fillers only containing one grain size group, shooting images of the coarse-grained soil fillers under various grain sizes according to different angles to form a coarse-grained soil filler image library covering various different sizes, different angles and different grain size groups, and marking the quality of the filler images of the coarse-grained soil particles of each grain size group to form a second sample library.
Second, construct the first neural network and train
With reference to fig. 1, feature extraction is performed using three convolutional layers, the first two convolutional layers are followed by a modified linear unit or ReLU to remove negative values, a maximum Pooling layer (Pooling layer) summarizes the characteristics found by each convolution step, three fully-connected layers are followed by the three convolutional layers, and the first two fully-connected layers use the ReLU to remove negative values. And finally, outputting a classification result of the particle size range through a classifier.
The input of the first neural network is a filler image collected (the gray scale is 32x 32), the particle size range in the image is identified through the first neural network, and the particle size range in the image is output.
Training is performed by using samples in the first sample library, and with reference to fig. 3, the training process includes:
(1) the first sample library comprises training images, verification images and test images, and the proportion of the training images, the verification images and the test images is 7:2: 1; inputting a training image;
(2) initializing parameters in a first neural network;
(3) performing a first round of training; performing iterative computation by using initial parameters set by the model in the first round of training process, and performing first comparison and verification;
(4) judging whether the error grade is accepted, namely whether the judgment result of the range of the particle size is accepted, if so, outputting a training model and finishing the training; if not, entering the step (5);
(5) adjusting the training parameters and setting the iteration times to be 0; the model parameters mentioned here include not only Convolutional Neural Network (CNN) architecture (e.g., number of layers, size, and number of kernels in each layer), but also loss function used to measure errors during training, learning rate parameters, and simple processing such as gray conversion and denoising of input images before network training. The adjustment may be achieved by increasing or decreasing these parameters.
(6) Judging whether the iteration times are the maximum iteration times, if not, entering the step (7); if the verification image arrives, the verification image is used for verification, and the step (4) is returned;
(7) and (5) inputting the test image into the first neural network and outputting a result, calculating an output error, adjusting the picture quality, adding 1 to the iteration number, and returning to the step (6).
The image quality is mainly adjusted by processing pixels, noise and the like, and although complex image processing algorithms and technologies are not needed, the neural network can identify the image more accurately.
And obtaining a first neural network meeting the error requirement after training, classifying the particle size range of the samples in the first sample library, and classifying and dividing the particle size range of the filler images acquired on site during on-site application.
In order to obtain a machine learning result with high accuracy, the number of image samples must be sufficiently large. During the initial exploration, 200 source images (about 3600 x 2400 resolution) were used for each level of analysis, from which about 8000 sub-images were extracted as the first sample library.
A certain percentage of image samples are randomly reserved for each grading class for verification, and the proportion of a training image, a verification image and a test image is suggested to be 7:2:1 in the scheme at present.
Thirdly, constructing a second neural network and training
FIG. 2 is a diagram of a neural network architecture of a second neural network according to the present invention. The architecture is a relatively simple architecture with three convolutional layers, the first two followed by a modified linear unit or ReLU to remove negative values, a max Pooling layer (Pooling layer) that summarizes the properties found by each convolution step, and three fully-connected layers after the convolutional layer. The total mass of the particles is finally output. The input of the model is a gray-scale image with 32x32 pixels, and the output is the mass of coarse soil particles in the image;
the second neural network is multiple, each model corresponds to a particle size range, and images in the particle size range are identified.
The coarse-grained soil containing only one grain size group of grains is artificially prepared, and a high-definition camera is adopted to take a plurality of groups of pictures at different positions and different angles. And preparing a certain amount of samples for each particle size group, and manually marking the quality.
The second neural network in each particle size range is trained through the samples respectively, the training process is the same, and the method comprises the following steps:
(1) selecting an image in a grain size range from a second sample library as a training image, a verification image and a test image; the ratio is 7:2: 1; inputting a training image;
(2) initializing parameters in a second neural network;
(3) performing a round of training;
(4) judging whether the error grade is accepted, namely whether the output particle size quality error is accepted, and if so, outputting a training model to finish training; if not, entering the step (5);
(5) adjusting training parameters, and setting the iteration number to be 0;
(6) judging whether the iteration times are the maximum iteration times, if not, entering the step (7); if the verification image arrives, the verification image is used for verification, and the step (4) is returned;
(7) and (4) inputting the test image into a second neural network, outputting a result through the second neural network, calculating an output error, adjusting the picture quality, adding 1 to the iteration number, and returning to the step (6).
In one embodiment, the quality is used as input information in the training stage, the volume of the particles in the particle size group can be obtained through conversion of the weight through the input quality information, and the second neural network can further calculate the area of the particles in the particle size group on the assumption that the third size of the particles is half of the sum of the maximum particle size and the minimum particle size of the particle size group, so as to obtain the proportion of the area of the particles in the image area. According to the training information, in the verification and application process, the second neural network can identify the area of a single particle size range group in the mixed particle size image through the input image, then calculates the particle area of the single particle size range through the particle area ratio, converts the particle area into the volume and the mass, and outputs the particle size information and the mass of the particles of the particle size group.
In another embodiment, the second sample library is not labeled with quality, but is obtained by a second neural network calculation.
The specific operation process is as follows:
the image of the coarse particle filler, which contains only a single particle size range, is input to a neural network, which fits the particles in the image using a series of circular templates whose diameters are approximately equal to the mean of the major and minor axes of the particles. When the circular template reaches the center position of the particle successfully fitted with the circular template, the circular template stays at the position, the neural network adopts the next circular template to fit the rest particles in the image, and the fitting can automatically cross the particles which are already fitted until the fitting process is finished when all the particles which can be identified are fitted.
And calculating the total area of all the particles by calculating the area of the fitted circle, assuming that the third size of the particles is half of the sum of the maximum particle size and the minimum particle size of the particle size group, further calculating the total volume of the particles, and calculating the total mass of the particles in the image by adopting the weight.
Fourth, statistic module
And obtaining the mass distribution, namely grading, of various particle sizes according to the mass output by each second neural network, and drawing a curve of the change of the mass along with the particle size, namely a grading curve.
The grading calculation adopts an embedded algorithm, such as a probability statistics algorithm in deep learning.
And calculating the mass of each particle size group by using a second neural network, and calculating the mass ratio of each particle size group.
And outputting grading information and a grading curve through the statistics of the mass ratio of the particles in each particle size group.
Referring to fig. 5, in one embodiment, the sizes are divided into n groups, 4.75-9.5 mm,9.5-19mm,19-22.4mm and 22.4-30mm … are used as the first neural network, the sizes of the identified images are in the range of 4.75-22.4mm, and the original images are respectively output to the second neural networks corresponding to the size ranges, namely, three second neural networks of 4.75-9.4mm,9.5-19mm and 19-22.4mm in fig. 5. And the second neural network analysis obtains and outputs the quality of each particle size range. And the grading and grading curve is obtained after statistics by the statistical module.
With training using two grading categories (their particle distributions are very similar), the model can achieve the best accuracy of 93%, and with training of about 12000 images and verification of 3000 images, the accuracy of the model will further improve with further increase in sample volume. In the actual use process, a certain number of grading curves can be selected and filler samples are configured as required, and images are obtained and used for training a neural network.
Fifth, model derivation
And leading out a package of the first neural network, the plurality of second neural networks and the statistical module which meet the precision requirement after training and using the package in field analysis.
Sixth, coarse grain soil subgrade filler grading identification system based on convolutional neural network
The coarse-grained soil subgrade filler grading identification system based on the convolutional neural network comprises an acquisition module, a first neural network module, a plurality of second neural network modules and a statistical module, and is combined with the graph 4.
The acquisition module acquires images of the compacted fillers before and after the intelligent rolling equipment works;
the classification result of the particle size range output by the first neural network module is stimulated by each group of single particle size range and calls a second neural network module with the corresponding particle size;
the second neural network module outputs the particle size information and the quality of the coarse soil particles with the particle size;
and the statistical module is used for counting the particle size information and the quality of the particles output by each second neural network, obtaining the mass distribution, namely the gradation, of the coarse-grained soil particles in various particle sizes, and drawing a curve of the change of the quality along with the particle size, namely a gradation curve.
The system can also comprise a communication module for sending results to an experimenter, or the field device can acquire and send images through a wireless network, and the analysis process can also realize real-time automatic processing based on a server.
Grading information of different compaction stages is analyzed through model calculation. And comparing with the real grading category, outputting an analysis result if the error meets the requirement, returning to the training process of the CNN network if the error does not meet the requirement, and re-training to improve the accuracy until the model can analyze and calculate grading information in different compaction processes with higher accuracy.
The model is mainly applied on site, and if the obtained result error is too large, training is performed to realize model optimization.
Recent developments in deep learning have caused great repercussions in the civil engineering and railroad engineering community, but research in this area is still in an early stage. Most published papers and issued patents describe laboratory studies that are not readily applicable in a real-world environment. One problem is that creating large numbers of labeled training data is costly and the method proposed by the present invention readily identifies specific particle classifications under controlled conditions. To obtain 200 images of a gradation, only one or two hours are required. The high resolution and large size of images of modern digital cameras means that tens of subsets can be extracted from one photograph. In addition, machine learning frameworks generally facilitate data expansion and may generate new training images through various transformations of the raw data.
Training the model is a slow process, but this approach separates learning from application. Meanwhile, a so-called transfer learning method is employed, which starts with a trained model and then learns progressive improvement. During field operation, the model can be optimized efficiently. The trained model can be quickly and conveniently continuously trained by adopting a new image library, the accuracy is continuously improved, the application range is continuously expanded, and the training process does not need to come from the beginning.
In summary, the invention relates to a coarse-grained soil filler grading identification method based on a convolutional neural network and an application system thereof, and a first neural network and a plurality of second neural networks are constructed; collecting a field filler image, and adopting a classification result of a particle size range output by a first neural network; inputting the original image and the particle size range into a second neural network corresponding to the particle size range to obtain the mass of the coarse-grained soil in the group of images input on site; and (4) counting the mass distribution of the coarse-grained soil in various particle size ranges to obtain gradation and a gradation curve. According to the invention, the range of the particle size is obtained through the first neural network, the quality of the particle size is obtained through the corresponding second neural network, the accuracy of quality calculation is improved, and the accuracy of grading calculation is further ensured. And the plurality of second neural networks are processed in parallel, so that the training efficiency and the field processing efficiency of the second neural networks are ensured. The automatic degree is high, complex image processing algorithms are not needed, manual intervention is not needed, experience of operators is not relied on, the environmental suitability is strong, and the precision is high.
It is to be understood that the above-described embodiments of the present invention are merely illustrative of or explaining the principles of the invention and are not to be construed as limiting the invention. Therefore, any modification, equivalent replacement, improvement and the like made without departing from the spirit and scope of the present invention should be included in the protection scope of the present invention. Further, it is intended that the appended claims cover all such variations and modifications as fall within the scope and boundaries of the appended claims or the equivalents of such scope and boundaries.
Claims (10)
1. A coarse-grained soil filler gradation automatic identification method based on a convolutional neural network is characterized by comprising the following steps:
collecting filler images containing coarse soil particles with various particle sizes, and marking particle size ranges to form a first sample library; constructing a second sample library containing samples in various particle size ranges by using a filler image containing coarse soil particles in a single particle size range and marking the quality as a sample;
constructing a first neural network and training by adopting samples in a first sample library;
constructing a plurality of second neural networks, and respectively training by taking the filler images of the coarse-grained soil particles with different grain size ranges in a second sample library as samples;
counting the mass distribution of coarse-grained soil in various particle size ranges by a probability statistical algorithm of a statistical module, and outputting grading and grading curves;
packaging the first neural network, the second neural network and the statistical module into a program capable of operating independently;
acquiring a filler image of a roadbed filling and rolling construction site, outputting a particle size range classification result of particles after the filler image is processed by a first neural network after being packaged, and inputting the filler image of the construction site and the classification result of the particle size range into a second neural network after being packaged in a corresponding particle size range to obtain different particle size range information and corresponding particle quality of coarse-grained soil in the filler image of the construction site; and the packaged statistic module outputs the grading and the grading curve.
2. The convolutional neural network-based automatic coarse-grained soil filler grading identification method according to claim 1, wherein the second sample library is obtained by artificially preparing a plurality of groups of coarse-grained soil fillers containing only particles within a particle size range and shooting at different positions and different angles.
3. The convolutional neural network-based coarse-grained soil filler gradation automatic identification method according to claim 1 or 2, characterized in that the first neural network comprises three convolutional layers, three fully-connected layers and a classifier; deleting negative values from the first two convolution layers through the corrected linear units or the ReLU, and performing down-sampling through the maximum pooling layer; and the first two full-connection layers adopt ReLU to delete negative values, and finally, a classification result of the particle size range is output through a classifier.
4. The convolutional neural network-based coarse-grained soil filler gradation automatic identification method according to claim 1 or 2, characterized in that the training of the first neural network comprises:
(1) selecting a training image, a verification image and a test image from a first sample library; inputting a training image;
(2) initializing parameters in a first neural network;
(3) performing a round of training;
(4) judging whether the error grade is accepted, and if so, outputting a training model; if not, entering the step (5);
(5) setting the iteration times to be 0;
(6) judging whether the iteration times are the maximum iteration times, if not, entering the step (7); if the verification image arrives, the verification image is used for verification, and the step (4) is returned;
(7) and (4) inputting the test image into the first neural network, outputting a result by the first neural network, calculating an output error, improving the definition of the test image or denoising the test image, adding 1 to the iteration number, and returning to the step (6).
5. The convolutional neural network-based coarse-grained soil filler gradation automatic identification method according to claim 1 or 2, characterized in that the second neural network comprises three convolutional layers and three fully-connected layers; deleting negative values from the first two convolution layers through the corrected linear units or the ReLU, and performing down-sampling through the maximum pooling layer; and the first two full connection layers adopt ReLU to delete negative values, and the output outputs the quality through a classifier.
6. The convolutional neural network-based automatic coarse-grained soil filler gradation identification method as claimed in claim 5, wherein the training of the second neural network comprises:
(1) selecting images in the same grain size range from a second sample library as training images, verification images and test images; inputting a training image;
(2) initializing parameters in a second neural network;
(3) performing a round of training;
(4) judging whether the error grade is accepted, namely whether the output particle size quality error is accepted, and if so, outputting a training model to finish training; if not, entering the step (5);
(5) setting the iteration times to be 0;
(6) judging whether the iteration times are the maximum iteration times, if not, entering the step (7); if the verification image arrives, the verification image is used for verification, and the step (4) is returned;
(7) and (4) inputting the test image into a second neural network, outputting a result through the second neural network, calculating an output error, improving the definition of the test image or denoising the test image, adding 1 to the iteration number, and returning to the step (6).
7. A coarse-grained soil filler grading identification system based on a convolutional neural network is characterized by comprising an acquisition module, a first neural network module, a plurality of second neural network modules and a statistical module;
the acquisition module acquires a filler image of a construction site;
the first neural network module outputs classification results of different particle size ranges based on the filler image collected on the construction site, and the calculated particle size ranges and the filler image on the construction site are input to the second neural network module in the corresponding particle size range;
each second neural network module outputs the particle size information and the quality of coarse soil particles in the input filler image of the construction site;
and the statistical module is used for counting the particle size information and the quality of the particles output by each second neural network module to obtain the mass distribution of the coarse soil particles in various particle size ranges and outputting grading and grading curves.
8. The convolutional neural network-based coarse-grained soil filler gradation identification system of claim 7, wherein the first neural network module is built-in with a first neural network; collecting filler images containing coarse soil particles in various particle size ranges, and marking the particle size ranges to form a first sample library; the first neural network is obtained by adopting sample training in a first sample library.
9. The convolutional neural network-based coarse-grained soil filler gradation identification system of claim 8, wherein the second neural network module embeds a second neural network; constructing a second sample library containing samples in various particle size ranges by taking a filler image containing coarse soil particles in a single particle size range as a sample; and each second neural network is obtained by training by respectively adopting the filler images of the coarse soil particles with different particle size ranges in the second sample library as samples.
10. The convolutional neural network-based coarse soil filler gradation identification system of claim 9, wherein the first neural network comprises three convolutional layers, three fully-connected layers, and a classifier; deleting negative values from the first two convolution layers through the corrected linear units or the ReLU, and performing down-sampling through the maximum pooling layer; the first two full-connection layers adopt ReLU to delete negative values, and finally, a classifier is used for outputting a classification result of a particle size range;
the first two convolutional layers are subjected to linear unit or ReLU correction to delete negative values, and then are subjected to down-sampling through the maximum pooling layer; and the first two full connection layers adopt ReLU to delete negative values, and the output outputs the quality through a classifier.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010404996.8A CN111523616B (en) | 2020-05-14 | 2020-05-14 | Coarse-grained soil filler grading identification method based on convolutional neural network and application system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010404996.8A CN111523616B (en) | 2020-05-14 | 2020-05-14 | Coarse-grained soil filler grading identification method based on convolutional neural network and application system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111523616A true CN111523616A (en) | 2020-08-11 |
| CN111523616B CN111523616B (en) | 2021-01-29 |
Family
ID=71905905
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010404996.8A Active CN111523616B (en) | 2020-05-14 | 2020-05-14 | Coarse-grained soil filler grading identification method based on convolutional neural network and application system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111523616B (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112162164A (en) * | 2020-09-24 | 2021-01-01 | 安徽德尔电气集团有限公司 | Cable life prediction system based on neural network |
| CN113129283A (en) * | 2021-04-19 | 2021-07-16 | 河北工业大学 | Digital image-based filler grading intelligent detection compaction control method |
| CN113361693A (en) * | 2021-06-30 | 2021-09-07 | 北京百度网讯科技有限公司 | Method and device for generating convolutional neural network and image recognition method and device |
| CN113487643A (en) * | 2021-07-19 | 2021-10-08 | 华电西藏能源有限公司大古水电分公司 | Cemented sand gravel stock yard sampling and determining method |
| CN113591740A (en) * | 2021-08-04 | 2021-11-02 | 清华大学 | Deep learning-based method and device for identifying silt particles in complex river environment |
| CN114324839A (en) * | 2022-01-07 | 2022-04-12 | 中电建路桥集团有限公司 | Low shallow roadbed construction integrated equipment that fills of high ground water level silt soil road |
| CN114858560A (en) * | 2022-06-06 | 2022-08-05 | 邢台路桥建设总公司 | Method for quickly measuring optimal water content of compacted soil of earthwork roadbed |
| CN114863351A (en) * | 2022-07-07 | 2022-08-05 | 河北工业大学 | Picture and sound fusion roadbed filling collaborative identification management system based on Web3.0 |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107341518A (en) * | 2017-07-07 | 2017-11-10 | 东华理工大学 | A kind of image classification method based on convolutional neural networks |
| CN108804971A (en) * | 2017-04-26 | 2018-11-13 | 联想新视界（天津）科技有限公司 | A kind of image identification system, augmented reality show equipment and image-recognizing method |
| CN109102109A (en) * | 2018-07-16 | 2018-12-28 | 四川大学 | A kind of piece of exploitation of stone explosion gradation prediction technique |
| CN109993766A (en) * | 2019-03-18 | 2019-07-09 | 南京理工大学 | Sandstone image droplet measurement method based on deep learning |
| US20190347467A1 (en) * | 2018-05-10 | 2019-11-14 | Juntendo Educational Foundation | Image analysis method, apparatus, non-transitory computer readable medium, and deep learning algorithm generation method |
| CN110458119A (en) * | 2019-08-15 | 2019-11-15 | 中国水利水电科学研究院 | A kind of aggregate gradation method for quickly identifying of non-contact measurement |
| CN110838128A (en) * | 2019-11-07 | 2020-02-25 | 华侨大学 | Image method aggregate stacking void ratio prediction method and system |
| CN110969636A (en) * | 2019-11-07 | 2020-04-07 | 华侨大学 | Method and system for predicting porosity by grading characterization based on three-dimensional image method measurement |
| CN111024922A (en) * | 2019-09-25 | 2020-04-17 | 中国铁道科学研究院集团有限公司铁道建筑研究所 | High-speed railway roadbed compaction quality continuous detection system and method |
-
2020
- 2020-05-14 CN CN202010404996.8A patent/CN111523616B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108804971A (en) * | 2017-04-26 | 2018-11-13 | 联想新视界（天津）科技有限公司 | A kind of image identification system, augmented reality show equipment and image-recognizing method |
| CN107341518A (en) * | 2017-07-07 | 2017-11-10 | 东华理工大学 | A kind of image classification method based on convolutional neural networks |
| US20190347467A1 (en) * | 2018-05-10 | 2019-11-14 | Juntendo Educational Foundation | Image analysis method, apparatus, non-transitory computer readable medium, and deep learning algorithm generation method |
| CN109102109A (en) * | 2018-07-16 | 2018-12-28 | 四川大学 | A kind of piece of exploitation of stone explosion gradation prediction technique |
| CN109993766A (en) * | 2019-03-18 | 2019-07-09 | 南京理工大学 | Sandstone image droplet measurement method based on deep learning |
| CN110458119A (en) * | 2019-08-15 | 2019-11-15 | 中国水利水电科学研究院 | A kind of aggregate gradation method for quickly identifying of non-contact measurement |
| CN111024922A (en) * | 2019-09-25 | 2020-04-17 | 中国铁道科学研究院集团有限公司铁道建筑研究所 | High-speed railway roadbed compaction quality continuous detection system and method |
| CN110838128A (en) * | 2019-11-07 | 2020-02-25 | 华侨大学 | Image method aggregate stacking void ratio prediction method and system |
| CN110969636A (en) * | 2019-11-07 | 2020-04-07 | 华侨大学 | Method and system for predicting porosity by grading characterization based on three-dimensional image method measurement |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112162164A (en) * | 2020-09-24 | 2021-01-01 | 安徽德尔电气集团有限公司 | Cable life prediction system based on neural network |
| CN113129283B (en) * | 2021-04-19 | 2022-11-29 | 河北工业大学 | Digital image-based filler grading intelligent detection compaction control method |
| CN113129283A (en) * | 2021-04-19 | 2021-07-16 | 河北工业大学 | Digital image-based filler grading intelligent detection compaction control method |
| CN113361693A (en) * | 2021-06-30 | 2021-09-07 | 北京百度网讯科技有限公司 | Method and device for generating convolutional neural network and image recognition method and device |
| CN113487643A (en) * | 2021-07-19 | 2021-10-08 | 华电西藏能源有限公司大古水电分公司 | Cemented sand gravel stock yard sampling and determining method |
| CN113591740A (en) * | 2021-08-04 | 2021-11-02 | 清华大学 | Deep learning-based method and device for identifying silt particles in complex river environment |
| CN113591740B (en) * | 2021-08-04 | 2024-05-14 | 清华大学 | Deep learning-based sediment particle identification method and device in complex river environment |
| CN114324839A (en) * | 2022-01-07 | 2022-04-12 | 中电建路桥集团有限公司 | Low shallow roadbed construction integrated equipment that fills of high ground water level silt soil road |
| CN114324839B (en) * | 2022-01-07 | 2024-01-05 | 中电建路桥集团有限公司 | Low shallow roadbed construction integrated equipment that fills of high groundwater level silt land road |
| CN114858560A (en) * | 2022-06-06 | 2022-08-05 | 邢台路桥建设总公司 | Method for quickly measuring optimal water content of compacted soil of earthwork roadbed |
| CN114858560B (en) * | 2022-06-06 | 2024-05-24 | 邢台路桥建设集团有限公司 | Method for rapidly measuring optimal water content of compacted soil of earth subgrade |
| CN114863351A (en) * | 2022-07-07 | 2022-08-05 | 河北工业大学 | Picture and sound fusion roadbed filling collaborative identification management system based on Web3.0 |
| CN114863351B (en) * | 2022-07-07 | 2022-09-20 | 河北工业大学 | Picture and sound fusion roadbed filling collaborative identification management system based on Web3.0 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111523616B (en) | 2021-01-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111523616B (en) | Coarse-grained soil filler grading identification method based on convolutional neural network and application system | |
| CN111652270B (en) | Coarse-grained soil filler gradation automatic identification method based on image matching and application system | |
| CN110779937B (en) | Casting product internal defect intelligent detection device | |
| CN110458119B (en) | Non-contact measurement concrete aggregate gradation rapid identification method | |
| CN110569730B (en) | Road surface crack automatic identification method based on U-net neural network model | |
| CN110163213B (en) | Remote sensing image segmentation method based on disparity map and multi-scale depth network model | |
| CN111428625A (en) | Traffic scene target detection method and system based on deep learning | |
| CN110569755B (en) | Intelligent accumulated water detection method based on video | |
| CN113009447B (en) | Road underground cavity detection and early warning method based on deep learning and ground penetrating radar | |
| CN110969183A (en) | Method and system for determining damage degree of target object according to image data | |
| CN111754463B (en) | Method for detecting CA mortar layer defects of ballastless track based on convolutional neural network | |
| CN109978032A (en) | Bridge Crack detection method based on spatial pyramid cavity convolutional network | |
| JP5658613B2 (en) | Method and system for dividing particle size of granular material | |
| JP2021117625A (en) | Method for estimating particle size distribution of soil | |
| CN114022474A (en) | Particle grading rapid detection method based on YOLO-V4 | |
| CN115205255A (en) | Stone automatic grading method and system based on deep learning | |
| CN111830070A (en) | Automatic defect identification and judgment system and method based on edge calculation | |
| CN113129283B (en) | Digital image-based filler grading intelligent detection compaction control method | |
| CN112800934B (en) | Behavior recognition method and device for multi-class engineering vehicle | |
| CN111339950B (en) | Remote sensing image target detection method | |
| CN117237639A (en) | Pavement crack detection system based on YOLOv8 semantic segmentation | |
| CN115830302B (en) | Multi-scale feature extraction fusion power distribution network equipment positioning identification method | |
| CN111325724A (en) | Tunnel crack area detection method and device | |
| CN114897214A (en) | Metal additive manufacturing time prediction system and method based on graphical processing | |
| CN112766260B (en) | Image identification method and system for positioning air reservoir for accelerating and relieving railway train |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |