CN112489008A - Reinforcing steel bar identification system, reinforcing steel bar identification method and reinforcing steel bar counting and acceptance system - Google Patents
Reinforcing steel bar identification system, reinforcing steel bar identification method and reinforcing steel bar counting and acceptance system Download PDFInfo
- Publication number
- CN112489008A CN112489008A CN202011355672.6A CN202011355672A CN112489008A CN 112489008 A CN112489008 A CN 112489008A CN 202011355672 A CN202011355672 A CN 202011355672A CN 112489008 A CN112489008 A CN 112489008A
- Authority
- CN
- China
- Prior art keywords
- steel bar
- image
- rebar
- training
- identification
- 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.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
-
- 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
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30108—Industrial image inspection
- G06T2207/30136—Metal
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30242—Counting objects in image
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Probability & Statistics with Applications (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
Abstract
In order to improve and solve the defects and problems of the prior art that the traditional manual counting method is adopted to complete the inspection and acceptance of the entering quantity of the steel bars, the invention provides a steel bar identification system, a steel bar identification method and a steel bar counting and acceptance system. The technical scheme of the steel bar identification system is as follows: the method comprises the steps of acquiring a reinforcement image data set and a reinforcement identification training model; the rebar image dataset comprises a training image dataset and a verification image dataset; the steel bar image sample is stored in a training image data set or a verification image data set; the rebar identification training model receives a rebar image dataset.
Description
Technical Field
The invention belongs to the technical field of building construction, and particularly relates to a steel bar identification system, an identification method thereof and a steel bar counting and acceptance system.
Background
Reinforcing steel bars are one of the most commonly used building materials in building engineering, and have various types, different sizes and specifications and large using scale. At the present stage, in an actual engineering project, the steel bar entering quantity checking work adopts a mode of manually counting in batches, dozens of workers usually spend five to six hours to complete one steel bar entering quantity checking work, the work content is single and repeated, a large amount of manpower is consumed, and the work efficiency is low. In addition, counting errors often exist in manual counting of steel bars, and accuracy and reliability of checking and accepting quantity results cannot be guaranteed. Especially when the severe weather conditions such as high temperature, low temperature, typhoon, rainstorm, snowstorm and the like are faced, the working environment is poor, the work of manually counting the reinforcing steel bars is difficult to be normally carried out, and meanwhile, the counting error is amplified due to artificial psychological factors and emotional factors. In summary, a new intelligent checking and accepting system for the entering quantity of the steel bars based on computer vision and deep learning is needed to be provided, the time-consuming and labor-consuming repetitive work can be completed manually, and the defects and problems existing in the conventional manual counting method for completing the checking and accepting work of the entering quantity of the steel bars can be improved and solved.
Disclosure of Invention
In order to improve and solve the defects and problems of the prior art that the traditional manual counting method is adopted to complete the inspection and acceptance of the entering quantity of the steel bars, the invention provides a steel bar identification system, a steel bar identification method and a steel bar counting and acceptance system.
The technical scheme of the steel bar identification system is as follows:
the method comprises the steps of acquiring a reinforcement image data set and a reinforcement identification training model; the rebar image dataset comprises a training image dataset and a verification image dataset; the steel bar image sample is stored in a training image data set or a verification image data set; the rebar identification training model receives a rebar image dataset.
When the reinforcing steel bar identification system is used, firstly, reinforcing steel bars in reinforcing steel bar image samples are marked, and the reinforcing steel bar image samples are input into a reinforcing steel bar image data set; then, setting a training verification proportion, and randomly dividing the steel bar image samples into a training image data set or a verification image data set according to the training verification proportion; then, establishing a steel bar identification training model; importing the reinforcing steel bar image samples in the training image data set and the verification image data set into a reinforcing steel bar recognition training model, converting the reinforcing steel bar image samples in the training image data set into input tensor values, and converting the reinforcing steel bar image samples in the verification image data set into target tensor values; finally, setting model parameters such as a loss value threshold, an mAP value threshold, a maximum iteration number and the like of the steel bar recognition training model, and carrying out formal training; during the training process, respectively drawing a development curve of which the loss value and the mAP value change along with the increase of the iteration times in real time; when the development curve meeting the loss value and the mAP value is converged before the maximum iteration number is reached, the loss value is smaller than the threshold of the loss value, and the mAP value is larger than the threshold of the mAP value, stopping training, and releasing and deploying a trained reinforcing steel bar identification training model; and when the development curves of the loss value and the mAP value do not converge when the maximum iteration number is reached, or the loss value is larger than the threshold of the loss value or the mAP value is smaller than the threshold of the mAP value, returning to S2, adjusting model parameters such as the training verification proportion, the threshold of the loss value, the threshold of the mAP value, the maximum iteration number and the like, and retraining the reinforcing steel bar recognition training model.
According to the steel bar identification system, the steel bar is identified by means of fusion of computer vision and deep learning, and the accuracy of results is greatly improved. Compare traditional manual identification reinforcing bar, every bundle of reinforcing bar only need single cost several seconds alright accomplish, has improved the recognition rate of reinforcing bar greatly.
Further, in the steel bar identification system, the ratio of the steel bar image samples in the training image data set and the verification image data set is 6:4 to 9: 1. The proportion of the reinforcing steel bar image samples in the training image data set and the verification image data set is set to be 6: 4-9: 1, so that the robustness and the applicability of the reinforcing steel bar recognition training model can be effectively improved, and the reinforcing steel bar recognition training model is easy to converge.
Further, in the steel bar recognition system, the basic architecture of the steel bar recognition training model is based on the YOLOv4 image recognition method, and comprises a plurality of convolution layers and 3 YOLO layers.
The steel bar recognition training model is established based on the optimized YOLOv4 model, and has the following advantages:
(1) adopts a plurality of image data enhancement methods, increases the data capacity of training images,
(2) the training image database range covers the environment and scene of almost all on-site counting steel bars;
(3) the training image database resolution covers the full range of low resolution of 200 x 200 to high resolution of 5000 x 5000;
(4) the training model comprises a plurality of convolution layers and 3 YOLO layers, the number of the neural network layers is large, and the depth is wide;
(5) the training model detection box adopts a k-means clustering algorithm to set the size in advance.
Based on the characteristics, the steel bar recognition training model based on the optimized YOLOv4 model has the accuracy proved to be more than 99%, and the time for counting the steel bars is saved on the premise of ensuring certain accuracy.
The invention also provides a steel bar identification method, which adopts the steel bar identification system, and the technical scheme is as follows, comprising the following steps:
s1, collecting a steel bar image sample, marking steel bars in the steel bar image sample, and inputting the steel bar image sample into a steel bar image data set;
s2, setting a training verification proportion, and randomly dividing the steel bar image samples into a training image data set or a verification image data set according to the training verification proportion;
s3, establishing a steel bar recognition training model;
s4, importing the reinforcing steel bar image samples in the training image data set and the verification image data set into a reinforcing steel bar recognition training model, converting the reinforcing steel bar image samples in the training image data set into input tensor values, and converting the reinforcing steel bar image samples in the verification image data set into target tensor values;
s5, setting a loss value threshold, an mAP value threshold and a maximum iteration number of the steel bar recognition training model, and carrying out formal training;
s6, respectively drawing a development curve of which the loss value and the mAP value change along with the increase of the iteration times in real time; when the development curve meeting the loss value and the mAP value is converged before the maximum iteration number is reached, the loss value is smaller than the threshold of the loss value, and the mAP value is larger than the threshold of the mAP value, stopping training, and releasing and deploying a trained reinforcing steel bar identification training model; and when the development curves of the loss value and the mAP value do not converge when the maximum iteration number is reached, or the loss value is larger than the threshold of the loss value or the mAP value is smaller than the threshold of the mAP value, returning to S2, adjusting the training verification proportion, the threshold of the loss value, the threshold of the mAP value and the maximum iteration number, and retraining the reinforcing steel bar identification training model.
In the steel bar identification method, the loss value is a loss function value and represents the difference value between the predicted value and the actual true value of the training model, and the smaller the numerical value is, the closer the prediction result of the training model is to the actual value is represented; the mAP value is the average precision of the mean value and represents the precision of the training model for identifying the reinforcing steel bars, the range is 0-100%, and the larger the numerical value is, the higher the identification precision of the training model is.
According to the reinforcement identification method, aiming at the particularity of reinforcement identification work, namely the three characteristics of single type of an identified object, small identified object and large identified quantity in an image to be identified, a loss value, an mAP value and a development curve which changes along with the increase of iteration times are taken as judgment bases, a reinforcement identification training model is trained, the reinforcement identification training model can be optimized more effectively, and the training efficiency and the identification accuracy of the reinforcement identification training model are improved.
Further, in the steel bar identification method, in S1, the steel bar image sample is subjected to image adjustment to obtain a new steel bar image sample. In order to increase the accuracy of the recognition result of the training model, the data set capacity should be enlarged as much as possible. The steel bar image sample can be collected by taking a picture of a construction site or searching a network image database. And then, image adjustment can be carried out on the steel bar image samples in the modes of image rotation, brightness adjustment, contrast adjustment, definition adjustment, image random erasing, image random splicing and the like, so that more steel bar image samples are obtained.
Further, in the steel bar identification method, the steel bar identification training model further comprises a learning rate, wherein the learning rate is a gradient function adjusted along with the increase of the training iteration times; in S5, the method further comprises dynamically modifying the learning rate by first calculating a first moment estimate and a second moment estimate of the learning rate gradient function and dynamically adjusting the learning rate using the first moment estimate and the second moment estimate of the learning rate gradient function.
The reinforcement identification method adopts a dynamic learning rate control method based on the adam method, and is different from the learning rate modification strategy based on experience manual adjustment in the prior art. First, the first moment estimation and the second moment estimation of the learning rate gradient function are calculated, and the learning rate is dynamically adjusted by utilizing the first moment estimation and the second moment estimation of the learning rate gradient function.
Further, in the steel bar identification method, specifically, the steel bar identification training model further includes an impulse value, and the impulse value is a control function of a gradient descent speed; s5, further comprising modifying the impulse value.
Further, in the steel bar identification method, the steel bar identification training model further comprises an image resolution random adjustment function; s5, modifying the image resolution random adjustment function. The random adjustment function can adjust the resolution of the steel bar image sample within a certain range so as to increase the scene adaptability of the steel bar recognition training model. The adjustment scale of the resolution stochastic adjustment function is typically 0.5 to 1.5.
Further, in the steel bar identification method, specifically, the steel bar identification training model further includes a weight attenuation value, a detection frame size and a convolution kernel size; in S5, modifying the weight attenuation value, the size of the detection box, and the size of the convolution kernel.
The invention also provides a system for counting and checking the reinforcing steel bars, which adopts the technical scheme as follows:
the system comprises an image acquisition module, a server, the steel bar identification system and a steel bar counting module which are sequentially connected through signals.
When the reinforcing steel bar counting and accepting system is used, the image acquisition module sends the acquired target reinforcing steel bar image to the server; the steel bar identification system reads a target steel bar image in the server, marks steel bars in the target steel bar image, outputs the steel bars to the steel bar counting module, and counts the number of the steel bars in the target steel bar image through the steel bar counting module.
According to the reinforcing steel bar counting and accepting system, the image acquisition module can be a portable device such as a mobile phone, a worker can finish the counting of reinforcing steel bars only by taking pictures through the mobile phone, the field operation is convenient, the working process is greatly simplified, and the complex procedures that each reinforcing steel bar needs to be counted in sequence, the counted reinforcing steel bars are marked, and the counting result of the reinforcing steel bars needs to be manually recorded in the traditional manual counting method are avoided.
According to the steel bar counting and acceptance system, the working efficiency of steel bar counting is high, counting work of each bundle of steel bars can be completed only by one person within a few seconds, and compared with the traditional manual counting method, the efficiency of steel bar entering field number acceptance is greatly improved.
In addition, the steel bar counting and acceptance system provided by the invention adopts a means of combining computer vision and deep learning to carry out steel bar entrance quantity acceptance, so that the accuracy of results is greatly improved.
When the reinforcing steel bar counting and acceptance system is used, the shooting object distance and the shooting angle during the acquisition of the target reinforcing steel bar image are required to be paid attention to, theoretically, the shooting object distance is not too large or too small, the shooting angle is kept parallel to the length direction of the reinforcing steel bar (vertical to the cross section of the reinforcing steel bar), and the requirement that all reinforcing steel bar front views (reinforcing steel bar cross section views) which are in the acquired image and only need to be accepted are met as much as possible is met. Meanwhile, the number of steel bar bundles to be checked in one time is not too large, and the increase of the number of the bundles can cause the increase of the point error of the intelligent checking system. Tests show that the number of steel bar bundles accepted in one time is preferably a single bundle, and more than two bundles are not recommended.
Furthermore, the steel bar counting and acceptance system also comprises a result display module; and the result display module is in signal connection with the server. The reinforcing steel bar counting module counts the number of reinforcing steel bars in the target reinforcing steel bar image and then sends the number of the reinforcing steel bars to the result display module; checking the target steel bar image and the number of the steel bars on the result display module by a worker; and after the steel bars are checked to be correct, the result display module uploads the target steel bar image and the number of the steel bars to the server to finish steel bar counting and acceptance work.
Further, in the steel bar counting and acceptance system, the server comprises a basic operating system, a CPU/GPU processor, a 16G video memory card supporting GPU acceleration and a 10T storage hard disk; and the image acquisition module and the server as well as the server and the result display module transmit data through the 5G transmission module. The server can be set up in the high in the clouds, is responsible for receiving the target reinforcing bar image that uploads through 5G transmission module, sets up the operation platform that comes on the line of deployment for reinforcing bar identification system and reinforcing bar count module simultaneously to backup the target reinforcing bar image and the corresponding reinforcing bar result of counting uploaded at every turn, so that field management personnel backtrack, track each project, each stage reinforcing bar admission quantity inspection working conditions.
Further, in the steel bar counting and acceptance system, the steel bar counting module is built based on a math module in a Python language. The reinforcing bar count module is built based on the math module in the Python language, and the work of counting the reinforcing bar quantity in the reinforcing bar image that has mainly been responsible for accomplishing discernment, through receiving the reinforcing bar discernment result that reinforcing bar discernment module transmitted, the number of statistics reinforcing bar detection frame is as reinforcing bar quantity, and the work of counting the reinforcing bar quantity is accomplished to the corresponding numerical value of output to prepare to count the reinforcing bar result and pass to result display module through 5G transmission module.
Drawings
FIG. 1 is a schematic view of a rebar counting and acceptance system of the present invention;
fig. 2 is a schematic view of a rebar identification system of the present invention.
Detailed Description
The invention is described in further detail below with reference to the figures and specific examples. Advantages and features of the present invention will become apparent from the following description and from the claims. It is to be noted that the drawings are in a very simplified form and are not to precise scale, which is merely for the purpose of facilitating and distinctly claiming the embodiments of the present invention.
Example 1:
referring to fig. 2, a technical solution of a rebar identification system 1 of the present embodiment is as follows:
the method comprises a steel bar image data set 11 and a steel bar recognition training model 12; the rebar image dataset 11 comprises a training image dataset 111, and a verification image dataset 112; the steel bar image sample is stored in a training image data set 111 or a verification image data set 112; the rebar identification training model 12 receives a rebar image dataset 11.
When the reinforcing steel bar identification system 1 of the embodiment is used, firstly, reinforcing steel bars in reinforcing steel bar image samples are marked, and the reinforcing steel bar image samples are input into a reinforcing steel bar image data set 11; then, setting a training verification proportion, and randomly dividing the steel bar image samples into a training image data set 111 or a verification image data set 112 according to the training verification proportion; then, a steel bar recognition training model 12 is established; importing the rebar image samples in the training image data set 111 and the verification image data set 112 into the rebar recognition training model 12, converting the rebar image samples in the training image data set 111 into input tensor values, and converting the rebar image samples in the verification image data set 112 into target tensor values; finally, setting model parameters such as a loss value threshold, an mAP value threshold and the maximum iteration number of the steel bar recognition training model 12, and carrying out formal training; during the training process, respectively drawing a development curve of which the loss value and the mAP value change along with the increase of the iteration times in real time; when the development curve meeting the loss value and the mAP value is converged before the maximum iteration number is reached, the loss value is smaller than the threshold of the loss value, and the mAP value is larger than the threshold of the mAP value, stopping training, and releasing and deploying the trained reinforcing steel bar identification training model 12; when the development curves of the loss value and the mAP value do not converge when the maximum iteration number is reached, or the loss value is greater than the threshold of the loss value, or the mAP value is less than the threshold of the mAP value, returning to S2, adjusting model parameters such as the training verification proportion, the threshold of the loss value, the threshold of the mAP value, the maximum iteration number and the like, and retraining the reinforcing steel bar identification training model 12.
The reinforcing steel bar identification system 1 of the embodiment adopts a means of combining computer vision and deep learning to identify the reinforcing steel bars, so that the accuracy of results is greatly improved. Compare traditional manual identification reinforcing bar, every bundle of reinforcing bar only need single cost several seconds alright accomplish, has improved the recognition rate of reinforcing bar greatly.
In a preferred embodiment, in the rebar identification system, the ratio of the rebar image samples in the training image data set 111 and the verification image data set 112 is 6:4 to 9: 1. The proportion of the rebar image samples in the training image data set 111 and the verification image data set 112 is set to be 6: 4-9: 1, so that the robustness and the applicability of the rebar identification training model 12 can be effectively improved, and the rebar identification training model 12 is more prone to convergence.
In the steel bar recognition system, the basic architecture of the steel bar recognition training model 12 is based on the YOLOv4 image recognition method, and includes a plurality of convolution layers and 3 YOLO layers.
The steel bar recognition training model 12 is established based on the optimized YOLOv4 model, and has the following advantages:
(1) adopts a plurality of image data enhancement methods, increases the data capacity of training images,
(2) the training image database range covers the environment and scene of almost all on-site counting steel bars;
(3) the training image database resolution covers the full range of low resolution of 200 x 200 to high resolution of 5000 x 5000;
(4) the training model comprises a plurality of convolution layers and 3 YOLO layers, the number of the neural network layers is large, and the depth is wide;
(5) the training model detection box adopts a k-means clustering algorithm to set the size in advance.
Based on the characteristics, the steel bar recognition training model 12 based on the optimized YOLOv4 model has the accuracy proved to be more than 99%, and the time for counting the steel bars is saved on the premise of ensuring certain accuracy.
Example 2:
referring to fig. 2, the present embodiment provides a method for identifying a reinforcing bar, which adopts the reinforcing bar identification system 1 described in embodiment 1, and adopts the following technical scheme, including the following steps:
s1, collecting a steel bar image sample, marking steel bars in the steel bar image sample, and inputting the steel bar image sample into a steel bar image data set 11;
s2, setting a training verification proportion, and randomly dividing the steel bar image samples into a training image data set 111 or a verification image data set 112 according to the training verification proportion;
s3, establishing a steel bar recognition training model 12;
s4, importing the rebar image samples in the training image dataset 111 and the verification image dataset 112 into the rebar recognition training model 12, converting the rebar image samples in the training image dataset 111 into input tensor values, and converting the rebar image samples in the verification image dataset 112 into target tensor values;
s5, setting a loss value threshold, an mAP value threshold and the maximum iteration number of the steel bar recognition training model 12, and carrying out formal training;
s6, respectively drawing a development curve of which the loss value and the mAP value change along with the increase of the iteration times in real time; when the development curve meeting the loss value and the mAP value is converged before the maximum iteration number is reached, the loss value is smaller than the threshold of the loss value, and the mAP value is larger than the threshold of the mAP value, stopping training, and releasing and deploying the trained reinforcing steel bar identification training model 12; and when the development curves of the loss value and the mAP value do not converge when the maximum iteration number is reached, or the loss value is greater than the threshold of the loss value or the mAP value is less than the threshold of the mAP value, returning to S2, adjusting the training verification proportion, the threshold of the loss value, the threshold of the mAP value and the maximum iteration number, and retraining the reinforcing steel bar identification training model 12.
In the method for identifying the steel bars, the loss value is a loss function value and represents a difference value between a predicted value and an actual true value of the training model, and the smaller the numerical value is, the closer the prediction result of the training model is to the actual value is; the mAP value is the average precision of the mean value and represents the precision of the training model for identifying the reinforcing steel bars, the range is 0-100%, and the larger the numerical value is, the higher the identification precision of the training model is.
According to the reinforcement identification method, aiming at the particularity of reinforcement identification work, namely three characteristics of single type of an identified object, small identified object and large identification quantity in an image to be identified, a loss value, an mAP value and a development curve which changes along with the increase of iteration times are taken as judgment bases, the reinforcement identification training model 12 is trained, the reinforcement identification training model 12 can be optimized more effectively, and the training efficiency and the identification accuracy of the reinforcement identification training model 12 are improved.
In a preferred embodiment, in the method for identifying a rebar, in S1, the image of the rebar image sample is adjusted to obtain a new rebar image sample. In order to increase the accuracy of the recognition result of the training model, the data set capacity should be enlarged as much as possible. The steel bar image sample can be collected by taking a picture of a construction site or searching a network image database. And then, image adjustment can be carried out on the steel bar image samples in the modes of image rotation, brightness adjustment, contrast adjustment, definition adjustment, image random erasing, image random splicing and the like, so that more steel bar image samples are obtained.
In a preferred embodiment, in the steel bar identification method, the steel bar identification training model 12 further includes a learning rate, where the learning rate is a gradient function that is adjusted as the number of training iterations increases; in S5, the method further comprises dynamically modifying the learning rate by first calculating a first moment estimate and a second moment estimate of the learning rate gradient function and dynamically adjusting the learning rate using the first moment estimate and the second moment estimate of the learning rate gradient function.
The reinforcement identification method of the embodiment adopts a dynamic learning rate control method based on the adam method, and is different from the learning rate modification strategy based on experience manual adjustment in the past. First, the first moment estimation and the second moment estimation of the learning rate gradient function are calculated, and the learning rate is dynamically adjusted by utilizing the first moment estimation and the second moment estimation of the learning rate gradient function.
In a preferred embodiment, in the steel bar identification method, specifically, the steel bar identification training model 12 further includes an impulse value, and the impulse value is a control function of a descending speed of a gradient descent method; s5, further comprising modifying the impulse value.
In a preferred embodiment, in the steel bar identification method, the steel bar identification training model 12 further includes an image resolution random adjustment function; s5, modifying the image resolution random adjustment function. The random adjustment function can adjust the resolution of the rebar image samples within a certain range to increase the scene adaptability of the rebar recognition training model 12. The adjustment scale of the resolution stochastic adjustment function is typically 0.5 to 1.5.
In a preferred embodiment, in the steel bar identification method, specifically, the steel bar identification training model 12 further includes a weight attenuation value, a detection frame size, and a convolution kernel size; in S5, modifying the weight attenuation value, the size of the detection box, and the size of the convolution kernel.
Example 3:
referring to fig. 1, the present embodiment provides a steel bar counting and acceptance system, and the technical scheme is as follows:
including image acquisition module 2, server 3, embodiment 1 of signal connection in proper order reinforcing bar identification system 1 and reinforcing bar count module 4.
In the steel bar counting and acceptance system of the embodiment, when the system is used, the image acquisition module 2 sends the acquired target steel bar image to the server 3; the steel bar identification system 1 reads a target steel bar image in the server 3, marks steel bars in the target steel bar image, outputs the steel bars to the steel bar counting module, and counts the number of the steel bars in the target steel bar image by the steel bar counting module 4.
The reinforcing bar counting and acceptance system of the embodiment is characterized in that the image acquisition module 2 can be portable equipment such as a mobile phone, a worker can complete the counting of reinforcing bars only by taking pictures through the mobile phone, the field operation is convenient, the work flow is greatly simplified, and the complex process that the traditional manual counting method needs to count each reinforcing bar in sequence, mark the counted reinforcing bars and manually record the counting result of the reinforcing bars is avoided.
The reinforcing bar count system of accepting of this embodiment, the work efficiency of reinforcing bar count is higher, and the work of counting only needs single cost several seconds alright completion for every bundle of reinforcing bar, compares with the artifical method of counting of tradition, has improved the efficiency that the reinforcing bar entered the field quantity was accepted greatly.
In addition, the steel bar counting acceptance system of the embodiment adopts a means of combining computer vision and deep learning to carry out steel bar entrance quantity acceptance, and greatly improves the accuracy of results.
The reinforcing bar count acceptance system of this embodiment, during the use, should pay attention to shooting object distance and shooting angle when gathering target reinforcing bar image, should not be too big or the undersize in the theoretical shooting object distance, should shoot the angle and should keep being parallel (perpendicular to reinforcing bar cross section) with reinforcing bar length direction, should satisfy as far as possible and gather all reinforcing bar elevation pictures (reinforcing bar cross section picture) that have in the image and only remain acceptance. Meanwhile, the number of steel bar bundles to be checked in one time is not too large, and the increase of the number of the bundles can cause the increase of the point error of the intelligent checking system. Tests show that the number of steel bar bundles accepted in one time is preferably a single bundle, and more than two bundles are not recommended.
In a preferred embodiment, the rebar counting and acceptance system further comprises a result display module 5; the result display module 5 is in signal connection with the server 3. The steel bar counting module 4 counts the number of steel bars in the target steel bar image and then sends the number of the steel bars to the result display module 5; checking the target steel bar image and the number of the steel bars on the result display module 5 by a worker; after the error is checked, the result display module 5 uploads the target steel bar image and the number of the steel bars to the server 3 to complete the steel bar counting and acceptance work.
In a preferred embodiment, in the rebar counting and acceptance system, the server 3 includes a basic operating system, a CPU/GPU processor, a 16G video memory card supporting GPU acceleration, and a 10T storage hard disk; and data are transmitted between the image acquisition module 2 and the server 3 and between the server 3 and the result display module 5 through a 5G transmission module. The server 3 can be arranged at the cloud end and is responsible for receiving the target reinforcing steel bar image uploaded by the 5G transmission module, meanwhile, a deployment online operation platform is set up for the reinforcing steel bar identification system 1 and the reinforcing steel bar counting module 4, and the target reinforcing steel bar image uploaded at each time and a corresponding reinforcing steel bar counting result are backed up, so that field management personnel can trace back and track the working conditions of checking the entering quantity of reinforcing steel bars in each project and stage.
As a preferred embodiment, in the steel bar counting and acceptance system, the steel bar counting module 4 is built based on a math module in Python language. Reinforcing bar count module 4 builds based on the math module in the Python language, and the work of counting of the reinforcing bar quantity in the reinforcing bar image that mainly is responsible for accomplishing the discernment, through receiving the reinforcing bar identification result that reinforcing bar identification module transmitted, the number of statistics reinforcing bar detection frame is as reinforcing bar quantity, and the work of counting of output corresponding numerical value accomplishes reinforcing bar quantity to prepare to pass to the result display module with reinforcing bar counting result through 5G transmission module.
The above description is only for the purpose of describing the preferred embodiments of the present invention, and is not intended to limit the scope of the present invention, and any variations and modifications made by those skilled in the art based on the above disclosure are within the scope of the appended claims.
Claims (13)
1. A rebar identification system (1) comprising a rebar image dataset (11), and a rebar identification training model (12); the rebar image dataset (11) comprises a training image dataset (111), and a verification image dataset (112); the rebar image sample is stored in a training image data set (111) or a verification image data set (112); the rebar identification training model (12) receives a rebar image dataset (11).
2. A rebar identification system (1) as claimed in claim 1, wherein the scale of rebar image samples in the training image dataset (111), validation image dataset (112) is 6:4 to 9: 1.
3. The reinforcement recognition system (1) of claim 1, wherein the basic architecture of the reinforcement recognition training model (12) is based on a YOLOv4 image recognition method, comprising a plurality of convolution layers and 3 YOLO layers.
4. A reinforcement identification method, characterized in that, using the reinforcement identification system (1) of claim 1, comprising the steps of:
s1, collecting a steel bar image sample, marking steel bars in the steel bar image sample, and inputting the steel bar image sample into a steel bar image data set (11);
s2, setting a training verification proportion, and randomly dividing the steel bar image samples into a training image data set (111) or a verification image data set (112) according to the training verification proportion;
s3, establishing a steel bar recognition training model (12);
s4, leading the rebar image samples in the training image data set (111) and the verification image data set (112) into a rebar recognition training model (12), converting the rebar image samples in the training image data set (111) into input tensor values, and converting the rebar image samples in the verification image data set (112) into target tensor values;
s5, setting a loss value threshold, an mAP value threshold and the maximum iteration number of the steel bar recognition training model (12), and carrying out formal training;
s6, respectively drawing a development curve of which the loss value and the mAP value change along with the increase of the iteration times in real time; when the development curve meeting the loss value and the mAP value is converged before the maximum iteration number is reached, the loss value is smaller than a threshold value of the loss value and the mAP value is larger than a threshold value of the mAP value, stopping training, and releasing and deploying a trained reinforcing steel bar identification training model (12); and when the development curves of the loss value and the mAP value do not converge when the maximum iteration number is reached, or the loss value is larger than the threshold of the loss value or the mAP value is smaller than the threshold of the mAP value, returning to S2, adjusting the training verification proportion, the threshold of the loss value, the threshold of the mAP value and the maximum iteration number, and retraining the reinforcing steel bar identification training model (12).
5. The reinforcing bar identification method as claimed in claim 4, wherein in S1, the reinforcing bar image sample is image-adjusted to obtain a new reinforcing bar image sample.
6. The reinforcement identification method according to claim 4, wherein the reinforcement identification training model (12) further comprises a learning rate which is a gradient function that is adjusted as the number of training iterations increases;
in S5, the method further comprises dynamically modifying the learning rate by first calculating a first moment estimate and a second moment estimate of the learning rate gradient function and dynamically adjusting the learning rate using the first moment estimate and the second moment estimate of the learning rate gradient function.
7. The reinforcement identification method according to claim 4, wherein the reinforcement identification training model (12) further comprises an impulse value which is a control function of a gradient descent speed;
s5, further comprising modifying the impulse value.
8. The reinforcement identification method according to claim 4, wherein the reinforcement identification training model (12) further comprises an image resolution stochastic adjustment function;
s5, modifying the image resolution random adjustment function.
9. The reinforcement identification method according to claim 4, wherein the reinforcement identification training model (12) further comprises weight attenuation values, detection box sizes, convolution kernel sizes;
in S5, modifying the weight attenuation value, the size of the detection box, and the size of the convolution kernel.
10. A rebar counting and acceptance system comprising an image acquisition module (2), a server (3), a rebar identification system (1) according to claim 1, and a rebar counting module (4) in signal connection in sequence.
11. A rebar count acceptance system according to claim 10 further comprising a result display module (5); and the result display module (5) is in signal connection with the server (3).
12. The rebar count acceptance system of claim 13, wherein the server (3) comprises a base operating system, a CPU/GPU processor, a 16G video memory card supporting GPU acceleration, and a 10T storage hard disk; and data are transmitted between the image acquisition module (2) and the server (3) and between the server (3) and the result display module (5) through a 5G transmission module.
13. A rebar count acceptance system according to claim 10, wherein the rebar count module (4) is built based on a math module in Python language.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011355672.6A CN112489008A (en) | 2020-11-27 | 2020-11-27 | Reinforcing steel bar identification system, reinforcing steel bar identification method and reinforcing steel bar counting and acceptance system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011355672.6A CN112489008A (en) | 2020-11-27 | 2020-11-27 | Reinforcing steel bar identification system, reinforcing steel bar identification method and reinforcing steel bar counting and acceptance system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112489008A true CN112489008A (en) | 2021-03-12 |
Family
ID=74935877
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011355672.6A Pending CN112489008A (en) | 2020-11-27 | 2020-11-27 | Reinforcing steel bar identification system, reinforcing steel bar identification method and reinforcing steel bar counting and acceptance system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112489008A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113762287A (en) * | 2021-09-17 | 2021-12-07 | 湖北工业大学 | Intelligent detection system and method for underwater target and storage medium |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20190068264A (en) * | 2017-12-08 | 2019-06-18 | 서한석 | Construction management system and method of installing reinforced bars |
| CN110032954A (en) * | 2019-03-27 | 2019-07-19 | 成都数之联科技有限公司 | A kind of reinforcing bar intelligent recognition and method of counting and system |
| CN110264466A (en) * | 2019-06-28 | 2019-09-20 | 广州市颐创信息科技有限公司 | A kind of reinforcing bar detection method based on depth convolutional neural networks |
| CN110706210A (en) * | 2019-09-18 | 2020-01-17 | 五邑大学 | Deep learning-based rebar counting method and device |
| CN110781755A (en) * | 2019-09-28 | 2020-02-11 | 深圳中物智建科技有限公司 | Intelligent identification and metering method and system for construction steel bars |
| CN111127457A (en) * | 2019-12-25 | 2020-05-08 | 上海找钢网信息科技股份有限公司 | Reinforcing steel bar number statistical model training method, statistical method, device and equipment |
| CN111507291A (en) * | 2020-04-22 | 2020-08-07 | 南京煜耀智能科技有限公司 | Reinforcing steel bar point identification software system based on AI machine learning |
| CN111639740A (en) * | 2020-05-09 | 2020-09-08 | 武汉工程大学 | Steel bar counting method based on multi-scale convolution neural network |
-
2020
- 2020-11-27 CN CN202011355672.6A patent/CN112489008A/en active Pending
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20190068264A (en) * | 2017-12-08 | 2019-06-18 | 서한석 | Construction management system and method of installing reinforced bars |
| CN110032954A (en) * | 2019-03-27 | 2019-07-19 | 成都数之联科技有限公司 | A kind of reinforcing bar intelligent recognition and method of counting and system |
| CN110264466A (en) * | 2019-06-28 | 2019-09-20 | 广州市颐创信息科技有限公司 | A kind of reinforcing bar detection method based on depth convolutional neural networks |
| CN110706210A (en) * | 2019-09-18 | 2020-01-17 | 五邑大学 | Deep learning-based rebar counting method and device |
| CN110781755A (en) * | 2019-09-28 | 2020-02-11 | 深圳中物智建科技有限公司 | Intelligent identification and metering method and system for construction steel bars |
| CN111127457A (en) * | 2019-12-25 | 2020-05-08 | 上海找钢网信息科技股份有限公司 | Reinforcing steel bar number statistical model training method, statistical method, device and equipment |
| CN111507291A (en) * | 2020-04-22 | 2020-08-07 | 南京煜耀智能科技有限公司 | Reinforcing steel bar point identification software system based on AI machine learning |
| CN111639740A (en) * | 2020-05-09 | 2020-09-08 | 武汉工程大学 | Steel bar counting method based on multi-scale convolution neural network |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113762287A (en) * | 2021-09-17 | 2021-12-07 | 湖北工业大学 | Intelligent detection system and method for underwater target and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112801230B (en) | Intelligent acceptance method for unmanned aerial vehicle of power distribution line | |
| CN113407651B (en) | Progress information feedback and deviation real-time display method | |
| CN110910440B (en) | Power transmission line length determination method and system based on power image data | |
| CN108848337A (en) | Long-range auditing method, device, terminal and the computer storage medium of line project | |
| CN116993665A (en) | Intelligent detection method for construction progress of construction engineering working face based on computer vision | |
| CN112489008A (en) | Reinforcing steel bar identification system, reinforcing steel bar identification method and reinforcing steel bar counting and acceptance system | |
| CN108072896B (en) | A kind of fully automatically seismic wave first break pickup method and system | |
| CN115035297A (en) | Automatic recording method, system, device and medium for drilling core RQD | |
| CN117114619B (en) | Project security management system based on big data analysis | |
| CN112004063B (en) | Method for monitoring connection correctness of oil discharge pipe based on multi-camera linkage | |
| CN116363088A (en) | Expressway concrete pavement crack detection method based on unmanned aerial vehicle, electronic equipment and storage medium | |
| CN111190975A (en) | River channel patrol method for calculating task track coverage rate according to GPS data | |
| CN114493291B (en) | Intelligent detection method and system for high fill quality | |
| CN108537828A (en) | A kind of shop data analysing method and system | |
| CN115775092A (en) | Construction process safety risk management and control system based on digital twin technology | |
| CN115241980A (en) | System and method for checking power supply radius of distribution network area based on unmanned aerial vehicle front end identification | |
| CN113810658A (en) | Construction management system and method for climbing project | |
| CN112627538A (en) | Intelligent acceptance method for binding quality of steel mesh binding wires based on computer vision | |
| CN111507432A (en) | Intelligent weighing method and system for agricultural insurance claims, electronic equipment and storage medium | |
| CN117670256B (en) | BIM technology-based engineering cost accurate control system and method | |
| CN116662318B (en) | Method for rapidly extracting house information data based on inclined three-dimensional model | |
| KR102635926B1 (en) | Facility Crack Measuring System | |
| CN117036611B (en) | Three-dimensional scene construction method and system of non-mine safety monitoring platform | |
| CN116862307B (en) | Building engineering service assisting system based on artificial intelligence | |
| CN115880644B (en) | Method and system for identifying coal quantity based on artificial intelligence |
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 |