CN114782669A - Digital instrument automatic identification, positioning and reading method based on deep learning - Google Patents

Digital instrument automatic identification, positioning and reading method based on deep learning Download PDF

Info

Publication number
CN114782669A
CN114782669A CN202210018923.4A CN202210018923A CN114782669A CN 114782669 A CN114782669 A CN 114782669A CN 202210018923 A CN202210018923 A CN 202210018923A CN 114782669 A CN114782669 A CN 114782669A
Authority
CN
China
Prior art keywords
training
model
network
digital
instrument
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
Application number
CN202210018923.4A
Other languages
Chinese (zh)
Inventor
杨延西
匡益
刘磊
韩乐
孙俏
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xian University of Technology
Original Assignee
Xian University of Technology
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Xian University of Technology filed Critical Xian University of Technology
Priority to CN202210018923.4A priority Critical patent/CN114782669A/en
Publication of CN114782669A publication Critical patent/CN114782669A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/04—Inference or reasoning models
    • G06N5/046—Forward inferencing; Production systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Computing Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Software Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

The invention discloses a method for realizing the identification, positioning and reading of a digital instrument based on a deep learning method, which comprises the steps of firstly shooting an image with the digital instrument, and collecting 1000 images to form a data set; labeling the data set by using Labelimg, and then dividing the data set into a training set, a testing set and a verification set according to the proportion of 6:2: 2; and (3) sending the marked image into a trained and tuned YOLOv3 network model, then intercepting the digital instrument image and inputting the digital instrument image into an improved CenterNet network to obtain a digital display area coordinate and an inclination angle, then carrying out inclination correction on the image according to the inclination angle and intercepting the digital display area image, and finally directly inputting the digital display area image into a CRNN network to obtain a final reading recognition result. Compared with the traditional algorithm, the method reduces the influence of the conditions such as instrument inclination, environmental change and the like on reading identification, and has the advantages of wider application range and better robustness.

Description

Digital instrument automatic identification, positioning and reading method based on deep learning
Technical Field
The invention belongs to the technical field of automatic reading of digital instruments, and particularly relates to a deep learning-based automatic digital instrument identification, positioning and reading method.
Background
The industrial instrument can be divided into a digital instrument and a pointer instrument according to a display mode, the digital instrument mainly displays the indicating number through various carriers such as liquid crystal, a nixie tube and the like, and the digital instrument has the advantages of simplicity in observation, convenience in maintenance and the like, and is more and more applied in industrial production. At present, with the development of computer technology, an automatic meter reading recognition algorithm based on an image recognition algorithm is gradually developed, for example, in chinese patent CN109255336A, "inspection robot-based lightning arrester recognition method", a bidirectional projection method is adopted in combination with an artificial neural network to perform digital segmentation and recognition. However, these conventional image processing methods have high requirements for captured images, and if the images have problems such as large illumination change, complex background, various instrument types, and instrument inclination, the recognition accuracy tends to be low in the subsequent processing.
Disclosure of Invention
The invention aims to provide a digital instrument automatic identification, positioning and reading method based on deep learning, which solves the influence of the problems of instrument inclination, environment change, various instruments and the like on digital instrument reading identification.
The technical scheme adopted by the invention is that the digital instrument automatic identification, positioning and reading method based on deep learning is implemented according to the following steps:
step 1, shooting an image with a digital instrument;
step 2, collecting 1000 images of the digital instrument by using a 2-type digital display thermometer and a data set expansion means, and manufacturing an instrument data set;
step 3, labeling the data set in the step 2 by using Labelimg labeling software, and then dividing the data set into a training set, a testing set and a verification set according to the ratio of 6:2: 2;
step 4, training and adjusting the YOLOv3 network model, sending the image labeled in the step 3 into the YOLOv3 network model, and outputting the image as the position information and the classification information of the digital instrument in the image through the network;
step 5, training and adjusting the CenterNet network model, inputting the position information of the digital instrument output in the step 4 into the improved CenterNet network model, and outputting the position information and the inclination angle of the display area of the digital instrument through the network;
step 6, carrying out inclination transformation on the display area according to the inclination angle of the digital instrument output in the step 5 to obtain a corrected display area image;
and 7, training and adjusting the CRNN model, and inputting the display area image corrected in the step 6 into the CRNN model to obtain the final digital instrument reading.
The present invention is also characterized in that,
the 2 types of digital display thermometers in the step 2 are a 0-200 ℃ digital display thermometer and a 0-500 ℃ digital display thermometer, and the data set expansion means is contrast and brightness random adjustment and channel random exchange.
In the step 4, training and tuning are performed by firstly using a K-means algorithm to perform cluster analysis on the instrument data set to obtain the anchors size of the best matched instrument data set;
during training, setting the network input size to be 618 multiplied by 3, the iteration number to be 50200, the initial learning rate to be 0.005, the learning rate to be attenuated by 10 times every 10000 times, setting the optimization algorithm to be a Mini-BGD algorithm, setting the momentum value to be 0.9, the attenuation coefficient to be 0.0005, setting the Batchsize to be 64, and finishing forward propagation in 8 times;
and by observing a loss value change curve of the verification set, when the loss value change curve shows a descending trend and tends to be stable after a certain number of iterations, the model is regarded as convergent, and if the model does not converge, the model is subjected to retraining by increasing or reducing the learning rate according to a multiple of 10.
In the step 5, training and optimization are performed by setting the number of iteration rounds to be 150 and the initial learning rate to be 0.000125, respectively reducing the learning rate by 10 times in the 90 th round and the 120 th round, setting the blocksize to be 8, setting the optimization algorithm to be Adam algorithm, observing a loss value change curve of the verification set, considering that the model converges when the model tends to be stable after presenting a descending trend and iterating for a certain number of times, and performing retraining according to the number of times of 10 if the model does not converge.
Step 5, the improved CenterNet network model is that on the basis of the original CenterNet network structure, the input size of a modified network is 256 multiplied by 3, a main extraction network selects ResNet-50, the last layer of downsampling operation in the ResNet-50 is replaced by hollow convolution, and an angle regression branch is added to the head part;
the multilayer ceramic material specifically comprises two convolution layers and a ReLU active layer, wherein each layer comprises the following components: the convolution kernel size of the first layer of convolution layer is 64 multiplied by 3, the characteristic graph output is 64 multiplied by 64, the second layer is a ReLU activation layer, the third layer is a convolution layer, namely an output layer, the convolution kernel size is 1 multiplied by 1, and the characteristic graph output is 1 multiplied by 64; because an angle regression branch is added, the corresponding loss function to be added into the rotation angle regression adopts the minimum absolute value as the loss function, namely the L1 loss function, and the specific formula of the improved CenterNet loss function is as follows:
Figure BDA0003459865550000041
wherein L isCenterNetAs a loss function of the CenterNet network, λAngleIs an angle backThe weight lost to the branch, taken 0.5 out of the network,
Figure BDA0003459865550000042
as angle prediction value, ynIs the true rotation angle value.
Step 6 oblique transformation to center point (x)0，y0) As the origin, the rotation matrix is expressed as
Figure BDA0003459865550000043
Where θ is the tilt angle output in step 5.
In the step 7, training and optimizing are carried out by firstly crawling images which take the digital display instrument and the digital display thermometer as key words in the network picture through a network crawler, screening and cutting the images to obtain images only containing digital parts, and then cutting the images collected in the experiment to obtain a data set;
the training is divided into two stages, the first stage only uses the processed network pictures for training, and the training parameters are set as follows: the image input size is 32 multiplied by 100 multiplied by 3, the batch size is 128, the number of iteration rounds is set to 72 rounds, the initial learning rate is 0.005, and the optimization algorithm is the Adam algorithm;
in the second stage, the data set is divided into a training set, a verification set and a test set according to the ratio of 8:1, the optimal model obtained in the first stage is used as a pre-training model, training is carried out on the model by using the training set, and the training parameters in the second stage are set as follows: the image input size is 32 × 100 × 3, the batchsize is 128, the iteration number is set to 144 rounds, the initial learning rate is 0.0005, the learning rate attenuation coefficient is 0.00004, the optimization algorithm is Adam algorithm, the first moment estimation and second moment estimation attenuation rates are 0.9 and 0.999, and L2 regularization is set;
by observing a loss value change curve of the verification set, when a descending trend appears and tends to be stable after a certain number of iterations, the model is regarded as convergent, and if the model does not converge, the model is retrained by increasing or reducing the learning rate according to a multiple of 10.
The beneficial effect of the invention is that,
(1) the invention realizes the identification, positioning and reading of the digital instrument by using a deep learning-based method, and introduces an angle prediction branch on the basis of a CenterNet network so that the model can directly carry out regression of the inclination angle without independently using Hough linear detection and other methods to calculate the inclination angle.
(2) Compared with the traditional algorithm, the method reduces the influence of the conditions such as instrument inclination, environmental change and the like on the reading number, and has wider application range and better robustness.
Drawings
FIG. 1 is a schematic diagram of an experimental scenario of the deep learning-based digital instrument automatic identification, positioning and reading method of the present invention;
FIG. 2 is a flow chart of the deep learning based digital meter auto-identification, location and reading method of the present invention;
FIG. 3 is a flow chart of coarse positioning and sorting of the digital instrument according to the method for automatic identification, positioning and reading of the digital instrument based on deep learning of the present invention;
FIG. 4 is a diagram of the results of coarse positioning and classification of a digital meter according to the deep learning-based method for automatically identifying, positioning and reading the digital meter;
FIG. 5 is a flow chart of the digital display area location and tilt angle detection of the deep learning-based method for automatically identifying, locating and reading a digital instrument according to the present invention;
FIG. 6 is a diagram of the result of positioning the digital display area of the deep learning-based method for automatically identifying, positioning and reading the digital instrument according to the present invention;
FIG. 7 is a flow chart of the digital display area recognition method of the present invention for deep learning based digital instrument automatic recognition, location and reading;
FIG. 8 is a diagram of the recognition result of the digital meter based on the deep learning automatic digital meter recognition, positioning and reading method of the present invention.
In the figure, 1 is a camera, 2 is an inspection robot, 3 is a digital instrument, 4 is a wall surface, and 5 is a server.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
According to the experiment scene schematic diagram of the digital instrument automatic identification, positioning and reading method based on deep learning, as shown in fig. 1, a camera 1 is connected with an inspection robot 2 and returns a shot image stream in real time, the inspection robot 2 can drive according to a pre-planned path, after the inspection robot reaches a specified position, the camera 1 is adjusted, an original image containing an instrument 3 on a wall surface 4 is shot, then the image is transmitted to a server 5, and the server 5 performs automatic reading identification on a liquid crystal digital instrument.
The invention relates to a flow chart of a deep learning-based digital instrument automatic identification, positioning and reading method, which is specifically implemented according to the following steps as shown in FIG. 2:
step 1, the inspection robot 2 moves to a designated position, adjusts the camera 1, and shoots an image with the digital instrument 3.
And 2, as shown in the flow chart of fig. 3, acquiring 1000 images of the digital instrument by using a 0-200 ℃ digital display thermometer and a 0-500 ℃ digital display thermometer and by using a data set expansion means of random adjustment of contrast and brightness and random channel exchange to manufacture an instrument data set.
And 3, labeling the data set in the step 2 by using Labelimg labeling software, and then dividing the data set into a training set, a test set and a verification set according to the ratio of 6:2: 2.
Step 4, training and adjusting the YOLOv3 network model, firstly, carrying out cluster analysis on the instrument data set by using a K-means algorithm to obtain the anchors size of the best matched instrument data set;
during training, setting the network input size to be 618 multiplied by 3, the iteration number to be 50200, the initial learning rate to be 0.005, the learning rate of each 10000 times is attenuated by 10 times, setting the optimization algorithm to be a Mini-BGD algorithm, setting the momentum value to be 0.9, the attenuation coefficient to be 0.0005, setting the batchsize to be 64, and finishing forward propagation in 8 times;
by observing a loss value change curve of the verification set, when the loss value change curve shows a descending trend and tends to be stable after a certain number of iterations, the model is regarded as convergent, if the model does not converge, retraining is performed according to a multiple of 10 to increase or reduce the learning rate, after training is completed, a test set mAP value and an average forward reasoning time are calculated, and the higher the mAP value is and the shorter the average forward reasoning time is, the better the model performance is;
and then, the image marked in the step 3 is sent into a Yolov3 network model, the network output is the position information and the classification information of the digital instrument in the image, as shown in figure 4, and finally, the instrument image is intercepted according to the position information.
Step 5, as shown in the flow chart of fig. 5, training and tuning the centret network model, setting the number of iteration rounds as 150 and the initial learning rate as 0.000125 for the data set divided in step 3, respectively reducing the learning rate by 10 times in the 90 th round and the 120 th round, and the batch size as 8, setting the optimization algorithm as an Adam algorithm, observing the loss value change curve of the verification set, and when the loss value change curve shows a descending trend and tends to be stable after a certain number of iterations, considering the model to be converged, and if the loss value change curve does not converge, performing retraining according to the multiple of 10 or reducing the learning rate;
inputting the position information of the digital instrument output in the step 4 into an improved CenterNet network model, namely modifying the network input size to be 256 multiplied by 3 on the basis of an original CenterNet network structure, selecting ResNet-50 by a trunk extraction network, replacing the last layer of down-sampling operation in the ResNet-50 with cavity convolution, and adding an angle regression branch in the head part;
the multilayer ceramic material specifically comprises two convolution layers and a ReLU active layer, wherein each layer comprises the following components: the convolution kernel size of the convolution layer of the first layer is 64 multiplied by 3, the characteristic diagram output is 64 multiplied by 64, the second layer is a ReLU layer, the third layer is a convolution layer, namely an output layer, the convolution kernel size is 1 multiplied by 1, and the characteristic diagram output is 1 multiplied by 64; because an angle regression branch is added, the corresponding loss function to be added into the rotation angle regression adopts the minimum absolute value as the loss function, namely the L1 loss function, and the specific formula of the improved CenterNet loss function is as follows:
Figure BDA0003459865550000081
wherein L isCenterNetAs a loss function of the CenterNet network, λAngleThe weight lost by the angle regression branch is taken as 0.5 in the network,
Figure BDA0003459865550000082
as angle prediction value, ynIs the true rotation angle value;
the network output is the position information and the inclination angle of the display area of the digital instrument, as shown in fig. 6.
Step 6, performing inclination transformation on the display area according to the inclination angle predicted in the step 5, wherein the central point (x) is used for the inclination transformation0，y0) As the origin, the rotation matrix is expressed as
Figure BDA0003459865550000083
Wherein theta is the inclination angle predicted in the step 5;
thereby obtaining a corrected display area image.
Step 7, as shown in fig. 7, the process trains and adjusts the CRNN network model, crawls images using a digital display instrument and a digital display thermometer as keywords in a network picture through a network crawler, screens and cuts the images to obtain images only containing digital parts, and cuts the images collected in the experiment to obtain a data set;
the training is divided into two stages, the first stage only uses the processed network pictures for training, and the training parameters are set as follows: the image input size is 32 multiplied by 100 multiplied by 3, the batch size is 128, the number of iteration rounds is set to 72 rounds, the initial learning rate is 0.005, and the optimization algorithm is Adam algorithm;
in the second stage, an experimental acquisition data set is divided into a training set, a verification set and a test set according to the ratio of 8:1:1, then the optimal model obtained after the training in the first stage is used as a pre-training model, the training set is used for fine adjustment on the model, and the training parameters in the second stage are set as follows: the image input size is 32 multiplied by 100 multiplied by 3, the batchsize is 128, the iteration number is set to 144 rounds, the initial learning rate is 0.0005, the learning rate attenuation coefficient is 0.00004, the optimization algorithm is Adam algorithm, the first moment estimation and second moment estimation attenuation rates are 0.9 and 0.999, and L2 regularization is set;
by observing a loss value change curve of the verification set, when a descending trend appears and tends to be stable after a certain number of iterations, the model is regarded as convergent, and if the model does not converge, the model is retrained again by increasing or reducing the learning rate according to a multiple of 10; and after training is finished, calculating the accuracy of the test set and the average forward reasoning time, wherein the higher the accuracy is and the shorter the average forward reasoning time is, the better the model performance is.
Then, the display area image corrected in step 6 is input into the CRNN network model to obtain the final digital meter reading, and the final result is shown in fig. 8.
According to the method, the angle prediction branch is introduced on the basis of the CenterNet network, so that the model can directly carry out regression on the inclination angle, and Hough linear detection and other methods are not required to be independently used for calculating the inclination angle.

Claims (7)

1. The digital instrument automatic identification, positioning and reading method based on deep learning is characterized by comprising the following steps:
step 1, shooting an image with a digital instrument (3);
step 2, collecting 1000 images of the digital instrument by using a 2-type digital display thermometer and a data set expansion means, and manufacturing an instrument data set;
step 3, labeling the data set in the step 2 by using Labelimg labeling software, and then dividing the data set into a training set, a testing set and a verification set according to the ratio of 6:2: 2;
step 4, training and adjusting the YOLOv3 network model, sending the image marked in the step 3 into the YOLOv3 network model, and outputting the position information and the classification information of the digital instrument in the image through a network;
step 5, training and adjusting the CenterNet network model, inputting the position information of the digital instrument output in the step 4 into the improved CenterNet network model, and outputting the position information and the inclination angle of the display area of the digital instrument through the network;
step 6, carrying out inclination transformation on the display area according to the inclination angle of the digital instrument output in the step 5 to obtain a corrected display area image;
and 7, training and adjusting the CRNN model, and inputting the display area image corrected in the step 6 into the CRNN model to obtain the final digital instrument reading.
2. The deep learning-based digital instrument automatic identification, positioning and reading method as claimed in claim 1, wherein the class 2 digital display temperature meter of the step 2 is a digital display temperature meter of 0-200 ℃ and a digital display temperature meter of 0-500 ℃, and the data set expansion means is random adjustment of contrast and brightness and random exchange of channels.
3. The deep learning-based digital instrument automatic identification, positioning and reading method according to claim 1, wherein the training and tuning in step 4 is to first perform cluster analysis on the instrument data set using a K-means algorithm to obtain the anchors size of the best-matching instrument data set;
during training, setting the network input size to be 618 multiplied by 3, the iteration number to be 50200, the initial learning rate to be 0.005, the learning rate to be attenuated by 10 times every 10000 times, setting the optimization algorithm to be a Mini-BGD algorithm, setting the momentum value to be 0.9, setting the attenuation coefficient to be 0.0005, setting the Batchsize to be 64, and finishing forward propagation in 8 times;
by observing a loss value change curve of the verification set, when a descending trend appears and tends to be stable after a certain number of iterations, the model is regarded as convergent, and if the model does not converge, the model is retrained by increasing or reducing the learning rate according to a multiple of 10.
4. The deep learning-based digital instrument automatic identification, positioning and reading method according to claim 1, wherein the training and optimization in step 5 is to divide the data set in step 3, set the number of iteration rounds to be 150, the initial learning rate to be 0.000125, respectively reduce the learning rate by 10 times in the 90 th round and the 120 th round, and the blocksize to be 8, and the optimization algorithm is set to be Adam algorithm, and by observing the loss value change curve of the verification set, when the loss value change curve shows a descending trend and tends to be stable after a certain number of iterations, the model is regarded as converging, and if the loss value change curve does not converge, the model is retrained by increasing or decreasing the learning rate by 10 times.
5. The method for automatically identifying, positioning and reading a digital instrument based on deep learning of claim 1, wherein the improved centrnet network model of step 5 is that based on the original centrnet network structure, the input size of the modified network is 256 x 3, the trunk extraction network selects ResNet-50, the last layer of downsampling operation in the ResNet-50 is replaced by hole convolution, and an angle regression branch is added to the head part;
the multilayer ceramic material specifically comprises two convolution layers and a ReLU active layer, wherein each layer comprises the following components: the convolution kernel size of the convolution layer of the first layer is 64 multiplied by 3, the feature map output is 64 multiplied by 64, the second layer is a ReLU active layer, the third layer is a convolution layer, namely an output layer, the convolution kernel size is 1 multiplied by 1, and the feature map output is 1 multiplied by 64; because an angle regression branch is added, the corresponding loss function to be added into the rotation angle regression adopts the minimum absolute value as the loss function, namely the L1 loss function, and the specific formula of the improved CenterNet loss function is as follows:
Figure FDA0003459865540000031
wherein L isCenterNetAs a loss function of the CenterNet network, λAngleThe weight of the angle regression branch loss, 0.5 is taken from the network,
Figure FDA0003459865540000032
as angle prediction value, ynIs the true rotation angle value.
6. The deep learning based digital instrument automatic identification, location and reading method of claim 1, wherein the tilt transformation in step 6 is to a central point (x)0，y0) As the origin, the rotation matrix is expressed as
Figure FDA0003459865540000033
Where θ is the tilt angle output in step 5.
7. The method for automatically identifying, positioning and reading the digital instrument based on the deep learning according to claim 1, wherein the training and optimization in the step 7 is to crawl images which take the digital display instrument and the digital display thermometer as key words in a network picture through a network crawler, screen and cut the images to obtain the images only containing digital parts, and then cut the images collected in the experiment to obtain a data set;
the training is divided into two stages, the first stage only uses the processed network pictures for training, and the training parameters are set as follows: the image input size is 32 multiplied by 100 multiplied by 3, the batch size is 128, the number of iteration rounds is set to 72 rounds, the initial learning rate is 0.005, and the optimization algorithm is Adam algorithm;
and in the second stage, dividing the data set into a training set, a verification set and a test set according to the ratio of 8:1:1, taking the optimal model obtained in the first stage as a pre-training model, and training on the model by using the training set, wherein the training parameters in the second stage are set as follows: the image input size is 32 × 100 × 3, the batchsize is 128, the iteration number is set to 144 rounds, the initial learning rate is 0.0005, the learning rate attenuation coefficient is 0.00004, the optimization algorithm is Adam algorithm, the first moment estimation and second moment estimation attenuation rates are 0.9 and 0.999, and L2 regularization is set;
by observing a loss value change curve of the verification set, when a descending trend appears and tends to be stable after a certain number of iterations, the model is regarded as convergent, and if the model does not converge, the model is retrained by increasing or reducing the learning rate according to a multiple of 10.
CN202210018923.4A 2022-01-07 2022-01-07 Digital instrument automatic identification, positioning and reading method based on deep learning Pending CN114782669A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210018923.4A CN114782669A (en) 2022-01-07 2022-01-07 Digital instrument automatic identification, positioning and reading method based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210018923.4A CN114782669A (en) 2022-01-07 2022-01-07 Digital instrument automatic identification, positioning and reading method based on deep learning

Publications (1)

Publication Number Publication Date
CN114782669A true CN114782669A (en) 2022-07-22

Family

ID=82423893

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210018923.4A Pending CN114782669A (en) 2022-01-07 2022-01-07 Digital instrument automatic identification, positioning and reading method based on deep learning

Country Status (1)

Country Link
CN (1) CN114782669A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115439849A (en) * 2022-09-30 2022-12-06 杭州电子科技大学 Instrument digital identification method and system based on dynamic multi-strategy GAN network
CN116380149A (en) * 2023-04-07 2023-07-04 深圳市兴源智能仪表股份有限公司 Method and system for testing rotation of instrument code wheel

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115439849A (en) * 2022-09-30 2022-12-06 杭州电子科技大学 Instrument digital identification method and system based on dynamic multi-strategy GAN network
CN115439849B (en) * 2022-09-30 2023-09-08 杭州电子科技大学 Instrument digital identification method and system based on dynamic multi-strategy GAN network
CN116380149A (en) * 2023-04-07 2023-07-04 深圳市兴源智能仪表股份有限公司 Method and system for testing rotation of instrument code wheel
CN116380149B (en) * 2023-04-07 2024-02-02 深圳市兴源智能仪表股份有限公司 Method and system for testing rotation of instrument code wheel

Similar Documents

Publication Publication Date Title
CN114782669A (en) Digital instrument automatic identification, positioning and reading method based on deep learning
CN108520206B (en) Fungus microscopic image identification method based on full convolution neural network
WO2020015330A1 (en) Enhanced neural network-based image restoration method, storage medium, and system
CN110059586B (en) Iris positioning and segmenting system based on cavity residual error attention structure
CN112686833B (en) Industrial product surface defect detection and classification device based on convolutional neural network
CN107680053A (en) A kind of fuzzy core Optimized Iterative initial value method of estimation based on deep learning classification
CN113034411B (en) Road disease picture enhancement method coupling traditional method and depth convolution countermeasure generation network
CN112767344A (en) Disease enhancement method based on vehicle-mounted camera shooting and coupling tradition and machine learning
CN111199245A (en) Rape pest identification method
CN113486963A (en) Density self-adaptive point cloud end-to-end sampling method
CN117576379B (en) Target detection method based on meta-learning combined attention mechanism network model
CN110751271A (en) Image traceability feature characterization method based on deep neural network
CN112967296B (en) Point cloud dynamic region graph convolution method, classification method and segmentation method
CN113408381B (en) Micro-expression classification method based on self-attention residual convolution neural network
CN116385293A (en) Foggy-day self-adaptive target detection method based on convolutional neural network
CN116310466A (en) Small sample image classification method based on local irrelevant area screening graph neural network
An Xception network for weather image recognition based on transfer learning
CN114067126A (en) Infrared image target detection method
CN115439669A (en) Feature point detection network based on deep learning and cross-resolution image matching method
CN111681176B (en) Self-adaptive convolution residual error correction single image rain removing method
CN110750672B (en) Image retrieval method based on deep measurement learning and structure distribution learning loss
CN111179224A (en) Joint learning-based reference-free evaluation method for aerial image restoration quality
CN109887014A (en) Method for registering images between a kind of stationary orbit spectral coverage neural network based
CN111882543A (en) Cigarette filter stick counting method based on AA R2Unet and HMM
CN117132592B (en) Industrial defect detection method based on entropy fusion

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