CN109657682B - Electric energy representation number identification method based on deep neural network and multi-threshold soft segmentation - Google Patents
Electric energy representation number identification method based on deep neural network and multi-threshold soft segmentation Download PDFInfo
- Publication number
- CN109657682B CN109657682B CN201811440923.3A CN201811440923A CN109657682B CN 109657682 B CN109657682 B CN 109657682B CN 201811440923 A CN201811440923 A CN 201811440923A CN 109657682 B CN109657682 B CN 109657682B
- Authority
- CN
- China
- Prior art keywords
- segmentation
- threshold
- probability
- electric energy
- image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 230000011218 segmentation Effects 0.000 title claims abstract description 137
- 238000000034 method Methods 0.000 title claims abstract description 33
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 17
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 11
- 238000001514 detection method Methods 0.000 claims description 18
- 238000005070 sampling Methods 0.000 claims description 8
- 230000002708 enhancing effect Effects 0.000 claims description 2
- 238000013135 deep learning Methods 0.000 abstract description 3
- 230000000694 effects Effects 0.000 abstract description 3
- 230000002411 adverse Effects 0.000 abstract description 2
- 230000009286 beneficial effect Effects 0.000 abstract description 2
- 238000013461 design Methods 0.000 abstract description 2
- 239000002699 waste material Substances 0.000 abstract description 2
- 238000005516 engineering process Methods 0.000 description 6
- 238000004458 analytical method Methods 0.000 description 4
- 238000011176 pooling Methods 0.000 description 4
- 238000012545 processing Methods 0.000 description 4
- 238000012549 training Methods 0.000 description 4
- 238000012015 optical character recognition Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 238000005286 illumination Methods 0.000 description 2
- 238000003709 image segmentation Methods 0.000 description 2
- 238000003909 pattern recognition Methods 0.000 description 2
- 241000209499 Lemna Species 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000003708 edge detection Methods 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/14—Image acquisition
- G06V30/148—Segmentation of character regions
- G06V30/153—Segmentation of character regions using recognition of characters or words
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to the technical field of automatic identification of electric power and electric energy meters, and provides an electric energy representation number identification method based on a deep neural network and multi-threshold soft segmentation. The architecture of the method contains two modules: a number area positioning module and a number identification module. The positioning module is based on a YOLO algorithm and adopts a network model suitable for the characteristics of an electric indicating area to perform rapid positioning, and the indicating area identification module designs an identification method combining multi-threshold soft segmentation and a deep neural network. The invention has the beneficial effects that: by adopting the method, the number indicating area in the ammeter picture can be quickly positioned and accurately identified, and manual meter reading errors and resource waste can be reduced; the method combines deep learning and multi-threshold segmentation, can effectively overcome the adverse effect of image noise, and improves the identification accuracy.
Description
Technical Field
The invention belongs to the technical field of automatic identification of electric energy meters, and relates to a technology for accurately positioning and identifying electric energy meter readings by using an image identification technology under a real shooting condition.
Background
The traditional manual in-process that acquires electric energy meter registration information often has a large amount of manpower extravagant and the error of checking meter, utilizes image recognition technology to carry out the analysis to the electric energy meter image of shooing, and then reads out registration information automatically, can effectively promote the efficiency of checking meter and reduce artificial error, has important meaning to the intellectuality of power technology. However, for the electric energy meter image shot in the real environment, due to the influences of factors such as illumination, shooting angle and stain on the electric energy meter panel, great challenges are brought to electric energy representation area detection and digital identification.
In the aspect of electric energy meter reading detection, due to multi-angle shooting, complex change of background images, uncontrollable illumination conditions and the like, the reading area of the electric energy meter is difficult to accurately and robustly position. Yanjuan et al (the university of Nanjing Miller, 2012), based on digital image processing, propose an electric energy meter recognition technique based on digital image processing. In the technology, an improved binarization algorithm and a cutting algorithm based on the geometric characteristics of the electric energy meter are adopted to position the area where the electric energy meter figures are located, but the method is difficult to obtain an accurate positioning result for electric energy meter images with complex shooting conditions. The Canny edge detection algorithm and the Union region analysis-based electric energy meter digital positioning method proposed by Lemna juncea et al ("OpenCV and LSSVM-based digital meter reading automatic identification", Fuzhou university, 2016.) cannot accurately position a real photo with a complex situation because of excessive dependence on the effect of electric energy meter image binarization. In recent years, the deep neural network has good performance in tasks such as image target detection, image text detection and the like, and provides important reference for designing an electric energy representation area detection system capable of adapting to various complex conditions. Zhou et al ("EAST: An Efficient and Accurate Scene Text Detector", Computer Vision and Pattern recognition. IEEE,2017:2642-2651.) propose a Text detection method EAST based on direct regression, which can accurately detect dense texts, but the method has the problem of incomplete detection area, and the direct application to the electric energy representation area detection will affect the subsequent recognition performance. Girshick et al ("Rich features technologies for Accurate Object Detection and sensing Segmentation", IEEE Conference on Computer Vision and Pattern registration. IEEE Computer Society,2014:580-587.) proposed a dual-stage Object Detection algorithm RCNN that can achieve higher accuracy, but has lower Detection efficiency due to the architectural design of dual-stage multitask. Based on the above problems, the invention is inspired by Redmon et al ("You Only Look one: Unifield, Real-Time Object Detection", Computer Vision and Pattern recognition. IEEE,2016:779-788.) target Detection method, and provides a single-stage electric energy representation area Detection method based on a deep neural network.
In terms of power representation number recognition, an actually photographed power meter image generally does not meet the clean and tidy background condition required by an Optical Character Recognition (OCR) system, and the complex noise information of the actually photographed power meter image makes the traditional OCR method not directly applicable to power representation number recognition. The electric energy meter number recognition method based on threshold segmentation and structural feature analysis by Yangjuan and the like can accurately recognize well-shot pictures, but under the condition of interference on shooting, the problems of segmentation errors, structural missing errors and the like can occur, so that the recognition fails. The Le-Net network proposed by Lecun et al ("Gradient-based left applied to document recognition", Proceedings of the IEEE,1998,86(11):2278-2324 ") has been used commercially for recognition of handwritten fonts with great success, but the input to the network is that an accurate single-character image is supposed to be segmented, and under the condition of high noise, it is generally difficult to obtain a complete digital image by segmenting the number indication area by using a single threshold, so the method cannot be directly applied.
Aiming at the problems, the invention provides an electric energy representation number recognition algorithm combining multi-threshold soft segmentation and a deep neural network, and the basic idea is to utilize a plurality of thresholds to carry out binarization on a representation area image and adopt a plurality of segmentation thresholds to obtain a reasonable digital segmentation candidate scheme as far as possible; then, carrying out segmentation point probability calculation on segmentation points of all candidate schemes, and finally, combining the neural network digital recognition prediction probability and the probability obtained by counting the image data set of the index region; and (3) calculating joint probability by using the three items of data to select a final segmentation scheme so as to effectively eliminate adverse effects caused by high noise. The algorithm is cooperated with a registration area detection module to realize the automatic identification method of the electric energy representation number.
Disclosure of Invention
The invention aims to provide an automatic meter reading method combining a multi-threshold soft segmentation mode with a deep neural network. Processing the target: the shot electric energy meter picture is processed for the purpose: and automatically identifying the indication information in the image of the electric energy meter.
The technical scheme of the invention is as follows:
a electric energy meter number identification method based on a deep neural network and multi-threshold soft segmentation is characterized in that the deep neural network is utilized to carry out number region positioning, and a multi-threshold soft segmentation mode is combined with the deep neural network to carry out number digital identification; the system specifically comprises two modules: a positioning module and an identification module;
(1) positioning module
Based on a target detection-YOLO algorithm, a deep neural network comprising 20 convolutional layers and 2 full-connection layers is designed to detect an index area; aiming at the geometrical characteristic that the electric energy representation area is a long rectangle, a rectangular receptive field is generated by adopting 1 x 4 convolution kernels on 8 th, 12 th, 15 th, 18 th and 20 th convolution layers to be close to the indication area, the output dimension of the last full-connection layer is 8 x 16, and the network architecture can improve the detection performance of the electric energy representation area under the condition of reducing the quantity of parameters;
(2) identification module
Step one, counting the ratio of the number of different numbers contained in the regions with different length-width ratios;
counting the number of digits contained in the correctly positioned pictures under different length-width ratios, combining the proportions with similar length-width ratios, and keeping q proportions, wherein the actual electric energy representation number is not more than p; dividing the above count by the total number of all pictures under the aspect ratio to obtain statistical data RpqDenotes the probability that the picture of the q-th ratio contains p numbers, where 1. ltoreq. p.ltoreq.8, 1. ltoreq. q.ltoreq.12, RpqThe composition set is named R;
secondly, enhancing the number indicating area image detected by the positioning module by using a histogram equalization method, and then carrying out binarization on the number indicating area image by using a Niblack algorithm by adopting a plurality of parameter settings to obtain a plurality of binarization results; the Niblack binarization operation is as formula 1; wherein m and s respectively represent a gray average value and a standard deviation in a range of w x w of the sliding window, and k is a controllable parameter which is manually set and has a value of 0.1-0.2;
TNiblackformula 1 of m + k ═ s
Thirdly, selecting segmentation points of the binarized index area by using a multi-threshold soft segmentation mode
The existing digital segmentation method generally adopts a single threshold value to segment the vertical projection graph of the registration area, and has low accuracy under the condition of large noise. The invention provides a multi-threshold soft segmentation mode, namely, a plurality of thresholds are utilized to segment an index region to obtain a plurality of segmentation schemes, and then the joint probability score of each segmentation is calculated to select the optimal segmentation result; three thresholds T are adopted, the first threshold is minimum, the third threshold is maximum, and the value ranges of the three thresholds are as follows: index image height [0.1,0.5 ]; segmenting the vertical projection graph; traversing the vertical projection graph according to the direction of the transverse axis of the image, regarding the point, corresponding to the projection value of which is lower than the threshold value, in the transverse axis as a segmentation candidate point, and regarding a continuous region of the segmentation candidate point as a segmentation range; and under the condition that the first and second thresholds are smaller, selecting the middle value of the segmentation range as a segmentation point. Under the condition of a third large threshold, sampling operation is carried out on points in the dividing range, and the obtained points are used as dividing points under the threshold. In the invention, 1% of the transverse length of the picture is used as a sampling interval, and a denser division point is obtained by using a smaller sampling interval.
Fourthly, determining a segmentation candidate scheme by using the segmentation points obtained above, and expanding the candidate scheme by using the failure condition of segmentation;
directly arranging 18 groups of segmentation points (k and w take three values respectively, and T takes the first two thresholds) obtained by the first threshold and the second threshold T in the third step on the horizontal axis of the image in sequence to obtain 18 segmentation schemes; and for the segmentation points obtained by the third threshold value T, randomly combining the segmentation points in sequence in the direction of the transverse axis of the image to obtain L segmentation schemes (the combination length does not exceed 7, because the maximum number of electric energy representations does not exceed 8), and the step can obtain total initial L +18 segmentation schemes.
Then, the obtained segmentation scheme is used for segmenting the registration area image, and a fixed threshold value T is used in the segmentation processwh(0.8) judging whether the segmentation of the aspect ratio of the picture obtained by segmentation is finished, and if the aspect ratio of the picture obtained by segmentation exceeds the threshold value, judging that the segmentation fails; selecting the probability R corresponding to the aspect ratio closest to the aspect ratio in the R according to the aspect ratio of the segmentation failure picturepqSelecting (p-1) mean value points for equidistant segmentation according to the number p of the highest and next highest numbers; randomly combining the mean segmentation points with the L +18 segmentation schemes obtained in the previous step in sequence in the direction of the transverse axis to obtain the final N segmentation schemes, wherein L +18 is less than or equal to N;
fifthly, counting the occurrence frequency a of each segmentation point obtained under N segmentation schemesiFind aiThe ratio of 27 (here, 3 k,3 sliding window sizes w, 3 slicing thresholds) is taken as the probability Pc of the slicing pointi(the same way of calculating the mean cut point of the previous step). In the process of segmentation, blank areas on two sides of a segmentation point are required to be merged into the segmented pictures to realize seamless connection between the segmented pictures, so that segmentation information is prevented from being lost, and identification missing errors are prevented;
and sixthly, sending the pictures obtained by each segmentation scheme into a trained network in sequence for recognition, and outputting a prediction result and probability. Then using the prediction probability Pr of each segmented picturenjAnd a segmentation point probability Pc corresponding to such a segmentation schemeiThe accumulated result is used as a temporary probability under the segmentation scheme (because a high-precision model usually has a higher prediction probability for a correctly identified result and a relatively lower prediction probability for a wrong identification, the probability can be used for eliminating the condition that a segmentation region is wrong to cause identification errors). Finally, multiplying the temporary probability by the probability R obtained by the second step of statisticspq(this is done to ensure that the identified length matches the true condition of the detection, preventing the output probability for long strings from being too low). In the invention, the product value is used as the joint probability of a segmentation scheme, and the prediction result O corresponding to the maximum value of the joint probability in each segmentation scheme is finally selectednjAs digital images of electric energy metersAnd (6) outputting the prediction.
The invention has the beneficial effects that: by adopting the method, the number indicating area in the ammeter picture can be accurately positioned and effectively identified, and the influence of high noise of the image is resisted; the manual meter reading error and the resource waste can be reduced; in addition, the invention combines deep learning and the traditional method, reduces the burden of the machine under the condition of ensuring the positioning and identifying accuracy, solves the defects of the existing problems and can assist the power analysis to a certain extent.
Drawings
FIG. 1 is a process flow of the present invention, wherein the blue blocks represent statistically derived data.
FIG. 2 is a diagram of multi-threshold vertical projection selection of slicing points.
Detailed Description
The method is used for building a network model based on a Pythroch deep learning framework and realizing multi-threshold soft segmentation processing by utilizing a Python language.
Step 1: a Labelimg tool was used to create a training data set representing the location of the number of zones of the electrical energy representation.
Step 2: and building a network architecture and training a network, wherein the hierarchical structure is as follows.
An input layer: the input picture is a shot electric energy meter image to be positioned, and the size is 448 x 448.
Convolution layer and full connection layer: a total of 20 convolutional layers, and 2 fully-connected layers; adopting 1 × 4 convolution kernels in 8 th, 12 th, 15 th, 18 th and 20 th convolution layers, adopting 3 × 3 convolution kernels in the rest convolution layers, and enabling the step size of all convolution layers to be 1; the penultimate layer full link layer outputs dimension 4096; the final fully-connected layer output dimension is 8 × 16;
pooling: selecting a maximum pooling mode, wherein the step length is 2;
an output layer: and outputting the predicted center coordinates and the length and width of the frame.
Step 3: and intercepting the electric energy representation area picture from the original picture according to the output information of the positioning network.
Step 4: counting the positive definite bitmap, recording the number of digits in the picture under different length-width ratios, and dividing the number by the number of digits in the picture under the length-width ratioCounting to obtain statistical data Rpq(wherein p is more than or equal to 1 and less than or equal to 8, and q is more than or equal to 1 and less than or equal to 12), the probability that the picture with the proportion of the qth contains p numbers is expressed. This set is named R.
Step 5: histogram equalization is carried out on the test image, and then the combination of three different k and three different sliding window sizes w is set according to formula 1, so that nine groups of binarization results are obtained. And selecting three threshold values T (the first threshold value is minimum, and the third threshold value is maximum) for the binarized index area, and performing segmentation point selection on the image by using vertical projection. For the results under the first and second threshold values, the median of the segmentation range (segmentation candidate point continuous region, as shown in fig. 2) is selected as the segmentation point. Under a third relatively large threshold, sampling is performed on points in the segmentation range, and the obtained points are used as segmentation points under the threshold (in the invention, 1% of the transverse length of the picture is used as a sampling interval, and a smaller sampling interval can be used for obtaining denser sampling).
And Step 6, determining a segmentation scheme according to the obtained segmentation points. And (4) directly arranging 18 groups of segmentation points (k and w take three values, and T takes the first two thresholds) obtained by the first threshold and the second threshold T according to the horizontal axis direction of the image to obtain 18 segmentation schemes. And for the segmentation points obtained by the third threshold value T, randomly combining the segmentation points in sequence in the direction of the transverse axis of the image to obtain L segmentation schemes (the combination length does not exceed 7, because the maximum number of electric energy representations does not exceed 8), and the step can obtain total initial L +18 segmentation schemes.
Step 7: and performing the number-indicating area image segmentation on the segmentation scheme obtained in the previous step, and reserving blank areas on two sides during segmentation in the process to prevent the segmentation information from being lost. While slicing, using a fixed threshold Twh(0.8 is used in the invention) to judge whether the segmentation of the picture obtained by segmentation is finished, if the segmentation is finished, the picture is regarded as failed segmentation, and the probability R corresponding to the closest length-width ratio in the R is selected according to the length-width ratio of the picture failed segmentationpqThe highest and next highest numbers p, and (p-1) mean points are selected for equidistant segmentation. The segmentation points and the previously obtained L +18 segmentation schemes are used for carrying out random combination in the direction of the transverse axis according to the sequence to obtain the final N types of segmentationThe scheme is that L +18 is less than or equal to N.
Step 8: counting the occurrence frequency a of each segmentation point obtained under all segmentation schemesi(i is the number of the dividing points) to obtain aiThe ratio of 27 (here, 3 k,3 sliding window sizes w, 3 slicing thresholds) is taken as the probability Pc of the slicing pointi(the same calculation method is applied to the mean cut point at Step 7).
Step 9: and manually segmenting the image of the electric energy meter number area positioned at Step 3, making a single character data set required by training and identifying the network, building the identifying network according to the figure 1, and training and identifying the network model by using the made data set.
An input layer: the input picture is a sliced electric energy meter image with the size of 32 × 32.
And (3) rolling layers: the convolution kernel size is 5 x 5, the step size is 1;
pooling: selecting a maximum pooling mode, wherein the step length is 2;
an output layer: outputting the predicted number category and the prediction probability thereof;
step 10: and performing index region image segmentation on each segmentation scheme of Step 7. Sending the divided pictures into a trained network for recognition to obtain a result O of prediction of each picturenjAnd probability Pr thereofnj. Calculating Pr according to horizontal segmentation ordernjMultiplied by the probability Pc of the cut point corresponding to such a cutting schemei. Finally multiplying by R closest to the aspect ratio of the imagepq. Obtaining a joint probability P under a segmentation schemenWherein N is more than or equal to 1 and less than or equal to N.
Step 11: calculating the joint probability P corresponding to all the N segmentation schemes according to the operation pair of Step 9nObtaining PnThe slicing scheme corresponding to the maximum value in the step (a) is used as the final slicing scheme, and all items output OnjThe decimal point is the final prediction result and is added in the output process when the decimal point is the second last digit.
Claims (1)
1. A electric energy representation number recognition method based on a deep neural network and multi-threshold soft segmentation is characterized in that the deep neural network is utilized to carry out registration area positioning, and a multi-threshold soft segmentation mode is combined with the deep neural network to carry out registration number recognition; the system specifically comprises two modules: a positioning module and an identification module;
(1) positioning module
Based on a target detection-YOLO algorithm, a deep neural network comprising 20 convolutional layers and 2 full-connection layers is designed to detect an index area; aiming at the geometrical characteristic that the electric energy representation area is a long rectangle, a rectangular receptive field is generated by adopting 1 x 4 convolution kernels on 8 th, 12 th, 15 th, 18 th and 20 th convolution layers to be close to the indication area, the output dimension of the last full-connection layer is 8 x 16, and the network architecture can improve the detection performance of the electric energy representation area under the condition of reducing the quantity of parameters;
(2) identification module
Step one, counting the ratio of the number of different numbers contained in the regions with different length-width ratios;
counting the number of digits contained in the correctly positioned pictures under different length-width ratios, combining the proportions with similar length-width ratios, and keeping q proportions, wherein the actual electric energy representation number is not more than p; dividing the above count by the total number of all pictures under the aspect ratio to obtain statistical data RpqDenotes the probability that the picture of the q-th ratio contains p numbers, where 1. ltoreq. p.ltoreq.8, 1. ltoreq. q.ltoreq.12, RpqThe composition set is named R;
secondly, enhancing the number indicating area image detected by the positioning module by using a histogram equalization method, and then carrying out binarization on the number indicating area image by using a Niblack algorithm by adopting a plurality of parameter settings to obtain a plurality of binarization results; the Niblack binarization operation is as formula 1; wherein m and s respectively represent a gray average value and a standard deviation in a range of w x w of the sliding window, and k is a controllable parameter which is manually set and has a value of 0.1-0.2;
TNiblackformula 1 of m + k ═ s
Thirdly, selecting segmentation points of the binarized index area by using a multi-threshold soft segmentation mode;
a multi-threshold soft segmentation mode is adopted, namely a plurality of thresholds are utilized to segment the index region to obtain a plurality of segmentation schemes, and then the joint probability score of each segmentation is calculated to select the optimal segmentation result; the method adopts three thresholds T, wherein the first threshold is minimum, the third threshold is maximum, and the value ranges of the three thresholds are as follows: index image height [0.1,0.5 ]; segmenting the vertical projection graph; traversing the vertical projection graph according to the direction of the transverse axis of the image, regarding the point, corresponding to the projection value of which is lower than the threshold value, in the transverse axis as a segmentation candidate point, and regarding a continuous region of the segmentation candidate point as a segmentation range; under the conditions of a first threshold and a second threshold, selecting a middle value of a segmentation range as a segmentation point; under the condition of a third large threshold value, sampling operation is carried out on points in the dividing range, and the obtained points are used as dividing points under the threshold value;
fourthly, determining a segmentation candidate scheme by using the segmentation points obtained above, and expanding the candidate scheme by using the failure condition of segmentation;
taking three values of k and w for 18 groups of segmentation points obtained by the first threshold value and the second threshold value T in the third step, and directly arranging the first two threshold values T on the horizontal axis of the image in sequence to obtain 18 segmentation schemes; for segmentation points obtained by the third threshold value T, randomly combining the segmentation points in the horizontal axis direction of the image in sequence to obtain L types of segmentation schemes, and obtaining a primary L +18 types of segmentation schemes;
then, the obtained segmentation scheme is used for segmenting the registration area image, and a fixed threshold value T is used in the segmentation processwhJudging whether the segmentation is finished or not for the aspect ratio of the image obtained by segmentation, and if the aspect ratio exceeds the threshold value, judging that the segmentation fails; selecting the probability R corresponding to the aspect ratio closest to the aspect ratio in the R according to the aspect ratio of the segmentation failure picturepqSelecting (p-1) mean value points for equidistant segmentation according to the number p of the highest and next highest numbers; randomly combining the mean segmentation points with the L +18 segmentation schemes obtained in the previous step in sequence in the direction of the transverse axis to obtain the final N segmentation schemes, wherein L +18 is less than or equal to N;
fifthly, counting the occurrence frequency a of each segmentation point obtained under N segmentation schemesiFind aiThe ratio of 27 is taken as the probability Pc of the cut pointi；
In the segmentation process, blank areas on two sides of the segmentation point are also merged into the segmentation pictures to realize seamless connection between the segmentation pictures, so that the loss of segmentation information is prevented, and the identification missing error is avoided;
sixthly, the pictures obtained by each segmentation scheme are sequentially sent to a trained network for recognition, and a prediction result and probability are output; then using the prediction probability Pr of each segmented picturenjAnd a segmentation point probability Pc corresponding to such a segmentation schemeiThe accumulated multiplication result is used as a temporary probability under the segmentation scheme; finally, multiplying the temporary probability by the probability R obtained by the first step of statisticspqTaking the product value as the joint probability of a segmentation scheme, and finally selecting a prediction result O corresponding to the maximum value of the joint probability in each segmentation schemenjAs the prediction output of the digital image of the electric energy meter.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811440923.3A CN109657682B (en) | 2018-11-29 | 2018-11-29 | Electric energy representation number identification method based on deep neural network and multi-threshold soft segmentation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811440923.3A CN109657682B (en) | 2018-11-29 | 2018-11-29 | Electric energy representation number identification method based on deep neural network and multi-threshold soft segmentation |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109657682A CN109657682A (en) | 2019-04-19 |
| CN109657682B true CN109657682B (en) | 2022-02-11 |
Family
ID=66111100
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811440923.3A Active CN109657682B (en) | 2018-11-29 | 2018-11-29 | Electric energy representation number identification method based on deep neural network and multi-threshold soft segmentation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109657682B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110688900A (en) * | 2019-08-27 | 2020-01-14 | 国网河北省电力有限公司电力科学研究院 | Withdrawal meter management method based on image recognition |
| CN112770080B (en) * | 2019-11-01 | 2023-01-03 | 中移物联网有限公司 | Meter reading method, meter reading device and electronic equipment |
| CN111461121A (en) * | 2020-05-18 | 2020-07-28 | 江苏电力信息技术有限公司 | Electric meter number identification method based on YO L OV3 network |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20040091271A (en) * | 2003-04-21 | 2004-10-28 | 주식회사 팬택 | Method for concluding threshold for image division |
| CN104050450A (en) * | 2014-06-16 | 2014-09-17 | 西安通瑞新材料开发有限公司 | Vehicle license plate recognition method based on video |
| CN104298976A (en) * | 2014-10-16 | 2015-01-21 | 电子科技大学 | License plate detection method based on convolutional neural network |
| CN108875735A (en) * | 2018-05-25 | 2018-11-23 | 昆山湖大机器人技术有限公司 | Plate Production line lattice coding character automatic testing method |
-
2018
- 2018-11-29 CN CN201811440923.3A patent/CN109657682B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20040091271A (en) * | 2003-04-21 | 2004-10-28 | 주식회사 팬택 | Method for concluding threshold for image division |
| CN104050450A (en) * | 2014-06-16 | 2014-09-17 | 西安通瑞新材料开发有限公司 | Vehicle license plate recognition method based on video |
| CN104298976A (en) * | 2014-10-16 | 2015-01-21 | 电子科技大学 | License plate detection method based on convolutional neural network |
| CN108875735A (en) * | 2018-05-25 | 2018-11-23 | 昆山湖大机器人技术有限公司 | Plate Production line lattice coding character automatic testing method |
Non-Patent Citations (2)
| Title |
|---|
| License Plate Detection Based on Rectangular Features and Multilevel Thresholding;Ihsan Ullah et al.;《Proceedings of the International Conference on Image Processing, Computer Vision, and Pattern Recognition (IPCV)》;20160731;153-157 * |
| 复杂背景下车牌识别算法的研究与实现;徐凯;《中国优秀硕士学位论文全文数据库 (工程科技Ⅱ辑)》;20180215;C034-1081 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109657682A (en) | 2019-04-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112884064B (en) | Target detection and identification method based on neural network | |
| CN110533084B (en) | Multi-scale target detection method based on self-attention mechanism | |
| CN111325203B (en) | American license plate recognition method and system based on image correction | |
| CN111709935B (en) | Real-time coal gangue positioning and identifying method for ground moving belt | |
| CN109190752A (en) | The image, semantic dividing method of global characteristics and local feature based on deep learning | |
| CN108334881B (en) | License plate recognition method based on deep learning | |
| CN109657682B (en) | Electric energy representation number identification method based on deep neural network and multi-threshold soft segmentation | |
| WO2021139197A1 (en) | Image processing method and apparatus | |
| CN110598698B (en) | Natural scene text detection method and system based on adaptive regional suggestion network | |
| CN111768365A (en) | Solar cell defect detection method based on convolutional neural network multi-feature fusion | |
| CN114972312A (en) | Improved insulator defect detection method based on YOLOv4-Tiny | |
| CN110969164A (en) | Low-illumination imaging license plate recognition method and device based on deep learning end-to-end | |
| CN111414938B (en) | Target detection method for bubbles in plate heat exchanger | |
| CN113780423A (en) | Single-stage target detection neural network based on multi-scale fusion and industrial product surface defect detection model | |
| CN116029979A (en) | Cloth flaw visual detection method based on improved Yolov4 | |
| CN111652117A (en) | Method and medium for segmenting multi-document image | |
| CN111178405A (en) | Similar object identification method fusing multiple neural networks | |
| CN111862223B (en) | Visual counting and positioning method for electronic element | |
| CN116433494B (en) | File scanning image automatic correction and trimming method based on deep learning | |
| CN112381794A (en) | Printing defect detection method based on deep convolution generation network | |
| CN110136098B (en) | Cable sequence detection method based on deep learning | |
| CN117132531A (en) | Lightweight-based YOLOv5 insulator defect detection method | |
| CN111832497A (en) | Text detection post-processing method based on geometric features | |
| CN111160262A (en) | Portrait segmentation method fusing human body key point detection | |
| CN116612398A (en) | Unmanned aerial vehicle inspection photo pole tower license plate character recognition method based on CTPN algorithm |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |