CN113191343A - Aviation wire identification code automatic identification method based on convolutional neural network - Google Patents

Aviation wire identification code automatic identification method based on convolutional neural network Download PDF

Info

Publication number
CN113191343A
CN113191343A CN202110346523.1A CN202110346523A CN113191343A CN 113191343 A CN113191343 A CN 113191343A CN 202110346523 A CN202110346523 A CN 202110346523A CN 113191343 A CN113191343 A CN 113191343A
Authority
CN
China
Prior art keywords
wire
character
neural network
convolutional neural
data
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
CN202110346523.1A
Other languages
Chinese (zh)
Inventor
许艾明
黎小华
刘倍铭
方亿
李仁宏
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.)
Chengdu Aircraft Industrial Group Co Ltd
Original Assignee
Chengdu Aircraft Industrial Group Co Ltd
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 Chengdu Aircraft Industrial Group Co Ltd filed Critical Chengdu Aircraft Industrial Group Co Ltd
Priority to CN202110346523.1A priority Critical patent/CN113191343A/en
Publication of CN113191343A publication Critical patent/CN113191343A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/22—Image preprocessing by selection of a specific region containing or referencing a pattern; Locating or processing of specific regions to guide the detection or recognition
    • 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
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/205—Parsing
    • G06F40/216—Parsing using statistical methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/279—Recognition of textual entities
    • G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
    • 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

Abstract

The invention belongs to the technical field of image recognition, and particularly relates to an aviation wire identification code automatic recognition method based on a convolutional neural network, which comprises the steps of collecting image data and forming a training set; carrying out sample training on the acquired image data, and cleaning six types of low-quality image data in a training set; carrying out enhancement processing on the cleaned image data; a Bi-GRU circulating layer network is connected in series behind a convolution layer of a convolutional neural network CNN, a Dropout layer is fused at the tail end, and a training strategy from easy to difficult is adopted to improve the identification precision and reduce the training time; and finally, generating a wire code noun dictionary based on the wire code character sequence in the data set, and matching the recognition result with the noun dictionary to obtain a final recognition result. The technical scheme is based on the convolutional neural network, the recognition accuracy is improved, the model has higher robustness and high recognition speed, and compared with a common end-to-end training scheme, the convolutional neural network is used as the main component of the network, and the parallelism is high.

Description

Aviation wire identification code automatic identification method based on convolutional neural network
Technical Field
The invention belongs to the technical field of image recognition, and particularly relates to an aviation wire identification code automatic recognition method based on a convolutional neural network.
Background
The aviation wire identification code contains important information such as the manufacturing standard, type, specification, manufacturing country, manufacturer, manufacturing date and the like of the cable wire, and the identification of the identification code is the basis of wire use and maintenance. In the background of the prior art, a text recognition technology by means of an image is common. The method is mainly classified into Optical Character Recognition (OCR) and Character Recognition in natural Scenes (STR). Optical character recognition, OCR, is primarily directed to optical character recognition of scanned documents and is well established in terms of theory and application. The character recognition STR in the natural scene aims at the character recognition contained in the natural scene image, and is limited by the image degradation condition of various reasons in the scene, such as complex image background, character region deformation and the like, and the recognition rate is low. The aviation wire rod identification code identification technology is character identification in a natural scene, and due to the reasons of shielding of other wire rods, change of scene illumination, deformation and bending of characters under a shooting angle and the like in the identification process, a network model is low in robustness, long in training period, low in wire rod identification code identification rate and large in image marking amount. The prior art center also has some technical solutions proposed to solve the aforementioned problems, such as: the Chinese patent application (201811272122.0) discloses a natural scene character recognition method and device, wherein the method comprises the following steps: on the basis of text detection, space transformation and text recognition, a natural scene graph is divided into a mask graph and a pixel graph, and finally the pixel graph is transformed into a rectifying graph, so that the problem of low accuracy rate of character sequence recognition caused by the interference and transformation effect of information except characters in a natural scene is solved. Whether the network model is in a training stage or not needs to be continuously judged, the method is used for calculating segmentation errors to optimize the network model, the efficiency is low, and the improvement of the recognition rate is not obvious. The Chinese patent application (201910112721.4) discloses a natural scene character recognition method for recognizing warehouse shelf nameplates, which is characterized in that a nameplate character detection network and a character recognition network are respectively built, and the method has good accuracy, precision and recall rate for recognizing the characters of warehouse shelf nameplates. However, the training sets are all images with clear and standard framing, and are only suitable for the cases of simple natural scenes such as shelves and small deformed characters.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides an aviation wire identification code automatic identification technology based on a convolutional neural network, which is used for cleaning low-quality images and enhancing data; constructing a CRNN network structure through a convolutional neural network and a superposition loop layer, and training by adopting a training strategy from easy to difficult; and finally, decoding and integrating based on the customized dictionary to recognize the identification code. The robustness and the safety of the system are improved, the recognition accuracy can be effectively improved, and meanwhile the training time is reduced.
The invention is realized by the following technical scheme:
firstly, the technical scheme relates to and provides a simple aviation wire identification code automatic identification device, which mainly comprises a data acquisition part, a data transmission part and a simulation training and identification part as shown in fig. 2. The data acquisition can adopt an industrial high-definition camera for acquiring the aerial wire identification code image to be identified, and the industrial high-definition camera can adjust the height and the angle, so that the imaging plane and the identification code position are parallel as much as possible to reduce the character deformation caused by improper shooting angle; the data transmission part can adopt a router, and further can enable the industrial high-definition camera to be in communication connection with the router through a data line cable, so that the acquired image data can be transmitted to the simulation training and recognition part through the router; further, the model training and identifying part can adopt an upper computer. Preferably, in practical applications, the data transmission mode is not limited to the wired transmission mode, and a wireless data transmission mode may also be adopted, specifically, which data transmission mode is adopted, and the embodiment of the present invention is not limited uniquely.
An aviation wire identification code automatic identification method based on a convolutional neural network is characterized in that: the method comprises the steps of data acquisition, data cleaning, data enhancement, network model training and decoding integration;
and the data acquisition step comprises the steps of acquiring image data of the aviation wire identification code by using a high-definition industrial camera, transmitting the acquired image data to an upper computer through a router, and collecting the image data to be identified and wire code character information contained in the image by using the upper computer to form a training set.
And the data cleaning is to perform sample training on the acquired image data by using an upper computer, and clean six types of low-quality image data which have label errors, vertically arranged characters, excessive image transverse compression, no identification code characters in the image and character shielding and/or character over-blurring caused by wires in a training set in the sample training process. Wherein, to label mistake, need artifical discernment wire rod characters and write back correct label information wherein. Aiming at the vertically arranged characters, mainly the vertically arranged aviation wire identification codes, firstly, the number n of the characters in the image, the width L of the characters and the whole width L of the character arrangement are obtainedsCalculating a minimum width threshold LminAnd L is determinedsWhether or not it is greater than Lmin(ii) a If L iss≤LminJudging that the characters are vertically arranged; if L iss＞LminThen, the character is judged to be non-vertically arranged. And for the image with the over-horizontal compression of the image, reserving the image in the data set, and writing the image into the data set after the image is subjected to horizontal reduction processing. And directly performing abandoning processing on characters without identification codes in the image. Aiming at the character shielding caused by the wire, the number and the continuous quantity of the shielded characters need to be judged, and the shielding quantity N of the continuous characters isc<3 and total number of shielding words Nt<Wire drawing of 4The sheet is retained, and other shielding wire pictures are judged to have too large shielding amount, so that automatic identification is not suitable. And (4) screening based on the variance threshold value after the Laplace transformation aiming at the character over-blurring, and deleting the picture if the actual variance value is smaller than the threshold variance value and the picture is judged to be over-blurring. Furthermore, a small amount of character shielding and transverse compression excessive data are not cleaned, so that training pictures in a data set are closer to a real scene, the situation which possibly occurs in the actual recognition process is restored, and the robustness of the CRNN model built in the later stage can be improved. And processing the retained images after data cleaning into images with uniform size so as to conveniently perform feature learning on the images with fixed size by utilizing the CRNN constructed in the later stage.
And the data enhancement is to enhance the cleaned image data and write the enhanced picture into the data set so as to expand the rare sample of the data set. The enhancement processing is that the expanded image is subjected to operations including 1-180 degree random rotation and image brightness random change, the situation that wire positions possibly exist in a real scene are disordered and the situation that spot light is too bright and too dark is identified is restored, so that the reality of data is increased, furthermore, in the process of identifying the image, a hacker can apply slight disturbance to an original picture by using model parameters, although the disturbed picture cannot cheat human eyes, the interior of deep learning is a high-dimensional space and is easily influenced by the high-dimensional space, so that characters are identified as wrong results, therefore, the technical scheme faces frequent malicious attacks in the modern network environment, and is blended with a countermeasure sample (the countermeasure sample is provided by Chtian Szegedy et al and is an input sample formed by intentionally adding slight interference in a data set, causing the model to give an erroneous output with high confidence. Under the regularization background, the error rate of an original independent and identically distributed test set is reduced through countertraining, namely a network is trained on a counterdisturbance training set sample), specifically, slight Gaussian noise is added around the wire rod characters in a data set picture to resist the disturbance attack applied to the character picture by a hacker, so that the algorithm safety coefficient is improved.
The network model is trained, a CRNN network structure is built, a Bi-GRU circulating layer network is connected in series behind a convolutional layer of a convolutional neural network CNN, a Dropout layer is fused at the tail end, and a training strategy from easy to difficult is adopted to improve the recognition accuracy and reduce the training time. Specifically, as shown in fig. 4, in order to construct a CRNN network model structure, the invention performs character recognition on an aviation wire identification code in a natural scene, and increases the number of convolution kernels to improve the feature extraction capability of the model; the convolutional neural network CNN does not have the capability of identifying the sequence and cannot convert the identified single characters into an effective wire number sequence, so that a Bi-GRU circulating layer network is connected in series after the convolutional layer of the convolutional neural network, and a Dropout layer is fused at the tail end of the convolutional neural network, so that the robustness of the model is improved. Further, the CRNN network structure specifically includes 1 Input layer, 6 convolutional layers Conv, 4 max pooling layers MaxPool, 1 full connection layer FC, 1 bidirectional full connection layer (FC 1 and FC 2), 1 bidirectional GRU layer (GRU 1 and Reverse-GRU), and 1 Dropout layer. The Input layer is arranged to facilitate the CRNN to receive an Input image; the convolutional layer Conv mainly functions to convert an input image into a feature matrix; the 4 maximum pooling layers MaxPool are used for compressing the input feature map, and mainly comprise the steps of reducing the feature map to simplify the network calculation complexity and performing feature compression to extract main features; after the feature extraction is finished, 1 full connection layer FC, 1 bidirectional full connection layer, 1 bidirectional GRU layer and 1 Dropout layer are used for sorting and automatically coding the identified features into an aviation wire feature digital sequence, and the output of the last pooling layer is flattened into a 2 x 1-dimensional feature vector for connecting two bidirectional full connection layers FC1 and FC2, so that the model can be finally output as a string of aviation wire standard code values. Furthermore, because a plurality of non-standard data exist in the sample and the noise of the normal data is very large, the model is difficult to converge at the initial training stage, so the recognition accuracy is improved and the training time is reduced by adopting a scheme from easy to difficult. The method enables the model to be more easily converged, so that the model has a good recognition effect on high-quality conventional pictures and also has a good detection effect on low-quality wire rod pictures. In a commonly used end-to-end Seq2Seq scheme, the recurrent neural network RNN accounts for a major component, resulting in low model parallelism. The invention takes the convolutional neural network CNN as the main component, and has high parallelism and high recognition speed. The convolutional neural network CNN is a typical deep learning method, realizes complex function approximation and input data characterization by learning a deep nonlinear network structure, and shows the learning capability of strong essential characteristics of a data set. The target features in a large number of data sets can be automatically learned without human involvement in the feature selection process. The weight sharing and local connection mechanism enables the device to have certain invariance to geometric transformation, deformation and illumination, and simultaneously has good fault-tolerant capability and learning capability. The advantages enable the convolutional neural network to have great advantages in processing the problems under the conditions that the environment information is unknown and the inference rule is not clear, and can adapt to the environment problem that the image to be recognized is complex.
The decoding integration: and generating a wire code noun dictionary based on the wire code character sequence in the aviation wire data set, and matching the recognition result of the CRNN with the noun dictionary, so that the accuracy of the recognized wire code character sequence is improved, the wire code is automatically recognized to obtain a better result, and the result is a final recognition result. Specifically, after the neural network identifies a specific character sequence in the aviation wire, in order to make an output result more accurate, an aviation wire identification scene wire code noun dictionary is set, and the process comprises a decoding process and an integration process. The decoding process comprises the steps of performing word segmentation processing on a wire code character label of an aviation wire data set by using Jieba word segmentation to form a noun dictionary special for the data set, then performing word segmentation on a character sequence identified by a Convolutional Neural Network (CNN), and searching a noun with the highest similarity in the noun dictionary based on an obtained word segmentation result to serve as a character identification result; the integration process is that when the identification code image is partially shielded, voting is carried out on each character before decoding, the character which has the most votes is selected as the shielded character, and the shielded character and the character identification result are integrated to obtain the final identification result.
The beneficial effect that this technical scheme brought:
(1) the technical scheme is based on the convolutional neural network, so that the identification accuracy is improved, the model has higher robustness and high identification speed, and compared with a common end-to-end training scheme, the convolutional neural network is used as the main component of the network, so that the parallelism is high;
(2) a few training samples need to be collected, and a part of training samples are generated by using a data enhancement mode, so that the sample collection quantity is reduced;
(3) a countermeasure sample is integrated in the training set for strengthening training, so that the safety coefficient of the model is improved, and external malicious attack is prevented;
(4) the industrial camera is used for carrying out image acquisition on the aviation wire identification code, the convolutional neural network model is used for identifying the wire identification code, and the upper computer is used for processing the whole process, so that the automation of the identification process is realized, the manual troubleshooting error is reduced, the efficiency is improved, and the management maintenance, the fault finding and the wire replacement which is equivalent when necessary are facilitated;
(5) the direct contact between an operator and electricity is reduced, and the operation safety is realized.
Drawings
FIG. 1 is a schematic flow chart of the present invention
FIG. 2 is a schematic view of a simple aviation wire identification code automatic identification device;
FIG. 3 is a schematic diagram of the image pair defense reinforcement of the aviation wire identification code of the present invention;
fig. 4 is a schematic diagram of a CRNN network structure according to an embodiment of the present invention;
fig. 5 is a table diagram of parameters of each layer in the CRNN network structure.
Detailed Description
The invention is further described in the following with reference to the drawings and examples, but it should not be understood that the invention is limited to the examples below, and variations and modifications in the field of the invention are intended to be included within the scope of the appended claims without departing from the spirit of the invention.
Example 1
The embodiment discloses an aviation wire identification code automatic identification method based on a convolutional neural network, which is used as a basic implementation scheme of the invention and comprises the steps of data acquisition, data cleaning, data enhancement, network model training and decoding integration;
acquiring data, namely acquiring image data of the aviation wire identification code by using a high-definition industrial camera, transmitting the acquired image data to an upper computer through a router, and collecting the image data to be identified and wire code character information contained in an image by using the upper computer to form a training set;
data cleaning, namely performing sample training on the acquired image data by using an upper computer, and cleaning six types of low-quality image data which have label errors, vertically arranged characters, excessive transverse image compression, no identification code characters in an image and character shielding and/or character over-blurring caused by wires in a training set in the sample training process;
and data enhancement, namely enhancing the cleaned image data, and writing the enhanced picture into a data set so as to expand the rare sample of the data set.
Training a network model, constructing a CRNN network structure, connecting a Bi-GRU circulating layer network in series after a convolutional layer of a Convolutional Neural Network (CNN), fusing a Dropout layer at the tail end, and improving the identification precision and reducing the training time by adopting a training strategy from easy to difficult;
decoding integration: and generating a wire code noun dictionary based on the wire code character sequence in the aviation wire data set, and matching the recognition result of the CRNN with the noun dictionary to obtain a final recognition result.
This technical scheme uses the industry camera to carry out image acquisition to aviation wire rod identification code, and the discernment of wire rod identification code is carried out to application convolution neural network model, uses upper computer processing whole process to realize the automation of identification process, reduce artifical investigation error, improved efficiency, so that the wire rod replacement that the management is maintained, is seeked trouble and is equated when necessary. The technical scheme is based on the convolutional neural network, so that the identification accuracy is improved, the model has higher robustness and high identification speed, and compared with a common end-to-end training scheme, the convolutional neural network is used as the main component of the network, so that the parallelism is high; furthermore, a part of training samples are generated in a data enhancement mode, and the number of collected samples is reduced.
Example 2
The embodiment discloses an aviation wire identification code automatic identification method based on a convolutional neural network, which is a basic implementation scheme of the invention, namely in the embodiment 1, aiming at an image with vertically arranged characters in the process of data cleaning, the number n of the characters in the image, the width L of the characters and the whole width L of the character arrangement are obtainedsCalculating a minimum width threshold LminAnd L is determinedsWhether or not it is greater than Lmin(ii) a If L iss≤LminJudging that the characters are vertically arranged; if L iss＞LminJudging that the characters are not vertically arranged; further, aiming at character shielding caused by the wire, the number and the continuous quantity of the shielded characters are judged, and the shielding quantity N of the continuous characters is judgedc<3 and total number of shielding words Nt<4, the wire rod pictures are reserved, and other shielded wire rod pictures are judged to have too large shielding amount, so that automatic identification is not suitable; and (4) screening based on the variance threshold value after the Laplace transformation aiming at the character over-blurring, and deleting the picture if the actual variance value is smaller than the threshold variance value and the picture is judged to be over-blurring. And finally, processing the image remained after the data cleaning into an image with a uniform size.
Example 3
The embodiment discloses an aviation wire identification code automatic identification method based on a convolutional neural network, which is a basic implementation scheme of the invention, namely in the data enhancement process, the enhancement processing is to restore the conditions of disordered wire positions and over-bright and over-dark identification of field lights possibly existing in a real scene by performing operations including random rotation of 1-180 degrees and random change of image brightness on an extended image, and integrate a countermeasure sample to improve the safety coefficient of an algorithm.
According to the technical scheme, the confrontation sample is integrated into the training set for strengthening training, the safety coefficient of the model is improved, and external malicious attack is prevented.
Example 4
As a basic implementation scheme of the present invention, in embodiment 1, a CRNN network structure includes 1 Input layer, 6 convolutional layers Conv, 4 maximum pooling layers MaxPool, 1 full-link layer FC, 1 bidirectional full-link layer, 1 bidirectional GRU layer, and 1 Dropout layer. Furthermore, the easy-to-go training strategy is to train the model accuracy by using the pictures with clear text identification and moderate picture angle and brightness, and then train the six sub-six low-quality pictures in the data cleaning process.
Example 5
The embodiment discloses an aviation wire identification code automatic identification method based on a convolutional neural network, which is used as a basic implementation scheme of the invention, namely in the embodiment 1, decoding integration comprises a decoding process and an integration process; the decoding process is to use Jieba word segmentation to perform word segmentation processing on a wire code character label of the aviation wire data set to form a noun dictionary special for the data set, then perform word segmentation on a character sequence identified by the CRNN network, and search a noun with the highest similarity in the noun dictionary based on the obtained word segmentation result to serve as a character identification result; the integration process is that when the identification code image is partially shielded, voting is carried out on each character before decoding, the character which has the most votes is selected as the shielded character, and the shielded character and the character identification result are integrated to obtain the final identification result.

Claims (9)

1. An aviation wire identification code automatic identification method based on a convolutional neural network is characterized in that: the method comprises the steps of data acquisition, data cleaning, data enhancement, network model training and decoding integration;
the data acquisition comprises the steps of acquiring image data of the aviation wire identification code by using a high-definition industrial camera, transmitting the acquired image data to an upper computer through a router, and collecting the image data to be identified and wire code character information contained in an image by using the upper computer to form a training set;
the data cleaning is to perform sample training on the acquired image data by using an upper computer, and in the sample training process, six types of low-quality image data, including label errors, vertically arranged characters, excessive transverse image compression, no identification code characters in an image, character shielding caused by wires and/or over-fuzzy characters, in a training set are cleaned;
the data enhancement is to enhance the cleaned image data and write the enhanced picture into the data set so as to expand the rare sample of the data set;
the network model is trained, a CRNN network structure is built, a Bi-GRU circulating layer network is connected in series behind a convolutional layer of a convolutional neural network CNN, a Dropout layer is fused at the tail end, and a training strategy from easy to difficult is adopted to improve the recognition accuracy and reduce the training time;
the decoding integration: and generating a wire code noun dictionary based on the wire code character sequence in the aviation wire data set, and matching the recognition result of the CRNN with the noun dictionary to obtain a final recognition result.
2. The aviation wire identification code automatic identification method based on the convolutional neural network as claimed in claim 1, characterized in that: in the process of data cleaning, the number n of characters in the image, the width L of the characters and the whole width L of the character arrangement are obtainedsCalculating a minimum width threshold LminAnd L is determinedsWhether or not it is greater than Lmin(ii) a If L iss≤LminJudging that the characters are vertically arranged; if L iss＞LminThen, the character is judged to be non-vertically arranged.
3. The aviation wire identification code automatic identification method based on the convolutional neural network as claimed in claim 1, characterized in that: in the data cleaning process, aiming at character shielding caused by the wire rods, the number and the continuous quantity of the shielded characters are judged, and the continuous character shielding quantity N is judgedc<3 and total number of shielding words Nt<And 4, the wire rod pictures are reserved, and other shielded wire rod pictures are judged to have overlarge shielding amount, so that automatic identification is not suitable.
4. The aviation wire identification code automatic identification method based on the convolutional neural network as claimed in claim 1, characterized in that: in the data cleaning process, the characters are over-fuzzy, screening is carried out based on the variance threshold value after the Laplace transformation, when the actual variance value is smaller than the threshold variance value, the pictures are judged to be over-fuzzy, and the pictures are deleted.
5. The aviation wire identification code automatic identification method based on the convolutional neural network as claimed in claim 1, characterized in that: and in the process of data cleaning, processing the images remained after the data cleaning into images with uniform size.
6. The aviation wire identification code automatic identification method based on the convolutional neural network as claimed in claim 1, characterized in that: in the data enhancement process, the enhancement processing is to restore the conditions of disordered wire positions and recognition of over-bright and over-dark field lights possibly existing in a real scene by performing operations including random rotation of 1-180 degrees and random change of image brightness on an extended image, and to integrate into a countermeasure sample to improve the safety coefficient of the algorithm.
7. The aviation wire identification code automatic identification method based on the convolutional neural network as claimed in claim 1, characterized in that: the CRNN network structure includes 1 Input layer, 6 convolutional layers Conv, 4 maximum pooling layers MaxPool, 1 full connection layer FC, 1 bidirectional full connection layer, 1 bidirectional GRU layer, and 1 Dropout layer.
8. The aviation wire identification code automatic identification method based on the convolutional neural network as claimed in claim 1, characterized in that: in the network model training process, the easy-to-go training strategy is to train the model precision by using pictures with clear text identification and moderate picture angle and brightness, and then train by using the six sub-classes of low-quality pictures in the data cleaning process.
9. The aviation wire identification code automatic identification method based on the convolutional neural network as claimed in claim 1, characterized in that: the decoding integration comprises a decoding process and an integration process; the decoding process is to use Jieba word segmentation to perform word segmentation processing on a wire code character label of the aviation wire data set to form a noun dictionary special for the data set, then perform word segmentation on a character sequence identified by the CRNN network, and search a noun with the highest similarity in the noun dictionary based on the obtained word segmentation result to serve as a character identification result; the integration process is that when the identification code image is partially shielded, voting is carried out on each character before decoding, the character which has the most votes is selected as the shielded character, and the shielded character and the character identification result are integrated to obtain the final identification result.
CN202110346523.1A 2021-03-31 2021-03-31 Aviation wire identification code automatic identification method based on convolutional neural network Pending CN113191343A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110346523.1A CN113191343A (en) 2021-03-31 2021-03-31 Aviation wire identification code automatic identification method based on convolutional neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110346523.1A CN113191343A (en) 2021-03-31 2021-03-31 Aviation wire identification code automatic identification method based on convolutional neural network

Publications (1)

Publication Number Publication Date
CN113191343A true CN113191343A (en) 2021-07-30

Family

ID=76974236

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110346523.1A Pending CN113191343A (en) 2021-03-31 2021-03-31 Aviation wire identification code automatic identification method based on convolutional neural network

Country Status (1)

Country Link
CN (1) CN113191343A (en)

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101571921A (en) * 2008-04-28 2009-11-04 富士通株式会社 Method and device for identifying key words
US20180137349A1 (en) * 2016-11-14 2018-05-17 Kodak Alaris Inc. System and method of character recognition using fully convolutional neural networks
CN109214382A (en) * 2018-07-16 2019-01-15 顺丰科技有限公司 A kind of billing information recognizer, equipment and storage medium based on CRNN
CN111382743A (en) * 2018-12-28 2020-07-07 上海大学 License plate character recognition method based on data enhancement and data generation
CN111539414A (en) * 2020-04-26 2020-08-14 梁华智能科技(上海)有限公司 OCR image character recognition and character correction method and system
CN111598089A (en) * 2020-05-16 2020-08-28 湖南大学 License plate correction and recognition method based on deep learning
CN111860590A (en) * 2020-06-15 2020-10-30 浙江大华技术股份有限公司 License plate voting method and device, computer equipment and storage medium
CN112446370A (en) * 2020-11-24 2021-03-05 东南大学 Method for recognizing text information of nameplate of power equipment
CN112463964A (en) * 2020-12-01 2021-03-09 科大讯飞股份有限公司 Text classification and model training method, device, equipment and storage medium

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101571921A (en) * 2008-04-28 2009-11-04 富士通株式会社 Method and device for identifying key words
US20180137349A1 (en) * 2016-11-14 2018-05-17 Kodak Alaris Inc. System and method of character recognition using fully convolutional neural networks
CN109214382A (en) * 2018-07-16 2019-01-15 顺丰科技有限公司 A kind of billing information recognizer, equipment and storage medium based on CRNN
CN111382743A (en) * 2018-12-28 2020-07-07 上海大学 License plate character recognition method based on data enhancement and data generation
CN111539414A (en) * 2020-04-26 2020-08-14 梁华智能科技(上海)有限公司 OCR image character recognition and character correction method and system
CN111598089A (en) * 2020-05-16 2020-08-28 湖南大学 License plate correction and recognition method based on deep learning
CN111860590A (en) * 2020-06-15 2020-10-30 浙江大华技术股份有限公司 License plate voting method and device, computer equipment and storage medium
CN112446370A (en) * 2020-11-24 2021-03-05 东南大学 Method for recognizing text information of nameplate of power equipment
CN112463964A (en) * 2020-12-01 2021-03-09 科大讯飞股份有限公司 Text classification and model training method, device, equipment and storage medium

Non-Patent Citations (6)

* Cited by examiner, † Cited by third party
Title
JUN LIU等: "Res-RNN Network and Its Application in Case Text Recognition", 《RSVT "19: PROCEEDINGS OF THE 2019 INTERNATIONAL CONFERENCE ON ROBOTICS SYSTEMS AND VEHICLE TECHNOLOGY》 *
YAN WAN等: "Robust Scene Text Recognition with Automatic Rectification Research on Scene Chinese Character Recognition Method Based on Similar Chinese Characters", 《2020 2ND INTERNATIONAL CONFERENCE ON MACHINE LEARNING, BIG DATA AND BUSINESS INTELLIGENCE (MLBDBI)》 *
崔循: "基于深度学习的集装箱箱号自动识别", 《中国优秀博硕士学位论文全文数据库(硕士) 信息科技辑》 *
朱皓: "视觉跟踪系统中基于人机交互的目标信息提取技术研究", 《中国优秀博硕士学位论文全文数据库(博士) 信息科技辑》 *
罗迤文: "基于深度学习的复杂场景下车牌识别系统", 《中国优秀博硕士学位论文全文数据库(硕士) 信息科技辑》 *
胡正平等: "先验采样约束结合扩展遮挡字典的细化稀疏人脸识别技术研究", 《信号处理》 *

Similar Documents

Publication Publication Date Title
CN108898137B (en) Natural image character recognition method and system based on deep neural network
US11809485B2 (en) Method for retrieving footprint images
WO2021042505A1 (en) Note generation method and apparatus based on character recognition technology, and computer device
CN112818951A (en) Ticket identification method
CN103279753B (en) A kind of English scene text block identifying method instructed based on tree construction
CN110929099A (en) Short video frame semantic extraction method and system based on multitask learning
CN113011253A (en) Face expression recognition method, device, equipment and storage medium based on ResNeXt network
CN111539417A (en) Text recognition training optimization method based on deep neural network
CN112686219B (en) Handwritten text recognition method and computer storage medium
CN109657682B (en) Electric energy representation number identification method based on deep neural network and multi-threshold soft segmentation
CN117115614B (en) Object identification method, device, equipment and storage medium for outdoor image
CN116524725B (en) Intelligent driving traffic sign image data identification system
CN112861840A (en) Complex scene character recognition method and system based on multi-feature fusion convolutional network
CN113191343A (en) Aviation wire identification code automatic identification method based on convolutional neural network
CN111242114A (en) Character recognition method and device
CN115116074A (en) Handwritten character recognition and model training method and device
CN116630369A (en) Unmanned aerial vehicle target tracking method based on space-time memory network
CN116152824A (en) Invoice information extraction method and system
CN115272242A (en) YOLOv 5-based optical remote sensing image target detection method
CN114611618A (en) Cross-modal retrieval-oriented data acquisition processing method and system
CN115147432A (en) First arrival picking method based on depth residual semantic segmentation network
CN114612907A (en) License plate recognition method and device
CN111126513A (en) Universal object real-time learning and recognition system and learning and recognition method thereof
CN111401356A (en) Express single-hand-written telephone number recognition method based on deep learning
CN111539952B (en) Scratch detection and result sharing method

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
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20210730