CN111626281B - Chinese annotation information identification method and system for paper image map based on adaptive learning - Google Patents
Chinese annotation information identification method and system for paper image map based on adaptive learning Download PDFInfo
- Publication number
- CN111626281B CN111626281B CN202010346076.5A CN202010346076A CN111626281B CN 111626281 B CN111626281 B CN 111626281B CN 202010346076 A CN202010346076 A CN 202010346076A CN 111626281 B CN111626281 B CN 111626281B
- Authority
- CN
- China
- Prior art keywords
- chinese
- information
- area
- map
- labeling information
- 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
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
- G06V10/225—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 based on a marking or identifier characterising the area
-
- 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
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/56—Extraction of image or video features relating to colour
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Biomedical Technology (AREA)
- Evolutionary Biology (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
- Character Input (AREA)
Abstract
The invention provides a method and a system for identifying Chinese marking information of a paper image map based on self-adaptive learning, wherein the method comprises the following steps of S1, preliminarily positioning a standard area; s2, initially positioning a standard area; s3, extracting the foreground of the Chinese annotation information; s4, identifying the Chinese marking information single character; s5, recognizing and filtering the Chinese marking information semantics; and S6, finally analyzing the Chinese marking information. According to the invention, through constructing a multi-level deep learning model comprising a universal Fast R-CNN rapid positioning model, and the self-adaptive accurate positioning method and the semantic filtering method provided by the invention, accurate extraction of Chinese identification information in the paper image map is realized, and finally, an electronic automatic processing flow of the paper map is realized.
Description
Technical Field
The invention relates to the field of image recognition and classification, in particular to a method and a system for recognizing Chinese annotation information of a paper image map based on self-adaptive learning.
Background
With the development of map making technology and geographic information technology in recent years, modern computer graphics technology has gradually replaced the traditional manual map making technology, and map products also realize diversified services based on electronic map products.
Some maps which are relatively long in the past or paper map products backed up by lost electronic maps cannot be directly and effectively superposed and analyzed with the basic data of the existing electronic maps, so that operators are often required to manually input the data of the paper maps into a computer for subsequent data analysis, use and storage, and the workload is complicated and the errors are easy to occur.
Therefore, how to effectively depend on the existing computer technology and artificial intelligence technology to realize the quick, highly automated and intelligent accurate identification and extraction of the paper map information has strong practical significance and application value.
Disclosure of Invention
The invention mainly solves the technical problems in the prior art, and provides a method and a system for identifying Chinese annotation information of a paper image map based on self-adaptive learning.
The technical problem of the invention is mainly solved by the following technical scheme:
a Chinese labeling information identification method for a paper image map based on self-adaptive learning comprises the following steps:
s1, primary positioning of a standard area: preliminarily positioning an area possibly containing Chinese labeling information in a paper image map, and selecting a discontinuous area in the paper image through a rectangular frame to identify the area;
s2, accurately positioning the marked area: on the basis of primarily selecting the Chinese identification information area in the step S1, further screening the accurate Chinese information identification area;
s3, extracting the foreground of the Chinese labeling information: on the basis of accurately screening out the Chinese identification information area in the step S2, segmenting the image and the Chinese identification information through a background modeling algorithm of a Gaussian mixture model, wherein the map image is taken as a background, and the Chinese identification information is taken as a foreground for extraction;
s4, identifying the Chinese labeling information single character: effective recognition of the single Chinese character with the foreground Chinese identification information extracted in the step S4 is realized by constructing massive Chinese training samples;
s5, semantic recognition and filtering of Chinese labeling information: filtering non-Chinese identification information which does not accord with Chinese semantic habits and is not in the database by constructing a map common Chinese character string database;
s6, final analysis of Chinese labeling information: and finally, analyzing and outputting the Chinese labeling information, and outputting a standard Chinese character string which accords with the Chinese semantic grammar.
Further, the substep of step S1 is:
1) Performing electronic processing on the paper map, namely scanning the paper map into an electronic map in high quality;
2) Filtering, namely filtering local information interference such as noise, fuzziness and the like caused by poor quality of paper images in the electronic image map;
3) The image color and contrast are adjusted by a color enhancement method, so that better identification and use effects are achieved;
4) Building a Tensorflow deep learning framework, training sample data of a standard Chinese character library, and realizing initial identification of Chinese characters in a map area by adopting a Fast R-CNN model;
5) Setting a super-parameter interface, manually setting and adjusting a threshold value according to the quality of a paper image, and identifying the area as a Chinese character area in a map if the threshold value is higher than the set threshold value, wherein the area is identified by a frame in the map;
6) And finally, the preliminary positioning of the Chinese standard information in the image map is realized.
Further, step S2 is implemented by using a ResNet neural network model, training samples are sampled from the target image and generated in real time, and sample labels are divided into two types: one is a Chinese identification information area and the other is a non-Chinese identification information area, wherein the first sample is produced by overlapping the non-Chinese area in the target image by randomly selecting Chinese characters in a standard Chinese character library; the second type sample directly selects a non-Chinese character area of the target area, and constructs a larger-scale training set to complete model training through random selection of Chinese characters and images in a standard library; and training the region generated in the step of preliminary positioning of the labeled region by the trained model to realize accurate positioning on the basis of preliminary positioning of the Chinese labeled region.
Further, the substep of step S3 is:
1) Constructing a background modeling frame of a Gaussian mixture model;
2) The separation of the image and the identification information in the image map is realized, the image is a background, and the identification information is extracted as a foreground;
3) Setting a super-parameter interface, manually setting and adjusting a threshold value according to the quality of a paper image, determining that the foreground extraction meets the requirement if the threshold value is higher than the set threshold value, stopping the extraction operation, and turning to the next step to continue the execution;
4) If the foreground extraction can not achieve the satisfactory effect all the time, the foreground extraction is achieved through a manual auxiliary mode, and the manual auxiliary method comprises the following steps: and manually drawing characters on the newly built map layer in the map, and storing the characters as the new map layer.
Further, the substep of step S4 is:
1) Building a Tensorflow deep learning framework, and configuring a RestNet neural network model training environment;
2) Configuring multifonts in a standard Chinese character library as training samples, wherein the multifonts comprise a Song style, an imitation Song style and a regular style;
3) Training a Chinese character recognition model through a RestNet neural network;
4) Identifying the foreground information extracted in the preorder step by using a Chinese character identification model generated by training;
5) And realizing the Chinese character recognition in the final foreground information.
Further, the substep of step S5 is:
1) Collecting and sorting a common Chinese character string database which comprises a place name address library, a map identification library and a place name library;
2) A multi-level index structure is constructed, so that rapid indexing and matching of various Chinese databases are facilitated;
3) Matching the recognition result of the single character recognition step of the Chinese labeling information with a Chinese character string database, and outputting a matching result, thereby filtering non-Chinese identification information which does not conform to Chinese semantic habits and is not in the database;
4) If the matching is successful, indicating that the identification is correct, and turning to the next step for processing; and if the matching fails, identifying the foreground area, and returning the result to the manual approval of the user.
A Chinese labeling information identification system of a paper image map based on self-adaptive learning comprises the following steps:
the Chinese labeling area positioning module is used for identifying an area with Chinese identification information on a paper image map to be identified for the paper image map to be identified;
the Chinese labeling information foreground extraction module is used for segmenting the image and the Chinese labeling information through a background modeling algorithm of a Gaussian mixture model on the basis that the Chinese labeling information area is accurately screened out by the Chinese labeling area positioning module 10, taking the map image as a background, and taking the Chinese labeling information as a foreground for extraction;
the Chinese labeling information single character recognition module is used for recognizing the foreground Chinese identification information extracted by the Chinese labeling information foreground extraction module and realizing the accurate recognition of the Chinese identification information by constructing massive Chinese training samples;
the Chinese marking information semantic recognition filtering module is used for filtering non-Chinese identification information which does not accord with Chinese semantic habits and is not in a database by constructing a map common Chinese character string database;
and the Chinese labeling information analysis module is used for analyzing and outputting the final Chinese labeling information and outputting a Chinese character string which is required to be standard and accords with the Chinese semantic grammar.
Furthermore, the Chinese labeling area positioning module comprises a Chinese labeling information preliminary positioning sub-module and a Chinese labeling information precise positioning sub-module,
the Chinese marking information preliminary positioning sub-module is used for preliminarily positioning an area where Chinese marking information possibly exists in a paper image map and selecting a discontinuous area in the paper image through a rectangular frame;
and the Chinese labeling information accurate positioning sub-module is used for eliminating the non-Chinese identification area in the Chinese labeling information area in the initial positioning through a machine learning algorithm.
Furthermore, the Chinese labeling information semantic recognition filtering module is based on the construction of a map common Chinese character string database, and realizes the semantic understanding and analysis of a character string consisting of single Chinese characters through similarity calculation, wherein a two-order threshold value is set for the Chinese labeling information character string, if the matching rate of a single character of the character string and the database is higher than the threshold value, further processing is carried out according to the condition of the threshold value, if the matching rate is higher than the second-order threshold value, the similarity is higher, and correction processing is directly carried out according to the correct character string in the keyword database; if the first-order threshold value is higher than the second-order threshold value, returning to the manual review of an operator; if the first-order threshold value is lower than the first-order threshold value, the identification error is directly calibrated, and the method returns to the preamble module for recalculation.
Further, the second-order threshold is set to 50%, and the first-order threshold is set to 30%.
The invention realizes the initial positioning of the area to be identified in the preliminary positioning step of the standard area by constructing a multi-level deep learning model, realizes the identification of Chinese marking information in the single character identification step of the Chinese marking information by a first-level deep learning model, and realizes the accurate extraction of the Chinese marking information in the paper image map and the electronic automatic processing flow of the paper map by the self-adaptive accurate positioning method and the semantic filtering method.
Drawings
FIG. 1 is a flow chart of the method for recognizing Chinese annotation information on a paper image map based on adaptive learning according to the present invention;
FIG. 2 is a block diagram of a system for recognizing Chinese annotation information on a paper image map based on adaptive learning according to the present invention.
Detailed Description
The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings.
In the embodiment of the invention, a paper map printed by high-resolution image data with the resolution of 1 m is taken as an example, the Chinese marking information comprises a place name, an address, important facilities and a building name, the standard Chinese character sample library comprises a Song body, an imitation Song GB2312, a regular body and a black body, the deep learning platform adopts Tensorflow2.0, and the system is realized in a software mode.
As shown in fig. 1: the method for identifying the Chinese marking information of the paper image map based on the self-adaptive learning comprises the following steps:
s1, initially positioning a standard area: in the paper image map, preliminarily positioning an area possibly containing Chinese labeling information, and selecting a discontinuous area in the paper image through a rectangular frame to identify the area, namely preliminarily positioning the rectangular area containing the Chinese labeling information from the paper image map to be extracted with the Chinese labeling information. Training sample data (including different fonts) of a standard Chinese character library by adopting a mature Fast R-CNN model, finally setting a threshold value, determining the sample data as Chinese labeling information if the sample data is higher than the threshold value, and immediately marking a rectangular area where the Chinese information is located by using a minimum external rectangle. The threshold value is set to 90% in this embodiment.
The sub-step of the preliminary positioning step of the marked region comprises the following steps:
1) Performing electronic processing on the paper map, namely scanning the paper map into an electronic map in high quality;
2) Filtering, namely filtering local information interference such as noise, fuzziness and the like caused by poor quality of paper images in the electronic image map;
3) The image color and the contrast are adjusted by a color enhancement method, so that better identification and use effects are achieved;
4) Building a Tensorflow deep learning framework, training sample data (including different fonts) of a standard Chinese character library, and realizing initial identification of Chinese characters in a map area by adopting a Fast R-CNN model;
5) Setting a super-parameter interface, manually setting and adjusting a threshold value according to the quality of a paper image, considering the area as a Chinese character area in a map if the threshold value is higher than the set threshold value, and marking the area as a box in the map;
6) And finally, the preliminary positioning of the Chinese standard information in the image map is realized.
S2, accurately positioning the marked area: on the basis of primarily selecting the Chinese identification information area in the step S1, the accurate Chinese information identification area is further screened, and the non-Chinese identification area is removed through a machine learning algorithm.
The step is realized by adopting a ResNet neural network model, training samples are sampled from target images and generated in real time, and sample labels are divided into two types: one is "Chinese identification information region", and the other is "non-Chinese identification information region". The first kind of sample is produced by overlapping non-Chinese character areas (preliminarily defined in the previous step) in a target image by adopting randomly selected Chinese characters in a standard Chinese character library; the second type sample directly selects the non-Chinese character area of the target area. Constructing a large-scale training set to finish model training through random selection of Chinese characters and images in a standard library; and training the region generated in the preliminary positioning step of the labeled region by the trained model to realize accurate positioning on the basis of the preliminary positioning of the Chinese labeled region.
S3, extracting the foreground of the Chinese annotation information: on the basis of accurately screening out the Chinese identification information area in the step S2, the image and the Chinese identification information are segmented through a background modeling algorithm of a Gaussian mixture model, wherein the map image is taken as a background, and the Chinese identification information is taken as a foreground for extraction.
The substep of extracting the Chinese labeling information foreground is as follows:
1) Constructing a background modeling frame of a Gaussian mixture model;
2) The separation of the image and the identification information in the image map is realized, the image is a background, and the identification information is extracted as a foreground;
3) Setting a super-parameter interface, manually setting and adjusting a threshold value according to the quality of a paper image, determining that the foreground extraction meets the requirement if the threshold value is higher than the set threshold value, stopping the extraction operation, and turning to the next step to continue the execution;
4) If the foreground extraction can not achieve the satisfactory effect all the time, the foreground extraction is achieved through a manual auxiliary mode, and the manual auxiliary method comprises the following steps: and manually drawing characters on the newly built map layer in the map, and storing the characters as the new map layer.
S4, identifying the Chinese labeling information single character: effective recognition of the single Chinese character with the foreground Chinese identification information extracted in the step S4 is realized by constructing massive Chinese training samples; in the step, a RestNet neural network model which is the same as that in the step S2 is adopted, and multifonts (Song style, imitation Song style, regular style and the like) in the standard Chinese character library in the step S1 are adopted as training samples. Unlike the step S1, the work object of the generated model trained in this step is a pure chinese information foreground with the background removed, and the chinese character is a standard font (non-handwritten font), so the recognition rate reaches 99.1% in this embodiment.
The substep of the Chinese labeling information single character recognition step is as follows:
1) Building a Tensorflow deep learning framework, and configuring a RestNet neural network model training environment;
2) Configuring multifonts (Song style, imitation Song style, regular script and the like) in a standard Chinese character library as training samples;
3) Training a Chinese character recognition model through a RestNet neural network;
4) Identifying the foreground information extracted in the preorder step by using a Chinese character identification model generated by training;
5) And realizing the Chinese character recognition in the final foreground information.
S5, semantic identification and filtering of Chinese labeling information: by constructing a map common Chinese character string database, filtering non-Chinese identification information which does not accord with Chinese semantic habits and is not in the database. The method mainly realizes the prospect that some ground objects with special geometric shapes are extracted by mistake as Chinese identification information and is judged mainly through semantic analysis and recognition; by constructing a map common Chinese character string database, the embodiment comprises a place name address, a common map identification name, a place name and the like of a selected area so as to judge whether a Chinese phrase and a sentence formed by a plurality of Chinese characters are correct or not, if so, the next step is carried out, and if not, the identification returns to manual approval.
The substep of the semantic recognition filtering step of the Chinese labeling information is as follows:
1) Collecting and sorting common Chinese databases including a place name address library, a map identification library and a place name library;
2) A multi-level index structure is constructed, so that rapid indexing and matching of various Chinese databases (including place name addresses, common map identification names, building names, company names and the like) are facilitated;
3) Matching the recognition result of the single character recognition step of the Chinese labeling information with a Chinese database, and outputting a matching result;
4) If the matching is successful, indicating that the identification is correct, and turning to the next step for processing; and if the matching fails, identifying the foreground area, and returning a result to the user for manual examination.
S6, final analysis of Chinese annotation information: and finally, analyzing and outputting the Chinese labeling information, and outputting the Chinese character string which is required to be standard and accords with the Chinese semantic grammar.
As shown in fig. 2: the embodiment of the invention provides a system for identifying Chinese labeling information of a paper image map based on self-adaptive learning, which comprises the following steps:
the Chinese labeling area positioning module 10 has a main function of identifying an area with Chinese identification information on a paper image map to be identified for the paper image map to be identified. For promoting recognition efficiency and precision, this embodiment will be identified the process and divide into two steps of thick location and accurate location, and with this correspondence, the module will be divided into two submodule pieces: a Chinese marking information preliminary positioning sub-module and a Chinese marking information accurate positioning sub-module.
The final output result of the Chinese labeling information primary positioning sub-module is a Chinese identification information area divided by the rectangular frame identification; the Chinese labeling information accurate positioning sub-module is used for further screening accurate Chinese information identification areas on the basis of primarily selecting the Chinese identification information areas, and non-Chinese identification areas are removed mainly through a machine learning algorithm.
The embodiment innovatively provides a self-adaptive training method based on the current map, which comprises the following steps: the method can be effectively applied to various factors such as the quality, the resolution, the color matching scheme, the degree of freshness and the like of the current map, and therefore the method has higher performance and positioning accuracy compared with a model trained by a traditional third party sample library.
The foreground extraction module 20 of the Chinese labeling information has the main function of segmenting the image and the Chinese identification information through a background modeling algorithm of a Gaussian mixture model, taking the map image as the background, and taking the Chinese identification information as the foreground for extraction, so as to further improve the accuracy of the subsequent single Chinese character recognition. The module adopts a Gaussian mixture model to realize background modeling, and then extracts foreground Chinese identification information to be output as a module.
Optionally, as for the output result of this step, because the quality of the paper image map is limited, the extracted foreground information may be missing or have other redundant information, which is represented as incomplete font of the extracted foreground image or having other speckled data redundancy. Under the condition, an image smoothing processing submodule and an image filtering submodule can be added into the module optionally, so that the quality of the extracted Chinese identification information data is improved.
The single character recognition module 30 of Chinese labeling information has the main function of recognizing foreground Chinese identification information extracted from the preamble module, and realizes the accurate recognition of the Chinese identification information by constructing massive Chinese training samples. The module adopts a ResNet classification model trained under a TensorFlow framework, and the sample set adopts Chinese characters with common fonts. In this embodiment, since all the chinese annotation information in the map data is in the standard font, the recognition rate is high.
Furthermore, the data of the module is foreground Chinese labeling information extracted by the preorder module, is limited by the image between the color of the Chinese labeling information and the color of the base image, and inevitably generates a small amount of redundancy and deficiency of the foreground extraction information. This case is divided into two treatment methods:
1) When the single character recognition model can uniquely determine the character through matching of the font part of the Chinese character, the determined character is taken as recognition output and is taken as input of a next functional module;
2) When key parts of foreground Chinese character identification information Chinese characters are missing and the Chinese characters cannot be accurately identified, the Chinese characters are specially marked, and a next step of function module is switched to, and the judgment and adjustment are carried out through context semantics.
The Chinese labeling information semantic recognition filtering module 40 has the main functions of realizing semantic understanding and analysis of character strings consisting of single Chinese characters through similarity calculation based on the construction of a map common Chinese character string database so as to further find recognition errors and improve recognition precision, namely filtering non-Chinese identification information which does not conform to Chinese semantic habits and is not in the database by constructing the map common Chinese character string database.
Furthermore, a two-step threshold value is set for the Chinese labeling information character string, if the matching rate of a single character of the character string and the database is higher than the threshold value, the situation shows that most Chinese characters in Chinese phrases and sentences are correctly identified, a few single characters are wrong, and at the moment, further processing is required according to the threshold value situation: if the similarity is higher than the second-order threshold value, indicating that the similarity is higher, directly performing correction processing according to correct character strings in the keyword database; if the first-order threshold value is higher than the second-order threshold value, returning to the manual review of an operator; if the first-order threshold value is lower than the first-order threshold value, the identification error is directly calibrated, and the method returns to the preamble module for recalculation. In this embodiment, since the chinese identification information in the paper map is mainly place name addresses, and the vast majority of the chinese identification information is place names with 2 characters or 3 characters, the second-order threshold is set to 50%, and the first-order threshold is set to 30%.
And a Chinese label information analysis module 50, which realizes the output of the finally extracted Chinese identification information and is connected with other external functional modules or databases.
Compared with the traditional manual data identification and extraction method, the method has the following advantages:
1) The manual identification is influenced by subjective main factors, and an evaluation standard of absolute consistency cannot be generated according to the age, the application, the quality and the like of the paper map, so that evaluation deviation is inevitably generated;
2) The map data volume is large, manual identification of all kinds of information of all areas in a browsing map which needs to be fully covered is time-consuming and labor-consuming, and the cost is high;
3) The manual identification is influenced by individuals, omission and error identification are easy to generate, and Chinese information without definite semantics such as place names, labels and the like is difficult to identify errors, so that the authenticity and the reliability of data are influenced finally.
Those of skill would further appreciate that the exemplary elements and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or combinations of both, and that the exemplary compositions and steps have been described above generally in terms of their functionality in order to clearly illustrate the interchangeability of hardware and software. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the implementation. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.
The steps of a method or algorithm described in connection with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module may reside in random access memory, read only memory, electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.
The above description is only for the specific embodiments of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention are also within the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims (10)
1. A paper image map Chinese annotation information identification method based on self-adaptive learning is characterized by comprising the following steps:
s1, primary positioning of a standard area: preliminarily positioning an area possibly containing Chinese marking information in a paper image map, and selecting a discontinuous area in the paper image through a rectangular frame to mark the area;
s2, accurately positioning a marked area: on the basis of primarily selecting the Chinese identification information area in the step S1, further screening the accurate Chinese information identification area;
s3, extracting the foreground of the Chinese annotation information: on the basis of accurately screening out the Chinese identification information area in the step S2, segmenting the image and the Chinese identification information through a background modeling algorithm of a Gaussian mixture model, wherein the map image is taken as a background, and the Chinese identification information is taken as a foreground for extraction;
s4, identifying the Chinese labeling information single character: effective recognition of the single Chinese character with the foreground Chinese identification information extracted in the step S4 is realized by constructing massive Chinese training samples;
s5, semantic recognition and filtering of Chinese labeling information: filtering non-Chinese identification information which does not accord with Chinese semantic habits and is not in the database by constructing a map common Chinese character string database;
s6, final analysis of Chinese labeling information: and finally, analyzing and outputting the Chinese labeling information, and outputting a standard Chinese character string which accords with the Chinese semantic grammar.
2. The method for recognizing Chinese labeling information on a paper image map based on adaptive learning as claimed in claim 1, wherein: the substeps of step S1 are:
1) E, electronization treatment of the paper image map: scanning a paper map into an electronic image map in high quality;
2) And (3) filtering treatment: filtering local information interference caused by poor paper image quality in the electronic image map;
3) Adjusting the color and contrast of the image by a color enhancement method;
4) Building a Tensorflow deep learning framework, training sample data of a standard Chinese character library, and realizing initial identification of Chinese characters in a map area by adopting a Fast R-CNN model;
5) Setting a super-parameter interface, manually setting and adjusting a threshold value according to the quality of a paper image, considering the area as a Chinese character area in a map if the threshold value is higher than the set threshold value, and marking the area as a frame in the map;
6) And finally, the preliminary positioning of the Chinese standard information in the image map is realized.
3. The method for recognizing Chinese labeling information on a paper image map based on adaptive learning as claimed in claim 1, wherein: step S2 is realized by adopting a ResNet neural network model, training samples are sampled from target images and generated in real time, and sample labels are divided into two types: one is a Chinese identification information area and the other is a non-Chinese identification information area, wherein the first sample is produced by overlapping the non-Chinese area in the target image by randomly selecting Chinese characters in a standard Chinese character library; the second type sample directly selects a non-Chinese character area of the target area, and constructs a larger-scale training set to complete model training through random selection of Chinese characters and images in a standard library; and training the region generated in the preliminary positioning step of the labeled region by the trained model to realize accurate positioning on the basis of the preliminary positioning of the Chinese labeled region.
4. The method for recognizing Chinese labeling information on a paper image map based on adaptive learning as claimed in claim 1, wherein: the substep of step S3 is:
1) Constructing a background modeling frame of a Gaussian mixture model;
2) The separation of the image and the identification information in the image map is realized, the image is a background, and the identification information is extracted as a foreground;
3) Setting a super-parameter interface, manually setting and adjusting a threshold value according to the quality of a paper image, determining that the foreground extraction meets the requirement if the threshold value is higher than the set threshold value, stopping the extraction operation, and turning to the next step to continue the execution;
4) If the foreground extraction can not achieve the satisfactory effect all the time, the foreground extraction is achieved through a manual auxiliary mode, and the manual auxiliary method comprises the following steps: manually drawing characters on a newly built layer in the map, and storing the characters as a new layer.
5. The method for recognizing Chinese labeling information on a paper image map based on adaptive learning as claimed in claim 1, wherein: the substep of step S4 is:
1) Building a Tensorflow deep learning framework, and configuring a RestNet neural network model training environment;
2) Configuring multifonts in a standard Chinese character library as training samples, wherein the multifonts comprise a Song style, an imitation Song style and a regular style;
3) Training a Chinese character recognition model through a RestNet neural network;
4) Identifying the foreground information extracted in the preorder step by using a Chinese character identification model generated by training;
5) And realizing the Chinese character recognition in the final foreground information.
6. The method for recognizing Chinese labeling information on a paper image map based on adaptive learning as claimed in claim 1, wherein: the substeps of step S5 are:
1) Collecting and sorting a common Chinese character string database which comprises a place name address library, a map identification library and a place name library;
2) A multi-level index structure is constructed, so that rapid indexing and matching of various Chinese databases are facilitated;
3) Matching the recognition result of the single character recognition step of the Chinese labeling information with a Chinese character string database, and outputting a matching result, thereby filtering non-Chinese identification information which does not conform to Chinese semantic habits and is not in the database;
4) If the matching is successful, indicating that the identification is correct, and turning to the next step for processing; and if the matching fails, identifying the foreground area, and returning the result to the manual approval of the user.
7. A Chinese annotation information identification system based on a paper image map of self-adaptive learning is characterized by comprising the following steps:
the Chinese marking area positioning module is used for identifying an area with Chinese identification information on a map of a paper image map to be identified;
the Chinese labeling information foreground extraction module is used for segmenting the image and the Chinese labeling information through a background modeling algorithm of a Gaussian mixture model on the basis that the Chinese labeling information area is accurately screened out by the Chinese labeling area positioning module, the map image is used as a background, and the Chinese labeling information is extracted as a foreground;
the Chinese labeling information single character recognition module is used for recognizing the foreground Chinese identification information extracted by the Chinese labeling information foreground extraction module and realizing the accurate recognition of the Chinese identification information by constructing massive Chinese training samples;
the Chinese marking information semantic recognition filtering module is used for filtering non-Chinese identification information which does not accord with Chinese semantic habits and is not in a database by constructing a map common Chinese character string database;
and the Chinese labeling information analysis module is used for analyzing and outputting the final Chinese labeling information and outputting a standard Chinese character string which accords with the Chinese semantic grammar.
8. The system for recognizing Chinese labeling information on a paper image map based on adaptive learning as claimed in claim 7, wherein: the Chinese labeling area positioning module comprises a Chinese labeling information preliminary positioning sub-module and a Chinese labeling information precise positioning sub-module,
the Chinese labeling information primary positioning sub-module is used for primarily positioning an area where Chinese labeling information possibly exists in a paper image map and selecting a discontinuous area in the paper image through a rectangular frame;
and the Chinese labeling information accurate positioning sub-module is used for eliminating the non-Chinese identification area in the Chinese labeling information area in the initial positioning through a machine learning algorithm.
9. The system for recognizing Chinese labeling information on a paper image map based on adaptive learning as claimed in claim 7, wherein: the Chinese labeling information semantic recognition filtering module is based on the construction of a map common Chinese character string database, and realizes the semantic understanding and analysis of a character string consisting of single Chinese characters through similarity calculation, wherein a two-order threshold value is set for the Chinese labeling information character string, if the matching rate of a single character of the character string and the database is higher than the threshold value, further processing is carried out according to the condition of the threshold value, and if the matching rate is higher than the second-order threshold value, the similarity is higher, correction processing is directly carried out according to the correct character string in the keyword database; if the first-order threshold value is higher than the second-order threshold value, returning to the manual review of an operator; if the first-order threshold value is lower than the first-order threshold value, the identification error is directly calibrated, and the method returns to the preamble module for recalculation.
10. The system for recognizing Chinese labeling information on a paper image map based on adaptive learning as claimed in claim 9, wherein: the second order threshold was set at 50% and the first order threshold at 30%.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010346076.5A CN111626281B (en) | 2020-04-27 | 2020-04-27 | Chinese annotation information identification method and system for paper image map based on adaptive learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010346076.5A CN111626281B (en) | 2020-04-27 | 2020-04-27 | Chinese annotation information identification method and system for paper image map based on adaptive learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111626281A CN111626281A (en) | 2020-09-04 |
| CN111626281B true CN111626281B (en) | 2022-12-02 |
Family
ID=72271760
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010346076.5A Active CN111626281B (en) | 2020-04-27 | 2020-04-27 | Chinese annotation information identification method and system for paper image map based on adaptive learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111626281B (en) |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104966097B (en) * | 2015-06-12 | 2019-01-18 | 成都数联铭品科技有限公司 | A kind of complex script recognition methods based on deep learning |
| CN109447055B (en) * | 2018-10-17 | 2022-05-03 | 中电万维信息技术有限责任公司 | OCR (optical character recognition) -based character similarity recognition method |
| CN109800761B (en) * | 2019-01-25 | 2022-11-11 | 厦门商集网络科技有限责任公司 | Method and terminal for creating paper document structured data based on deep learning model |
-
2020
- 2020-04-27 CN CN202010346076.5A patent/CN111626281B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN111626281A (en) | 2020-09-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109947836B (en) | English test paper structuring method and device | |
| KR100411697B1 (en) | Character reading method and address reading method | |
| KR100248917B1 (en) | Pattern recognizing apparatus and method | |
| CN109308476A (en) | Billing information processing method, system and computer readable storage medium | |
| CN110276077A (en) | The method, device and equipment of Chinese error correction | |
| CN111274239B (en) | Test paper structuring processing method, device and equipment | |
| US20050185841A1 (en) | Automatic document reading system for technical drawings | |
| CN114005123A (en) | System and method for digitally reconstructing layout of print form text | |
| CN112163424A (en) | Data labeling method, device, equipment and medium | |
| CN103902993A (en) | Document image identification method and device | |
| CN112766255A (en) | Optical character recognition method, device, equipment and storage medium | |
| CN113468317A (en) | Resume screening method, system, equipment and storage medium | |
| CN115311666A (en) | Image-text recognition method and device, computer equipment and storage medium | |
| CN111626281B (en) | Chinese annotation information identification method and system for paper image map based on adaptive learning | |
| CN111859956B (en) | Address word segmentation method for financial industry | |
| CN114581932A (en) | Picture table line extraction model construction method and picture table extraction method | |
| JP3394694B2 (en) | Format information registration method and OCR system | |
| CN117496521A (en) | Method, system and device for extracting key information of table and readable storage medium | |
| CN117274969A (en) | Seal identification method, device, equipment and medium | |
| CN117111890A (en) | Software requirement document analysis method, device and medium | |
| CN114564942B (en) | Text error correction method, storage medium and device for supervision field | |
| CN113627892B (en) | BOM data identification method and electronic equipment thereof | |
| JP3145071B2 (en) | Character recognition method and device | |
| CN118278385B (en) | Test method and device based on chapter rolling surface analysis and readable storage medium | |
| JPH08287188A (en) | Character string recognition device |
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 |