CN111274858A - Business license identification method in network transaction supervision - Google Patents
Business license identification method in network transaction supervision Download PDFInfo
- Publication number
- CN111274858A CN111274858A CN201811514261.XA CN201811514261A CN111274858A CN 111274858 A CN111274858 A CN 111274858A CN 201811514261 A CN201811514261 A CN 201811514261A CN 111274858 A CN111274858 A CN 111274858A
- Authority
- CN
- China
- Prior art keywords
- character
- network
- blocks
- watermark
- identification
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/40—Document-oriented image-based pattern recognition
-
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield 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
- G06N3/04—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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/14—Image acquisition
- G06V30/148—Segmentation of character regions
- G06V30/153—Segmentation of character regions using recognition of characters or words
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
Abstract
The invention discloses a business license identification method in network transaction supervision. The license identification includes: s1, removing watermarks by a generating type countermeasure network; s2, fixing the width to detect small character blocks; s3, removing the overlapped small character blocks with fixed width; s4, connecting the small character blocks to obtain character blocks and angles thereof, and removing overlapped character blocks; s5, oblique cutting is carried out on the character blocks, and the sizes are normalized; s6, character block feature extraction; s7, identifying a recurrent neural network; and S8, post-processing the identification result. The invention obtains the watermark information to automatically add the watermark, and the countermeasure type neural network trains the watermark removing model, thereby obtaining better effect on the problems that the character block can not be detected due to the watermark interference, the character recognition rate is low, and the like, and improving the accuracy rate of character recognition of pictures of business licenses.
Description
The invention discloses a business license identification method in network transaction supervision. The license identification includes: s1, removing watermarks by a generating type countermeasure network; s2, fixing the width to detect small character blocks; s3, removing the overlapped small character blocks; s4, connecting the small character blocks to obtain character blocks and angles thereof, and removing overlapped character blocks; s5, oblique cutting is carried out on the character blocks, and the sizes are normalized; s6, character block feature extraction; s7, identifying a recurrent neural network; and S8, post-processing the identification result. The invention obtains the watermark information to automatically add the watermark, and the countermeasure type neural network trains the watermark removing model, thereby obtaining better effect on the problems that the character block can not be detected due to the watermark interference, the character recognition rate is low, and the like, and improving the accuracy rate of character recognition of pictures of business licenses.
Technical Field
The invention relates to the technical field of internet, in particular to a picture watermark removing method and a business license identification method based on the picture watermark removing method.
Background
In OCR, for the image added with the watermark, the watermark is also a character, the character watermark in the image brings great interference to the detection and identification of subsequent character blocks, and the traditional business license identification mainly detects, cuts and identifies a single character. The method cannot eliminate the interference of the watermark, only can identify the text with clear background and more regular and simple characters, cannot adapt to the picture with complex background, and has poor adaptability to the picture of a business license containing the watermark. A method for removing the watermark based on priori knowledge and the like cannot obtain effective effects aiming at the position change, inclination and the like of the watermark.
Disclosure of Invention
The invention aims to provide a business license identification method in network transaction supervision.
In order to achieve the purpose, the invention adopts the following technical method:
the method for identifying the business license in the network transaction supervision is characterized by comprising the following steps:
s1, removing watermarks by a generating type countermeasure network;
s2, fixing the width to detect small character blocks;
s3, removing the overlapped small character blocks;
s4, connecting the small character blocks to obtain character blocks and angles thereof, and removing overlapped character blocks;
s5, oblique cutting is carried out on the character blocks, and the sizes are normalized;
s6, extracting character block characteristics;
s7, cyclic neural network character recognition;
and S8, post-processing the identification result.
Further, the dehydration model described in step S1 splits the generator into two sub-networks G ═ G1, G2 }: global generator network G1 and local enhancement network G2, where G1 consists of a convolution front-end, a series of residual blocks and a transposed convolution back-end, and G2 also consists of a convolution front-end, a series of residual blocks and a transposed convolution (sometimes called deconvolution) back-end. The input and output of the network are three-channel color pictures. Wherein training the dehydrated impression type comprises:
s11, obtaining watermark information, such as inclination, length-width ratio, color, space of watermarks in the picture and the like, and making a watermark picture template;
s12, collecting business license pictures without watermarks; establishing a confrontation type neural network watermark removing model, and modifying corresponding training parameters;
s13, adding a watermark picture template, setting information such as transparency, inclination and the like, and pasting the manufactured watermark template into a business license picture;
and S14, training a watermark removing model, executing a network to obtain an output picture, calculating a loss function value, and updating the network until the network is converged.
Further, in step S2, first, a convolutional neural network is used to extract picture features from the input picture, and a fixed number of window features are taken from each position of the feature map; adopting a bidirectional cyclic neural network as the characteristic identification of the sequence, inputting the characteristics corresponding to all windows of each row into the cyclic neural network, and obtaining an output category calibration characteristic vector; and inputting the category calibration feature vector into a classification and regression layer to obtain category information and position information. And mapping the recognition result on the feature map back to the original image to obtain the recognition result of the small blocks in the original image with the fixed width.
Further, in step S3, a non-maximum suppression method is used to filter out small text blocks with fixed width, whose overlap ratio is greater than a threshold value and whose confidence coefficient is smaller.
Further, step S4 obtains a set of text blocks with a fixed width by a simple text line construction algorithm, and obtains an angle of a text block and positions of four points of a quadrangle by fitting a straight line with a midpoint of each small block according to the midpoint set, the upper left point set, and the lower left point set.
Further, step S5 rotates the picture according to the four vertices and the angles of the text block, and captures the quadrilateral picture, and normalizes the quadrilateral picture into a picture with a fixed height and a width scaled by the height.
Further, step S6 obtains local information by convolution, and integrates the local information at a higher layer to obtain global information; and (3) carrying out dimensionality reduction operation by utilizing downsampling, processing the height of the character block into only 1 pixel, increasing the number of characteristic channels corresponding to the width, and expressing the characteristic channels as characteristic vectors.
Further, step S7 identifies the feature vector using the bidirectional recurrent neural network, obtains the probability distribution of each row of features, obtains the maximum value in the probability, and obtains the most likely identification result corresponding to the feature vector, which is also the identification result of the position.
Further, step S8 transcribes the probabilities into corresponding recognition characters, removing repeated consecutive identical characters, and obtaining the final result.
Compared with the background technology, the invention has the following advantages:
the invention automatically generates the watermark picture through analyzing the watermark, trains the watermark removing model, then detects and identifies the character of the watermark removing picture, eliminates the watermark interference and improves the accuracy of character detection and identification. The method has strong adaptability, can identify the watermark pictures of the business licenses of most of online platforms, and optimizes the efficiency of manual input.
Drawings
FIG. 1 is a flow chart of license identification;
FIG. 2 is a diagram of a photo watermarking network architecture;
FIG. 3 is a flow chart of text detection and recognition;
FIG. 4 is a drawing of a license and a watermark removal effect of the license;
FIG. 5 is a block detection and recognition effect picture;
Detailed Description
In order to make the objects, technical means and advantages of the present invention more apparent, the present invention will be further described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Examples
Referring to fig. 2 and 4, the method for removing watermark from a picture, which synthesizes a watermark picture, trains a watermark removal model to improve the detection and identification of license text blocks, includes the following steps:
and R1, acquiring watermark information, such as inclination, length-width ratio, color, space of watermarks in the picture and the like, and making a watermark picture template.
The watermark picture information is obtained through a business license watermark picture downloaded from a network platform, specific parameters are obtained through visual observation or by means of an image processing tool, random parameters are obtained through Gaussian distribution during actual use, and each pair of parameter values corresponds to one type of watermark.
And R2, collecting the business license pictures without the watermarks.
The business license pictures are collected through manual screening and actual shooting by using other browsers such as Baidu browsers, Google browsers and the like and using a web crawler to obtain a considerable part of pictures.
And R3, establishing a confrontation type neural network watermark removing model and modifying corresponding training parameters.
Image synthesis and processing at high resolution using a antagonistic neural network, splitting the generator into two sub-networks G ═ G1, G2 }: global generator network G1 and local enhancement network G2, where G1 consists of a convolution front-end, a series of residual blocks and a transposed convolution back-end, and G2 also consists of a convolution front-end, a series of residual blocks and a transposed convolution (sometimes called deconvolution) back-end. The input and output of the network are three channels of RGB color pictures.
And R4, adding the watermark picture template, setting information such as transparency, inclination and the like, and pasting the manufactured watermark template into a business license picture.
Adding a business license watermark picture synthesis module, adding a watermark template addition module in front of a network, setting information such as a watermark inclination angle, a width-height ratio, a horizontal distance, transparency and the like, sticking the watermark to the business license picture, and synthesizing the watermark picture.
R5. training the watermark removing model, executing the network to obtain the output picture, calculating the loss function value, and updating the network until the network is converged.
And adopting multi-scale discriminators which have the same network structure, introducing characteristic matching loss, and updating the network by calculating a loss function until convergence.
Referring to fig. 1, fig. 3 and fig. 5, the detection and identification of the text block of the watermark removed picture includes the following steps:
A1. and detecting the text block.
The text block detection specifically comprises:
A11. adopting CNN to extract features, and taking a fixed number of window features at each position on the feature map;
A12. adopting a bidirectional cyclic neural network as the characteristic identification of the sequence, inputting the characteristics corresponding to all windows of each row into the cyclic neural network, and obtaining an output category calibration characteristic vector;
A13. the category-specific feature vectors are input into the classification and regression layer. Acquiring category information and position information;
A14. and combining the fixed small rectangular boxes of the characters obtained by classification into a text line by using a text line construction algorithm.
A2. Character block feature extraction: extracting image features by using a standard CNN and expressing the image features into feature vectors;
adopting 5Conv +3 MaxPholing +2bn, obtaining local information by convolution, and integrating the local information at a higher layer to obtain global information; and (5) performing dimensionality reduction operation by using Pooling.
A3. Character block recognition: the bidirectional recurrent neural network obtains the probability of belonging to each category;
and (4) identifying the feature vectors by using the bidirectional LSTM to obtain the probability distribution of each row of feature identification results.
A4. And (4) performing recognition result post-processing to remove non-character and repeated recognition results.
And analyzing the identification result as hh e ll o, wherein the non-character is'll'. Removing repeated characters, and knowing that adjacent repeated results, such as hh, ll, cannot occur in normal recognition results, even if there are adjacent repeated results in the image, such as l in hello, there will be non-character separation between two l in the recognition results; removing non-characters, the final recognition result is hello.
The above description is only for the preferred embodiment 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 included in the scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.
Claims (9)
1. The method for identifying the business license in the network transaction supervision is characterized by comprising the following steps:
s1, removing watermarks by a generating type countermeasure network;
s2, fixing the width to detect small character blocks;
s3, removing the overlapped small character blocks with fixed width;
s4, connecting the small character blocks to obtain character blocks and angles thereof, and removing overlapped character blocks;
s5, obliquely cutting character blocks and normalizing the size of the picture;
s6, extracting character block characteristics;
s7, cyclic neural network character recognition;
and S8, post-processing the identification result.
2. A method of license identification in network transaction supervision as claimed in claim 1, wherein: in step S1, the watermarking model splits the generator into two sub-networks G ═ G1, G2: global generator network G1 and local enhancement network G2, where G1 consists of a convolution front-end, a series of residual blocks and a transposed convolution back-end, and G2 also consists of a convolution front-end, a series of residual blocks and a transposed convolution back-end. The input and output of the network are three-channel color pictures.
3. A method of license identification in network transaction supervision as claimed in claim 1, wherein: in step S2, first, a convolutional neural network is used to extract picture features from an input picture, and a fixed number of window features are taken at each position of a feature map; adopting a bidirectional cyclic neural network as the characteristic identification of the sequence, inputting the characteristics corresponding to all windows of each row into the cyclic neural network, and obtaining an output category calibration characteristic vector; and inputting the category calibration feature vector into a classification and regression layer to obtain category information and position information. And mapping the recognition result on the feature map back to the original image to obtain the recognition result of the small blocks in the original image with the fixed width.
4. A method of license identification in network transaction supervision as claimed in claim 1, wherein: in step S3, the non-maximum suppression method is used to filter out small text blocks with fixed width, whose overlap ratio is greater than the threshold and whose confidence is smaller.
5. A method of license identification in network transaction supervision as claimed in claim 1, wherein: in step S4, a set of text blocks with a fixed width is obtained by a simple text line construction algorithm, and a straight line is fitted according to the midpoint set, the upper left point set, and the lower left point set by using the midpoint of each small block to obtain the angle of the text block and the positions of four points in the quadrangle.
6. A method of license identification in network transaction supervision as claimed in claim 1, wherein: in step S5, the image is rotated according to the four vertices and angles of the text block, and the quadrilateral image is clipped and normalized into an image with a fixed height and a width scaled according to the height scaling ratio.
7. A method of license identification in network transaction supervision as claimed in claim 1, wherein: in step S6, local information is obtained by convolution, and local information is integrated at a higher layer to obtain global information; and (3) carrying out dimensionality reduction operation by utilizing downsampling, processing the height of the character block into only 1 pixel, increasing the number of characteristic channels corresponding to the width, and representing the characteristic channels as characteristic vectors.
8. A method of license identification in network transaction supervision as claimed in claim 1, wherein: in step S7, the feature vector is identified using the bidirectional recurrent neural network, the probability distribution of each row of features is obtained, the maximum value in the probability is obtained, and the maximum possible identification result corresponding to the feature vector is obtained, which is also the identification result of the position.
9. A method of license identification in network transaction supervision as claimed in claim 1, wherein: in step S8, the probability is transcribed into a corresponding recognition character, and repeated consecutive identical characters are removed.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811514261.XA CN111274858A (en) | 2018-12-04 | 2018-12-04 | Business license identification method in network transaction supervision |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811514261.XA CN111274858A (en) | 2018-12-04 | 2018-12-04 | Business license identification method in network transaction supervision |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111274858A true CN111274858A (en) | 2020-06-12 |
Family
ID=71001344
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811514261.XA Pending CN111274858A (en) | 2018-12-04 | 2018-12-04 | Business license identification method in network transaction supervision |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111274858A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112949523A (en) * | 2021-03-11 | 2021-06-11 | 兴业银行股份有限公司 | Method and system for extracting key information from identity card image picture |
| CN113239910A (en) * | 2021-07-12 | 2021-08-10 | 平安普惠企业管理有限公司 | Certificate identification method, device, equipment and storage medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180025256A1 (en) * | 2015-10-20 | 2018-01-25 | Tencent Technology (Shenzhen) Company Limited | Method and apparatus for recognizing character string in image |
| CN107993190A (en) * | 2017-11-14 | 2018-05-04 | 中国科学院自动化研究所 | Image watermark removal device |
| CN108399405A (en) * | 2017-02-07 | 2018-08-14 | 腾讯科技（上海）有限公司 | Business license recognition methods and device |
| CN108805789A (en) * | 2018-05-29 | 2018-11-13 | 厦门市美亚柏科信息股份有限公司 | A kind of method, apparatus, equipment and readable medium removing watermark based on confrontation neural network |
-
2018
- 2018-12-04 CN CN201811514261.XA patent/CN111274858A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180025256A1 (en) * | 2015-10-20 | 2018-01-25 | Tencent Technology (Shenzhen) Company Limited | Method and apparatus for recognizing character string in image |
| CN108399405A (en) * | 2017-02-07 | 2018-08-14 | 腾讯科技（上海）有限公司 | Business license recognition methods and device |
| CN107993190A (en) * | 2017-11-14 | 2018-05-04 | 中国科学院自动化研究所 | Image watermark removal device |
| CN108805789A (en) * | 2018-05-29 | 2018-11-13 | 厦门市美亚柏科信息股份有限公司 | A kind of method, apparatus, equipment and readable medium removing watermark based on confrontation neural network |
Non-Patent Citations (1)
| Title |
|---|
| 大熊吃土豆: "一种去水印的营业执照识别方法", Retrieved from the Internet <URL:https://blog.csdn.net/u013293750/article/details/80896420#comments_21659389> * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112949523A (en) * | 2021-03-11 | 2021-06-11 | 兴业银行股份有限公司 | Method and system for extracting key information from identity card image picture |
| CN113239910A (en) * | 2021-07-12 | 2021-08-10 | 平安普惠企业管理有限公司 | Certificate identification method, device, equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109154978B (en) | System and method for detecting plant diseases | |
| CN109583483B (en) | Target detection method and system based on convolutional neural network | |
| CN108596197B (en) | Seal matching method and device | |
| KR101617681B1 (en) | Text detection using multi-layer connected components with histograms | |
| CN103824091B (en) | A kind of licence plate recognition method for intelligent transportation system | |
| CN112819772B (en) | High-precision rapid pattern detection and recognition method | |
| CN107563377A (en) | It is a kind of to detect localization method using the certificate key area of edge and character area | |
| Shen et al. | Improving OCR performance with background image elimination | |
| CN111062931B (en) | Detection method of spliced and tampered image | |
| AU2014346263B2 (en) | Complex background-oriented optical character recognition method and device | |
| CN105787943B (en) | SAR image registration method based on multi-scale image block feature and rarefaction representation | |
| US20140301608A1 (en) | Chemical structure recognition tool | |
| CN109858438B (en) | Lane line detection method based on model fitting | |
| CN111274858A (en) | Business license identification method in network transaction supervision | |
| EP2994884A2 (en) | Method and apparatus for image matching | |
| CN112232249B (en) | Remote sensing image change detection method and device based on depth characteristics | |
| CN109086738B (en) | Character recognition method and device based on template matching | |
| CN114529715B (en) | Image identification method and system based on edge extraction | |
| Lee et al. | Backbone alignment and cascade tiny object detecting techniques for dolphin detection and classification | |
| CN114373178A (en) | Picture character detection and identification method and system | |
| Deepak et al. | Leaf detection application for android operating system | |
| Abraham et al. | A fuzzy based automatic bridge detection technique for satellite images | |
| Rani et al. | Object Detection in Natural Scene Images Using Thresholding Techniques | |
| Zhang et al. | Extraction and recognition of bridges over water in high resolution SAR image | |
| Yang et al. | A skeleton based binarization approach for video text recognition |
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 |