CN111898606B - Night imaging identification method for superimposing transparent time characters in video image - Google Patents
Night imaging identification method for superimposing transparent time characters in video image Download PDFInfo
- Publication number
- CN111898606B CN111898606B CN202010422410.0A CN202010422410A CN111898606B CN 111898606 B CN111898606 B CN 111898606B CN 202010422410 A CN202010422410 A CN 202010422410A CN 111898606 B CN111898606 B CN 111898606B
- Authority
- CN
- China
- Prior art keywords
- time
- characters
- night
- character
- transparent
- 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
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/14—Image acquisition
- G06V30/148—Segmentation of character regions
- G06V30/153—Segmentation of character regions using recognition of characters or words
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Character Discrimination (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to the field of computer vision, and relates to a method for identifying transparent time annotation information in a night video image, which comprises the following steps: improving the training process configuration of the CRNN; making a transparent time character and overlapping a night background image as a training input sample; a two-dimensional attention mechanism module is introduced to train a recognition model suitable for the characters at the transparent time at night; and outputting a time identification result from the image to be detected at night by using the identification model and carrying out logic verification. The method aims at the identification requirement of the transparent time marking information on the mainstream camera equipment, uses the night imaging environment, preliminarily improves the difference between the transparent time characters and the background image, further enhances the feature extraction capability of the transparent time characters, constructs a recognition model of a natural scene 'substrate-free' superposed transparent feature text, and realizes the transparent time character recognition method with the feature weighting extraction capability in the night imaging environment.
Description
Technical Field
The invention belongs to the field of computer vision, and can be used for detecting the overlapping time characters in pictures of video monitoring systems in public security and related industries. In particular to a method for identifying transparent time marking information in a night video image.
Background
In the design process of a recognition scheme for superimposing time characters in video monitoring images, it is found that some mainstream brands of cameras may adopt a stroke transparent (i.e. non-black and non-white gray) display style when superimposing characters, which brings new challenges to the character recognition of a 'substrate-free' background which faces many difficulties. Experimental data show that under the same natural scene, the recognition accuracy of the character strokes is reduced by more than 20% by using a transparent style of the character strokes than an opaque (pure black or pure white) style, and the expected application effect of the scheme is seriously influenced.
In order to obtain a more ideal effect, the recognition algorithm of overlapping time characters in an image based on a deep learning technology generally assumes a constraint condition of an application scene, namely, the overlapped characters are marked by adopting a standard, pure black or pure white and non-transparent style in the image. Because of the inherent low contrast property of the transparent characters, the transparent characters are easily confused with the natural background of the image, and become more difficult to identify, and the performance of the training model is directly influenced.
Through experimental data analysis, the recognition accuracy of stroke transparent characters is improved by only depending on algorithm training, and a bottleneck is met.
Disclosure of Invention
The invention aims to solve the technical problem of providing a method for identifying transparent time characters in night video images aiming at the identification requirement of stroke transparent time characters.
The basic technical concept of the invention is to improve the training process configuration of CRNN (a general text recognition neural network based on image sequences); making a transparent time character and overlapping a night background image as a training input sample; a two-dimensional attention mechanism module is introduced to train a recognition model suitable for the characters at the transparent time at night; and outputting a time identification result from the image to be detected at night by using the identification model and carrying out logic verification.
In order to solve the technical problem, the invention provides a method for recognizing transparent time characters in night imaging, which comprises the following steps of:
step i, improving a training method of the CRNN universal text recognition network;
step ii, making a CRNN night transparent time character training sample;
step iii, introducing an attention module to strengthen and train a night transparent time character recognition model;
and iv, identifying transparent time characters in the image to be detected at night and checking the reasonability of an output value.
Preferably, in the step i, the specific steps of improving the training method of the CRNN universal text recognition network include:
1-1) in a feature extraction link of a main ResNet classification network (a mainstream CNN convolution network), 3-layer sampling is adopted to retain more transverse features of characters;
1-2) when extracting the character height features, using maximum pooling;
1-3) selecting a single-layer LSTM (long-short time memory network), and using the output hidden layer vector for attention vector calculation.
Preferably, in the step ii, the specific step of making the CRNN night transparent time character training sample includes:
2-1) generating time characters with black and white colors, various fonts and various time formats on the background transparent image by using a character pixel rendering function;
2-2) generating random spots for the time character stroke pixels by using a PerlinNoise noise point diagram;
2-3) performing transparency processing on the time characters with random spots, and superposing the time characters on a random night background image to be used as an input sample for training;
2-4) taking the text form of the time characters in the step 2-1), and carrying out ignoring and forced replacing treatment on the characters in the non-standard time format to serve as a recognition target sample matched with the step 2-3);
preferably, the step iii of introducing the attention module to intensively train the night transparent time character recognition model specifically includes:
3-1) recording the matching probability vector of each time character feature decoded from the feature map and the target time character text by adopting a CTC coding system;
3-2) when the matching probability vector is decoded and output, the hidden layer vector inside the LSTM is superposed on the processed feature map to generate an attention weight map, and the feature of the time character at the corresponding position is enhanced;
3-3) in the attention module, reducing the attention weight map into an attention vector, and simply adding the attention vector and the matching probability vector to obtain a final target character matching probability vector;
and 3-4) obtaining the time character recognition model when the matching probability is integrally optimal.
Preferably, the step vi, the specific steps of identifying the time character in the image to be detected at night and checking the reasonability of the output value include:
4-1) collecting an image to be detected of the video monitoring equipment at night, and inputting the image to the identification model;
4-2) when the number of the output digital characters is not enough than the standard, judging that the output digital characters cannot be verified, and forcibly converting the output digital characters into a specific time value;
4-3) for recognizing that the number of the output numeric characters exceeds the standard, filling or deleting redundant time connection symbols '-and';
4-4) completing time segmentation and correction according to logic constraints of 'year, month and day' and 'hour, minute and second' and combining the matching probability recorded in 3-3), and converting into a final time value.
So far, the CRNN night character recognition model training is completed, the image to be detected at night is input, the time character recognition result is obtained, and the technical scheme of the invention is realized.
The beneficial effects of the invention include:
1) Aiming at the identification requirement of transparent time marking information on mainstream camera equipment, an identification model of a natural scene 'substrate-free' superposed transparent characteristic text is realized, and a special solution is formed.
2) The problem of feature learning of the transparent time characters is turned to a new idea of further improving the character and background distinguishing degree in a specific imaging mode, and a key technology is provided for the recognition requirements of similar application fields.
Drawings
The technical solution of the present invention will be further specifically described with reference to the accompanying drawings and the detailed description.
FIG. 1 is a basic flow diagram of the process of the present invention.
Fig. 2 is a training input sample of the CRNN network night recognition model.
FIG. 3 is a diagram illustrating the display effect of representative time characters in the night image to be examined.
FIG. 4 is a diagram of an attention-based mechanism for enhancing character feature extraction.
Fig. 5 is a time character format standardized example pattern of the recognition result, in which 5 (a) denotes ": "time character format of symbol is schematically illustrated, and 5 (b) represents deletion of redundant": schematic diagram of time character format of symbol.
FIG. 6 is an exemplary pattern of time character logic verification of recognition results.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
As shown in fig. 1, the present invention provides an overall flowchart of a method for recognizing transparent time characters in night imaging, which mainly comprises the following steps:
step i, self-defining a CRNN text recognition network structure;
step ii, making a night transparent time character training sample;
step iii, introducing an attention module to strengthen training and recognizing the model;
and iv, identifying transparent time characters in the image to be detected at night and checking the reasonability of an output value.
More specifically, step i comprises the following subdivision steps:
1-1) in a feature extraction link of a main ResNet classification network (a mainstream CNN convolution network), adopting 3-layer sampling setting to reserve more transverse features of characters;
in this embodiment, 3-layer sampling obtains a 64-pixel width feature map from an original image with a width of 512 pixels.
1-2) when extracting the character height characteristics, using maximum value pooling;
mf＝max({f 0,y |y∈{0,1,2,3}})
and selecting the largest feature from the 4 feature values in the height direction, thereby improving the character recognition accuracy.
1-3) selecting a single-layer LSTM (long-short time memory network), and using the output hidden layer vector for attention vector calculation.
Step ii comprises the following subdivision steps:
2-1) generating time characters with black and white colors, multiple fonts and multiple time formats on the background transparent image by using a character pixel rendering function;
Ω i ＝{0,1，...，K}
drawing K characters, where Ω i Representing the range of pixels to be drawn for the ith character (containing only the stroke pixels of that character), Ω 0 Representing background pixels other than character pixels.
I[x,y]Representing a pixel [ x, y ] in an image]RGB average luminance value of [ omega ] i Indicating the number of background pixels covered by the ith character, and D (i) indicating the average brightness of the covered background pixels of the ith character.
The rendering function for character generation is as follows:
using RGBA coding, [0,0,0,0] represents a colorless transparent pixel, [0,0,0,1] represents a black pixel, and [1,1,1,1] represents a white pixel.
2-2) generating random spots for the time character stroke pixels by using a PerlinNoise noise point diagram;
wherein eta is the generated noise point diagram, and eta [ x, y ] belongs to {0,1}.
The larger the attenuation control coefficient k > =0,k is, the more obvious the attenuation effect is;
k =8 in the present embodiment.
2-3) performing transparency processing on the time characters with random spots, and superposing the time characters to a random night background image to be used as an input sample for training;
transparency blending the temporal character pixel and the background pixel b using the formula:
f”＝(1.0-λα)οb+λf'
wherein alpha represents a character image with a median value of a transparency channel of f being 1, lambda is a mixing parameter, and the smaller lambda is, the more transparent the character in the output image is and the invisible the character is;
in the present embodiment λ =0.9,
as shown in fig. 2, the training sample time character strokes are transparent and have a defect effect, which simulates the situation that the transparent time character strokes of the image to be detected are mixed with background pixels in practical application as shown in fig. 3, typically as follows: '0', '3') the partial stroke area of '5'.
2-4) taking the text form of the time characters in the step 2-1), and performing neglect and forced replacement processing on the characters in the non-standard time format to serve as a recognition target sample matched with the step 2-3);
Γ＝{"0","1","2","3","4","5","6","7","8","9","-",":"}
Γ 1 = year and month
Γ 2 = time and minute
Γ 3 ＝U-Γ-Γ 1 -Γ 2
Γ denotes the set of supported standard characters, Γ 1 Indicating that a character set should be replaced with "-",Γ 2 indicating the character set, Γ, which should be replaced by 3 Representing all unsupported character sets, including chinese characters.
The character replacement formula is as follows:
s'＝replace(s,Γ 1 ,"-")
s”＝replace(s',Γ 2 ,":")
s”'＝delete(s”,Γ 3 )
step iii comprises the following subdivision steps:
3-1) recording the matching probability vector of each time character feature decoded from the feature map and the target time character text by adopting a CTC coding system;
p[i,c]
where i represents the ith character of the predicted string being output, c represents the probability that the ith character is predicted as character c, e.g., p [3, "8" ] represents the probability that the 3 rd character is predicted as character "8".
3-2) when the matching probability vector is decoded and output, the hidden layer vector inside the LSTM is superposed on the processed feature map to generate an attention weight map, and the feature of the time character at the corresponding position is enhanced;
wherein, W v 、
W h 、/>
Representing a linear transformation requiring learning, v representing a feature map, N i,j Eight adjacent fields near the (i, j) coordinate in the feature map, namely eight grid areas around the center grid in the nine-grid chart, h t Representing a hidden layer vector corresponding to the t character; a is i,j Is the attention vector corresponding to the (i, j) coordinate in the attention weight map, all a i,j The whole of the compositionBody is the attention weight graph a.
As shown in fig. 4, the processing portion within the dashed box.
3-3) in the attention module, reducing the attention weight map into an attention vector, and simply adding the attention vector and the matching probability vector to obtain a final target character matching probability vector;
as shown in FIG. 4, the Attention module outputs an Attention vector.
And 3-4) obtaining the time character recognition model when the matching probability is integrally optimal.
Step iv comprises the following subdivision steps:
4-1) collecting an image to be detected of the video monitoring equipment at night, and inputting the image to the identification model;
in this embodiment, the night period is 20 days — 04 days.
4-2) when the number of the recognized and output digital characters is less than that of the standard 14 (the number of digits in the standard format YYYY-MM-DD hh: MM: ss should reach 14), judging that the verification cannot be carried out, and forcibly converting the number into a specific time value;
identifying the image to be detected at night by using the model in 3-4), outputting a character set s,
z＝|delete(s,U-N)|
z represents the number of numeric characters output for recognition, where N represents the set of numeric characters and U-N is the set of all characters except the number.
When z <14, it is determined that verification is impossible, and the time is converted into a specific time value, and the present embodiment is collectively designated as time information "2000-01-01" 00 and outputs.
4-3) connecting symbols 'and' are added or deleted for the time reaching more than 14 numbers to be output;
s'＝pre(s,c dash ,2)
the redundant "-" connectors are deleted, where the pre function indicates that the character is deleted, leaving only the top 2 designated characters.
Padding a "-" connector, where ist indicates inserting a character in a designated sequence number of a string s, c dash Representing a "-" character, the d (c) function determines whether the character c is a number.
In the same way, the method for preparing the composite material,
s'＝re(pre(re(s),c colon ,2))
deleting redundant ': characters, wherein the re function represents that the character string is reversed in order, the character string is reversed in order and then restored, namely the last two': characters can be reserved, and c colon Representing a ": character.
m＝re(s)
s'＝re(n)
And (4) completing the ": connector, wherein the two times of reverse operations are completed by taking the last": as a reference.
As shown in FIG. 5, 5 (a) indicates that after the 'number is filled, 5 (b) indicates that the' number is deleted.
4-4) according to logic constraints of 'year, month and day' and 'hour, minute and second' and combining the matching probability obtained in 3-3), finishing time segmentation and correction, and converting into a final time value.
N 1 ∈{c 1 |(10c 1 +c 2 )∈[a,b],c 1 ,c 2 ∈{0,1,...,9}}
A correction method of two-digit numbers, wherein a and b represent the range of two-digit numbers, c 1 c 2 Representing the number in ten and one of the current prediction, p [ i, c 1 ]Indicates that the ith character is predicted to be c 1 The correction process is performed on this reasonable number set N 1 And (5) finding the predicted character with the highest probability as a correction result again. When c is going to 1 After the correction is finished, the correction is carried out by
For the correct ten digits, a set of reasonable numbers N is determined 2 Then c is carried out 2 And (4) correcting.
After verification, the correct time value is finally output, as shown in fig. 6.
So far, training of a night character recognition model is completed, a night image to be detected is input, a time character recognition result is obtained, and the technical scheme of the invention is realized.
In this embodiment, 8215 nighttime random image recognition results are counted, and 91.86% accuracy (full matching rate) is obtained.
The method achieves ideal application effect.
It will be clear to those skilled in the art that the specific values of the above parameters or thresholds can be adjusted according to the strictness of the sample training method and the specification implementation, and do not limit the present invention.
Finally, it should be noted that the above-mentioned embodiments are only preferred embodiments of the present invention, and not intended to limit the present invention, and although the present invention has been described in detail with reference to the foregoing embodiments, it will be apparent to those skilled in the art that modifications and equivalents can be made in the technical solutions described in the foregoing embodiments, or some technical features of the present invention may be substituted. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (3)
1. A character recognition method for superimposing transparent time characters in a night image is characterized by comprising the following steps:
step i, improving a training method of the CRNN universal text recognition network;
step ii, making a CRNN night image transparent time character training sample;
step iii, introducing an attention module to strengthen and train a night image transparent time character recognition model;
step iv, identifying transparent time characters in the night image to be detected and checking the reasonability of an output value;
in the step i, the specific steps of the training method for improving the CRNN universal text recognition network include:
1-1) in the feature extraction link of a main ResNet classification network, adopting 3-layer sampling to reserve more transverse features of characters;
1-2) when extracting the character height features, using maximum pooling;
1-3) selecting a single-layer LSTM, and using the output hidden layer vector thereof for attention vector calculation;
step iii, the specific steps of the attention module for strengthening training of the night image transparent time character recognition model include:
3-1) recording the matching probability vector of each time character feature decoded from the feature map and the target time character text by adopting a CTC coding system;
3-2) when the matching probability vector is decoded and output, the hidden layer vector inside the LSTM is superposed on the processed feature map to generate an attention weight map, and the feature of the time character at the corresponding position is enhanced;
3-3) in the attention module, reducing the attention weight map into an attention vector, and adding the attention vector and the matching probability vector to obtain a final target character matching probability vector;
and 3-4) obtaining the time character recognition model when the matching probability is integrally optimal.
2. The method for character recognition by superimposing a transparent time character on a night image according to claim 1, wherein the step ii of preparing the CRNN night image transparent time character training sample comprises the steps of:
2-1) generating time characters with black and white colors, multiple fonts and multiple time formats on the background transparent image by using a character pixel rendering function;
2-2) generating random spots for the time character stroke pixels by using a PerlinNoise noise point diagram;
2-3) performing transparency processing on the time characters with random spots, and superposing the time characters to a randomly generated night background image to be used as an input sample of training;
2-4) taking the text form of the time characters in the step 2-1), and performing neglect and forced replacement processing on the time characters in the non-standard time format to serve as recognition target samples matched with the step 2-3).
3. The method for recognizing characters by superimposing transparent time characters on night images according to claim 2, wherein the step iv, the specific steps of recognizing the time characters in the image to be detected at night and checking the reasonableness of the output value, comprises:
4-1) collecting an image to be detected of the video monitoring equipment at night, and inputting the image to the identification model;
4-2) when the number of the output digital characters is not enough than the standard, judging that the output digital characters cannot be verified, and forcibly converting the output digital characters into a specific time value;
4-3) for recognizing that the number of the output numeric characters exceeds the standard, filling or deleting redundant time connection symbols '-and';
4-4) completing time segmentation and correction according to logic constraints of 'year, month and day' and 'hour, minute and second' and combining the matching probability recorded in 3-3), and converting into a final time value.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010422410.0A CN111898606B (en) | 2020-05-19 | 2020-05-19 | Night imaging identification method for superimposing transparent time characters in video image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010422410.0A CN111898606B (en) | 2020-05-19 | 2020-05-19 | Night imaging identification method for superimposing transparent time characters in video image |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111898606A CN111898606A (en) | 2020-11-06 |
| CN111898606B true CN111898606B (en) | 2023-04-07 |
Family
ID=73207553
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010422410.0A Active CN111898606B (en) | 2020-05-19 | 2020-05-19 | Night imaging identification method for superimposing transparent time characters in video image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111898606B (en) |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112381057A (en) * | 2020-12-03 | 2021-02-19 | 上海芯翌智能科技有限公司 | Handwritten character recognition method and device, storage medium and terminal |
| CN113157322A (en) * | 2020-12-29 | 2021-07-23 | 深圳微步信息股份有限公司 | Method and system for automatically displaying software compiling time |
| CN113065550B (en) * | 2021-03-12 | 2022-11-11 | 国网河北省电力有限公司 | Text recognition method based on self-attention mechanism |
| CN113065561A (en) * | 2021-03-15 | 2021-07-02 | 国网河北省电力有限公司 | Scene text recognition method based on fine character segmentation |
| CN113723337B (en) * | 2021-09-07 | 2024-09-24 | 武汉东智科技股份有限公司 | Monitoring image place information identification method based on DDT deep nerve model structure |
| CN113869426B (en) * | 2021-09-29 | 2024-07-26 | 北京搜狗科技发展有限公司 | Formula identification method and device |
| CN113902804B (en) * | 2021-10-14 | 2024-10-25 | 北京达佳互联信息技术有限公司 | Method, device, equipment and medium for identifying transparent area in image |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102663675A (en) * | 2012-03-28 | 2012-09-12 | 广东威创视讯科技股份有限公司 | Method and apparatus for scaling image processed by character superimposition |
| CN109344831A (en) * | 2018-08-22 | 2019-02-15 | 中国平安人寿保险股份有限公司 | A kind of tables of data recognition methods, device and terminal device |
| CN110378334A (en) * | 2019-06-14 | 2019-10-25 | 华南理工大学 | A kind of natural scene text recognition method based on two dimensional character attention mechanism |
| CN110533030A (en) * | 2019-08-19 | 2019-12-03 | 三峡大学 | Sun film image timestamp information extracting method based on deep learning |
| CN110781881A (en) * | 2019-09-10 | 2020-02-11 | 腾讯科技（深圳）有限公司 | Method, device, equipment and storage medium for identifying match scores in video |
| CN111027562A (en) * | 2019-12-06 | 2020-04-17 | 中电健康云科技有限公司 | Optical character recognition method based on multi-scale CNN and RNN combined with attention mechanism |
-
2020
- 2020-05-19 CN CN202010422410.0A patent/CN111898606B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102663675A (en) * | 2012-03-28 | 2012-09-12 | 广东威创视讯科技股份有限公司 | Method and apparatus for scaling image processed by character superimposition |
| CN109344831A (en) * | 2018-08-22 | 2019-02-15 | 中国平安人寿保险股份有限公司 | A kind of tables of data recognition methods, device and terminal device |
| CN110378334A (en) * | 2019-06-14 | 2019-10-25 | 华南理工大学 | A kind of natural scene text recognition method based on two dimensional character attention mechanism |
| CN110533030A (en) * | 2019-08-19 | 2019-12-03 | 三峡大学 | Sun film image timestamp information extracting method based on deep learning |
| CN110781881A (en) * | 2019-09-10 | 2020-02-11 | 腾讯科技（深圳）有限公司 | Method, device, equipment and storage medium for identifying match scores in video |
| CN111027562A (en) * | 2019-12-06 | 2020-04-17 | 中电健康云科技有限公司 | Optical character recognition method based on multi-scale CNN and RNN combined with attention mechanism |
Non-Patent Citations (4)
| Title |
|---|
| Improving CRNN with EfficientNet-like feature extractor and multi-head attention for text recognition;Dinh Viet Sang 等;《SoICT 2019》;20191206;全文 * |
| Natural Scene Text Recognition Based on Encoder-Decoder Framework;LING-QUN ZUO 等;《IEEE Access》;20190514;全文 * |
| 基于Attention-CTC的自然场景文本识别算法;和文杰 等;《电子科技》;20191215;第32卷(第12期);全文 * |
| 基于注意力增强网络的场景文字识别;徐富勇;《现代计算机》;20200325(第09期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111898606A (en) | 2020-11-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111898606B (en) | Night imaging identification method for superimposing transparent time characters in video image | |
| CN111028308B (en) | Steganography and reading method for information in image | |
| CN109472270A (en) | Image style conversion method, device and equipment | |
| CN110866879B (en) | Image rain removing method based on multi-density rain print perception | |
| CN116052016A (en) | Fine segmentation detection method for remote sensing image cloud and cloud shadow based on deep learning | |
| CN112084859B (en) | Building segmentation method based on dense boundary blocks and attention mechanism | |
| CN111738113A (en) | Road extraction method of high-resolution remote sensing image based on double-attention machine system and semantic constraint | |
| CN115376024A (en) | Semantic segmentation method for power accessory of power transmission line | |
| CN113449691A (en) | Human shape recognition system and method based on non-local attention mechanism | |
| CN101950405B (en) | Watermark adding method based on video content | |
| CN114677722A (en) | Multi-supervision human face in-vivo detection method integrating multi-scale features | |
| CN111696021A (en) | Image self-adaptive steganalysis system and method based on significance detection | |
| CN115410081A (en) | Multi-scale aggregated cloud and cloud shadow identification method, system, equipment and storage medium | |
| CN111899202B (en) | Enhancement method for superimposed time character in video image | |
| CN116682082A (en) | Vehicle digital twin method suitable for automatic driving scene | |
| CN116703725A (en) | Method for realizing super resolution for real world text image by double branch network for sensing multiple characteristics | |
| CN116152173A (en) | Image tampering detection positioning method and device | |
| CN114202775A (en) | Transformer substation dangerous area pedestrian intrusion detection method and system based on infrared image | |
| CN112950599B (en) | Large intestine cavity area and intestine content labeling method based on deep learning | |
| CN116310376A (en) | Double-light fusion method, device and equipment based on multiscale regional attention mechanism | |
| CN115482463A (en) | Method and system for identifying land cover of mine area of generated confrontation network | |
| CN114943204A (en) | Chinese character font synthesis method based on generation countermeasure network | |
| CN113781376A (en) | High-definition face attribute editing method based on divide-and-conquer fusion | |
| CN110781884A (en) | Method for realizing intelligent reading of electric meter data | |
| CN115761444B (en) | Training method of incomplete information target recognition model and target recognition 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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| PE01 | Entry into force of the registration of the contract for pledge of patent right | ||
| PE01 | Entry into force of the registration of the contract for pledge of patent right |
Denomination of invention: A Nighttime Imaging Recognition Method for Overlaying Transparent Time Characters in Video Images Effective date of registration: 20230615 Granted publication date: 20230407 Pledgee: Industrial Bank Limited by Share Ltd. Wuhan branch Pledgor: WUHAN EASTWIT TECHNOLOGY CO.,LTD. Registration number: Y2023420000235 |