CN106599892A - Television station logo identification system based on deep learning - Google Patents
Television station logo identification system based on deep learning Download PDFInfo
- Publication number
- CN106599892A CN106599892A CN201611151654.XA CN201611151654A CN106599892A CN 106599892 A CN106599892 A CN 106599892A CN 201611151654 A CN201611151654 A CN 201611151654A CN 106599892 A CN106599892 A CN 106599892A
- Authority
- CN
- China
- Prior art keywords
- station symbol
- sample
- station
- deep learning
- system based
- 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
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
- G06V10/26—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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/09—Recognition of logos
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to a computer vision technology, and discloses a television station logo identification system based on deep learning to improve the ability of station logo identification. The system comprises a sample collection module used for collecting station logo samples, a sample screening module used for screening samples to remove inappropriate samples, a station logo segmentation module used for segmenting a station logo from a background, a sample synthesis module used for artificially synthesizing different samples, a model training module used for training a station logo identification model based on the collected samples and the artificially synthesized samples, and a station logo identification module used for identifying station logos based on the identification model. The system of the invention is suitable for identifying normal station logos, rebroadcast station logos and extremely similar station logos.
Description
Technical field
The present invention relates to computer vision technique, more particularly to the TV TV station symbol recognition system based on deep learning.
Background technology
With flourishing for broadcast medium, radio and television have penetrated into the various aspects of daily life and work.Consider
To important function of television station's station symbol in terms of television station is distinguished, realize that Computer Automatic Recognition television station station symbol has great
Meaning.First, television station's station symbol is the distinct mark that many television stations and other television stations are distinguished, and can be used to protect its business profit
Benefit, and TV signal whether occur illegally intercutting, break or blank screen important embodiment;2nd, television station's TV station symbol recognition can be used for
The quick-searching of wired pay channel;3rd, contribute to image processing tool removing for station symbol, so as to improve the matter of video
Amount.4th, relevant information can pointedly be recommended to user according to the other result of TV station identification by television operator or businessman,
Such as in TV programme emerging commodity or service etc.;5th, television station's TV station symbol recognition can be detected by satellite transmission
The signal stabilization of the cable television of signal, it is ensured that broadcast TV program is not disturbed by illegal signals, especially external illegal group
Attack microwave tower int the sky is knitted, illegal reaction signal etc. is intercutted.
The big multipair situations below discrimination of existing TV station symbol recognition method is low or does not just account for situations below at all:
(1) Similar color and transparent station symbol are difficult;(2) extremely similar provincial station symbol, station symbol geometry are the same and only Arabic
The different station symbol of numeral, such as Sichuan 2-7 platforms;(3) station symbol relayed overlaps, and such as Sichuan satellite TV relays CCTV1.Cause
This, we need to design a kind of method and system of efficient, high-accuracy and TV station symbol are identified.
The content of the invention
The technical problem to be solved in the present invention is:A kind of TV TV station symbol recognition system based on deep learning is provided, is improved
The identification ability of station symbol.
To solve the above problems, the technical solution used in the present invention is:Based on the TV TV station symbol recognition system of deep learning,
Including:
Sample collection module, for picking platform standard specimen sheet;
Screening sample module, for screening sample, rejects inappropriate sample;
Station symbol splits module, for splitting station symbol and background;
Sample synthesis module, for the sample of synthetic diversity;
Model training module, the sample of sample and synthetic based on collection, training station target identification model;
TV station symbol recognition module, recognizes station symbol based on identification model.
Further, when acquisition module carries out sample collection, possesses certain time interval.
Further, inappropriate sample is:The same frame that collected due to network interim card and station symbol serious deformation
Platform sample and the sample without station symbol caused due to zapping.
Further, the step of splitting station symbol and background includes:
The station symbol image of a black background is selected first, according to five points of 1/3rd and high-order artwork of a width of artwork
One of intercept the upper left corner subimage, by obtain subimage adopt two value filterings, eliminate noise, obtain station symbol foreground image,
Finally both horizontally and vertically project respectively, be partitioned into station symbol itself.
Further, the method for synthetic sample includes
Fixed channel logo position synthetic method:Intercepting a fixed size first does not have the background image of station symbol;Secondly select
Station symbol foreground image;Finally image overlay by two kinds together;
Any channel logo position synthetic method:Intercepting a fixed size first does not have the background image of station symbol;Secondly select
Station symbol image itself, and determining table target size;Finally in random selection one above background image as station symbol size
Region, and this region and station symbol are superimposed.
Further, the mode of superposition is:Pixel position not for 0 in station symbol foreground image is recorded first;Secondly will the back of the body
In scape image, the pixel value of these positions is set to the pixel value of these positions of station symbol foreground image.
Further, Gaussian smoothing marginal information is adopted to the sample of synthetic.
Further, TV station symbol recognition module adopts CNN algorithms as TV station symbol recognition algorithm;Wherein, CNN model structures include
Four layers of convolutional layer and two-layer are linked entirely, and have activation primitive and max-pooling layers after convolutional layer.
Further, activation primitive is Max-Feature-Map, and form is as follows:
Wherein, C ∈ Rh×w×2n, the quantity of C is 2n, and 2n represents the quantity of representative feature collection of illustrative plates;The span of k is 1-n;
The real number set that R is represented;H represents the height of convolution characteristic spectrum;W represents the width of convolution characteristic spectrum;I represents convolution feature
The position in the y-axis direction in collection of illustrative plates, span is 1-h;J represents the position in the x-axis direction in convolution characteristic spectrum, value model
Enclose for 1-w.
Further, the convolution kernel and stride (i.e. the spaced pixels number of convolution every time) of four layers of convolutional layer is respectively:
9X9X96/1、5X5X192/1、5X5X256/1、4X4X384/1;All of max-pooling cores and stride are 2x2/2;
The number of first neuron for connecting full layer is 256;Depending on the number of the neuron of second full articulamentum is according to station symbol quantity.
The invention has the beneficial effects as follows:The present invention can high real-time and high-accuracy solves the problems, such as TV station symbol recognition, including
Normal station symbol (such as Sichuan satellite TV), relay station symbol (as Sichuan satellite TV relays CCTV1), extremely similar station symbol (such as Sichuan 2-7 platforms);
The model for simultaneously optimizing deep learning solves the problems, such as deep learning algorithm terminal operating efficiency is low, occupancy cpu resource is high
Problem.
Description of the drawings
Fig. 1 is the training mode flow process of system;
Fig. 2 is station symbol segmentation flow chart;
Fig. 3 is the design sketch after station symbol segmentation;
Fig. 4 is artificial sample synthesis flow;
Fig. 5 (a) is any channel logo position synthetic effect figure;
Fig. 5 (b) is to fix channel logo position synthetic effect figure;
Fig. 6 is model training flow chart.
Specific embodiment
In order that the object, technical solutions and advantages of the present invention are clearer, below implementation steps of the present invention are carried out
It is further to describe in detail.
According to system need set up all modules of TV TV station symbol recognition system, i.e. sample collection module, screening sample
Module, station symbol segmentation module, sample synthesis module, model training module and TV station symbol recognition module.Below according to the work of system
Pattern is illustrating the workflow of this system.
(1) training mode of the invention
Training flow process is as shown in figure 1, concrete steps are described as follows:
(1) sample collection
Sample collection module is for each a number of sample of collection, and has between the regular hour between every sample
Every.As the video flowing of TV programme has certain seriality and dependency, if be not spaced or time interval not enough,
Between the sample of collection, similarity is high, even if the help that the sufficient training to following model of sample is provided is also little.
(2) screening sample
Screening sample module reject the platform sample of the same frame and station symbol serious deformation collected due to network interim card with
And the sample without station symbol caused due to zapping, retain the sample of station symbol normal sample and station symbol somewhat deformation, to increase sample
Multiformity.
(3) split station symbol
As shown in Fig. 2 the flow process of the station symbol segmentation of station symbol segmentation module:The station symbol figure of a black background is selected first
Picture, according to the subimage in 1/5th intercepting upper left corners of 1/3rd and high-order artwork of a width of artwork.Experiment is proved so
Intercepting method be rational:TV station symbol is nearly all in the upper left corner first;Next to that can intercept according to this ratio
To complete station symbol;It is finally that this method can bluntly obtain the image containing station symbol, without with special calculation
Method obtains station symbol, so as to simplify whole flow process.The subimage for obtaining is adopted into two value filterings, noise is eliminated, station symbol prospect is obtained
Image.Finally both horizontally and vertically project respectively, be partitioned into station symbol itself, the effect after station symbol segmentation is as schemed
Show.
(4) sample synthesis
Due to the position of station symbol it is substantially changeless, even if changing, and small range.For such case,
The mode of sample synthesis module synthesis sample has two kinds:One kind is to fix channel logo position, and another is any channel logo position.People
Work sample synthesis flow is as shown in Figure 4.
Fixed channel logo position synthetic method:Intercepting a fixed size first does not have the background image of station symbol;Secondly select
Station symbol foreground image;Finally image overlay by two kinds together.The mode of superposition:Picture in station symbol foreground image is recorded first
Element position not for 0;Secondly the pixel value of these positions in background image is set to the picture of these positions of station symbol foreground image
Element value;Gaussian smoothing marginal information is finally adopted, more natural sample is obtained.Station symbol in the artificial sample for so obtaining
Position is consistent with the channel logo position in station symbol foreground image, such as shown in Fig. 5 (b).
Any channel logo position synthetic method:Intercepting a fixed size first does not have the background image of station symbol;Secondly select
Station symbol image itself, and determining table target size;Finally in random selection one above background image as station symbol size
Region, this region and station symbol are superimposed.Stacked system and the side of superposition in fixed channel logo position synthetic method
Formula is consistent, such as shown in Fig. 5 (a).
(5) model training
In recent years, deep learning has all significantly surmounted traditional algorithm image recognition, speech recognition etc. are multi-field, and which is special
Levy with very high identification.And field of image recognition is directed to, the effect of convolutional neural networks (CNN) is best, therefore we will
CNN algorithms are used as TV station symbol recognition algorithm.Station symbol image to needing identification need not do any early stage pretreatment, be sent directly into
Recognize in CNN networks, not only omit unnecessary human intervention process, while the raw information of the image for retaining, carries
The high generalization of model.CNN model structures employ four layers of convolutional layer and two-layer is linked entirely, and finally classification adopts softmax
Grader.And there are activation primitive and max-pooling layers, the full linking layer of two-layer after convolutional layer.Activation primitive is Max-
Feature-Map, form are as follows:
Wherein, C ∈ Rh×w×2n, the quantity of C is 2n, and 2n represents the quantity of representative feature collection of illustrative plates;The span of k is 1-n;
The real number set that R is represented;H represents the height of convolution characteristic spectrum;W represents the width of convolution characteristic spectrum;I represents convolution feature
The position in the y-axis direction in collection of illustrative plates, span is 1-h;J represents the position in the x-axis direction in convolution characteristic spectrum, value model
Enclose for 1-w.Then the derivative of f is:
The convolution kernel and stride of four one-tenth convolutional layers is respectively:9X9X96/1、5X5X192/1、5X5X256/1、
4X4X384/1;All of max-pooling cores and stride are 2x2/2;The number of first neuron for connecting full layer is
256, depending on the number of the neuron of second full articulamentum is according to station symbol quantity, than needing identification if any 80 class station symbols, then
The number of the neuron of second full articulamentum is exactly 80.Configured by such model number of plies, ensureing the same of accuracy of identification
When improve recognition efficiency.
Deep learning algorithm needs substantial amounts of data, and 1,000,000 grades are proper, and within the short time in TV station symbol recognition task
So many sample cannot be collected.If training sample is very few, Expired Drugs just occur, i.e., in training set and checking collection
There is good effect above, but effect is very poor in actual test.The main stream approach for solving this problem now is:It is first
First trained in a huge data set (this data set can be unrelated with current task) above, obtain a basic model,
Then (finetune) is finely tuned with the training data of current task on this basic model again.Do so also have one it is excellent
Point can be just the convergence rate for accelerating model.Model training flow process is as shown in Figure 6.
Embodiment 1
The training process of TV station symbol recognition model is illustrated as a example by training a model comprising 80 station symbols.
(1) sample per class station symbol, index are gathered:The every class 3000 of sample number, between the acquisition time between every sample
It is divided into 3s;
(2) delete the improper sample for training in sample;
(3) sample of an approximately black background is selected in each class station symbol, station symbol foreground image and platform is partitioned into
Specimen body image;
(4) synthetic sample, is respectively synthesized sample according to fixed channel logo position mode and any channel logo position mode
2000/per class;
(5) basic model is trained on Imagenet data sets;
(6) using the station symbol sample of the station symbol sample and synthetic of actual acquisition as last sample, and will be all of
Sample is randomly divided into 4:1, respectively as training set and checking collection;
(7) the training set in step (6) and checking collection are finely tuned on the basic model in step (5)
(finetune) terminate model training when, the accuracy of identification when model in checking collection above is held essentially constant, obtain platform
Mark identification model.
(2) test pattern of the invention
The model for training and identification interface are deployed in into local side (television set, TV box etc.), at regular intervals
Program sectional drawing will be recognized automatically, return recognition result.
Embodiment 2
(1) model and recognition result are deployed to above 5508 machine core board of external Changhong;
(2) signal source is connected to the input of this machine core board, and the outfan of this machine core board is connected to into TV set terminal;
(3) identification 5s interval time is set;
(4) if current truncated picture is Sichuan nternational Channel, then the output result of model be one 80 dimension moral to
Amount, program can calculate the position that the score of maximum in vector is located.If this score is more than 0.95, then the identification of program
As a result it is the station symbol information represented by maximum position, i.e. Sichuan nternational Channel;No person does not export.
In sum, the present invention is by local side programmed acquisition station symbol sample, screening sample, synthetic sample, training
Identification model realizes a kind of high efficiency and high-precision real-time television TV station symbol recognition method and system.
It will appreciated by the skilled person that all or part of flow process in realizing above-described embodiment method can be
What logical various algorithm routines were realized, described program can be stored in computer read/write memory medium, and the program is being performed
When, it may include the as above flow process of the embodiment of each method.Wherein, described storage medium can be magnetic disc, CD, read-only storage note
Recall body (Read-Only Memory, ROM) or random access memory (Random AccessMemory, RAM) etc..
The ultimate principle and main feature of the present invention are the foregoing described, the description of description simply illustrates the original of the present invention
Reason, without departing from the spirit and scope of the present invention, the present invention also has various changes and modifications, these changes and improvements
Both fall within scope of the claimed invention.
Claims (10)
1. the TV TV station symbol recognition system based on deep learning, it is characterised in that include:
Sample collection module, for picking platform standard specimen sheet;
Screening sample module, for screening sample, rejects inappropriate sample;
Station symbol splits module, for splitting station symbol and background;
Sample synthesis module, for the sample of synthetic diversity;
Model training module, the sample of sample and synthetic based on collection, training station target identification model;
TV station symbol recognition module, recognizes station symbol based on identification model.
2. the TV TV station symbol recognition system based on deep learning according to claim 1, it is characterised in that acquisition module enters
During row sample collection, possesses certain time interval.
3. the TV TV station symbol recognition system based on deep learning according to claim 1, it is characterised in that inappropriate sample
Originally it is:The same frame collected due to network interim card and the platform sample of station symbol serious deformation and due to zapping cause without platform
Target sample.
4. the TV TV station symbol recognition system based on deep learning according to claim 1, it is characterised in that segmentation station symbol and
The step of background, includes:
The station symbol image of a black background is selected first, according to 1/5th of 1/3rd and high-order artwork of a width of artwork
The subimage in the upper left corner is intercepted, the subimage for obtaining is adopted into two value filterings, noise is eliminated, station symbol foreground image is obtained, finally
Both horizontally and vertically project respectively, be partitioned into station symbol itself.
5. the TV TV station symbol recognition system based on deep learning according to claim 1, it is characterised in that synthetic sample
This method includes
Fixed channel logo position synthetic method:Intercepting a fixed size first does not have the background image of station symbol;Secondly select station symbol
Foreground image;Finally image overlay by two kinds together;
Any channel logo position synthetic method:Intercepting a fixed size first does not have the background image of station symbol;Secondly select station symbol
Image itself, and determining table target size;Finally in one region as station symbol size of random selection above background image,
And this region and station symbol are superimposed.
6. the TV TV station symbol recognition system based on deep learning according to claim 5, it is characterised in that the mode of superposition
For:Pixel position not for 0 in station symbol foreground image is recorded first;Secondly the pixel value of these positions in background image is arranged
For the pixel value of these positions of station symbol foreground image.
7. the TV TV station symbol recognition system based on deep learning according to claim 1, it is characterised in that to synthetic
Sample adopt Gaussian smoothing marginal information.
8. the TV TV station symbol recognition system based on deep learning according to claim 1, it is characterised in that TV station symbol recognition mould
Block is using CNN algorithms as TV station symbol recognition algorithm;Wherein, CNN model structures include that four layers of convolutional layer and two-layer are linked entirely, and roll up
There are activation primitive and max-pooling layers after lamination.
9. the TV TV station symbol recognition system based on deep learning according to claim 8, it is characterised in that activation primitive is
Max-Feature-Map, form are as follows:
Wherein, C ∈ Rh×w×2n, the quantity of C is 2n, and 2n represents the quantity of representative feature collection of illustrative plates;The span of k is 1-n;R is represented
Real number set;H represents the height of convolution characteristic spectrum;W represents the width of convolution characteristic spectrum;I represents convolution characteristic spectrum
In y-axis direction position, span is 1-h;J represents the position in the x-axis direction in convolution characteristic spectrum, and span is
1-w。
10. the TV TV station symbol recognition system based on deep learning according to claim 8, it is characterised in that four layers of convolution
The convolution kernel and stride of layer is respectively:9X9X96/1、5X5X192/1、5X5X256/1、4X4X384/1;All of max-
Pooling cores and stride are 2x2/2;The number of first neuron for connecting full layer is 256;The god of second full articulamentum
Depending on the number of Jing units is according to station symbol quantity.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611151654.XA CN106599892A (en) | 2016-12-14 | 2016-12-14 | Television station logo identification system based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611151654.XA CN106599892A (en) | 2016-12-14 | 2016-12-14 | Television station logo identification system based on deep learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106599892A true CN106599892A (en) | 2017-04-26 |
Family
ID=58802294
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201611151654.XA Pending CN106599892A (en) | 2016-12-14 | 2016-12-14 | Television station logo identification system based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106599892A (en) |
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107330027A (en) * | 2017-06-23 | 2017-11-07 | 中国科学院信息工程研究所 | A kind of Weakly supervised depth station caption detection method |
CN107346546A (en) * | 2017-07-17 | 2017-11-14 | 京东方科技集团股份有限公司 | A kind of image processing method and device |
CN107808373A (en) * | 2017-11-15 | 2018-03-16 | 北京奇虎科技有限公司 | Sample image synthetic method, device and computing device based on posture |
CN108009629A (en) * | 2017-11-20 | 2018-05-08 | 天津大学 | A kind of station symbol dividing method based on full convolution station symbol segmentation network |
CN108460414A (en) * | 2018-02-27 | 2018-08-28 | 北京三快在线科技有限公司 | Generation method, device and the electronic equipment of training sample image |
CN108596221A (en) * | 2018-04-10 | 2018-09-28 | 江河瑞通(北京)技术有限公司 | The image-recognizing method and equipment of rod reading |
CN109117768A (en) * | 2018-07-30 | 2019-01-01 | 上海科江电子信息技术有限公司 | A kind of TV station symbol recognition method based on deep learning |
CN109344904A (en) * | 2018-10-16 | 2019-02-15 | 杭州睿琪软件有限公司 | Generate method, system and the storage medium of training sample |
CN110378999A (en) * | 2019-06-24 | 2019-10-25 | 南方电网科学研究院有限责任公司 | Target frame marking method and device for target object in training sample and storage medium |
CN110378336A (en) * | 2019-06-24 | 2019-10-25 | 南方电网科学研究院有限责任公司 | Semantic level labeling method and device for target object in training sample and storage medium |
CN111076809A (en) * | 2019-12-31 | 2020-04-28 | 四川长虹电器股份有限公司 | Convolutional neural network-based equipment abnormal sound identification method and system |
CN111462162A (en) * | 2019-01-18 | 2020-07-28 | 上海大学 | Foreground segmentation algorithm for specific class of pictures |
CN114595778A (en) * | 2022-03-15 | 2022-06-07 | 北京达佳互联信息技术有限公司 | Identification pattern recognition method and device, electronic equipment and storage medium |
US11922600B2 (en) | 2018-08-31 | 2024-03-05 | Samsung Display Co., Ltd. | Afterimage compensator, display device having the same, and method for driving display device |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103902987A (en) * | 2014-04-17 | 2014-07-02 | 福州大学 | Station caption identifying method based on convolutional network |
CN104809443A (en) * | 2015-05-05 | 2015-07-29 | 上海交通大学 | Convolutional neural network-based license plate detection method and system |
-
2016
- 2016-12-14 CN CN201611151654.XA patent/CN106599892A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103902987A (en) * | 2014-04-17 | 2014-07-02 | 福州大学 | Station caption identifying method based on convolutional network |
CN104809443A (en) * | 2015-05-05 | 2015-07-29 | 上海交通大学 | Convolutional neural network-based license plate detection method and system |
Non-Patent Citations (2)
Title |
---|
XIANG WU ET AL.: "A Lightened CNN for Deep Face Representation", 《COMPUTER SCIENCE》 * |
李慧: "台标的自动提取和识别", 《中国优秀硕士学位论文全文数据库》 * |
Cited By (20)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107330027A (en) * | 2017-06-23 | 2017-11-07 | 中国科学院信息工程研究所 | A kind of Weakly supervised depth station caption detection method |
CN107330027B (en) * | 2017-06-23 | 2020-05-22 | 中国科学院信息工程研究所 | Weak supervision depth station caption detection method |
CN107346546A (en) * | 2017-07-17 | 2017-11-14 | 京东方科技集团股份有限公司 | A kind of image processing method and device |
CN107808373A (en) * | 2017-11-15 | 2018-03-16 | 北京奇虎科技有限公司 | Sample image synthetic method, device and computing device based on posture |
CN108009629A (en) * | 2017-11-20 | 2018-05-08 | 天津大学 | A kind of station symbol dividing method based on full convolution station symbol segmentation network |
CN108460414B (en) * | 2018-02-27 | 2019-09-17 | 北京三快在线科技有限公司 | Generation method, device and the electronic equipment of training sample image |
CN108460414A (en) * | 2018-02-27 | 2018-08-28 | 北京三快在线科技有限公司 | Generation method, device and the electronic equipment of training sample image |
CN108596221B (en) * | 2018-04-10 | 2020-12-01 | 江河瑞通(北京)技术有限公司 | Image recognition method and device for scale reading |
CN108596221A (en) * | 2018-04-10 | 2018-09-28 | 江河瑞通(北京)技术有限公司 | The image-recognizing method and equipment of rod reading |
CN109117768A (en) * | 2018-07-30 | 2019-01-01 | 上海科江电子信息技术有限公司 | A kind of TV station symbol recognition method based on deep learning |
US11922600B2 (en) | 2018-08-31 | 2024-03-05 | Samsung Display Co., Ltd. | Afterimage compensator, display device having the same, and method for driving display device |
CN109344904A (en) * | 2018-10-16 | 2019-02-15 | 杭州睿琪软件有限公司 | Generate method, system and the storage medium of training sample |
CN109344904B (en) * | 2018-10-16 | 2020-10-30 | 杭州睿琪软件有限公司 | Method, system and storage medium for generating training samples |
CN111462162A (en) * | 2019-01-18 | 2020-07-28 | 上海大学 | Foreground segmentation algorithm for specific class of pictures |
CN111462162B (en) * | 2019-01-18 | 2023-07-21 | 上海大学 | Foreground segmentation algorithm for specific class pictures |
CN110378999A (en) * | 2019-06-24 | 2019-10-25 | 南方电网科学研究院有限责任公司 | Target frame marking method and device for target object in training sample and storage medium |
CN110378336A (en) * | 2019-06-24 | 2019-10-25 | 南方电网科学研究院有限责任公司 | Semantic level labeling method and device for target object in training sample and storage medium |
CN111076809A (en) * | 2019-12-31 | 2020-04-28 | 四川长虹电器股份有限公司 | Convolutional neural network-based equipment abnormal sound identification method and system |
CN111076809B (en) * | 2019-12-31 | 2021-08-31 | 四川长虹电器股份有限公司 | Convolutional neural network-based equipment abnormal sound identification method and system |
CN114595778A (en) * | 2022-03-15 | 2022-06-07 | 北京达佳互联信息技术有限公司 | Identification pattern recognition method and device, electronic equipment and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106599892A (en) | Television station logo identification system based on deep learning | |
Al-Ghaili et al. | Vertical-edge-based car-license-plate detection method | |
CN104751142B (en) | A kind of natural scene Method for text detection based on stroke feature | |
JP4942510B2 (en) | Vehicle image recognition apparatus and method | |
CN107679502A (en) | A kind of Population size estimation method based on the segmentation of deep learning image, semantic | |
CN108596166A (en) | A kind of container number identification method based on convolutional neural networks classification | |
CN109409384A (en) | Image-recognizing method, device, medium and equipment based on fine granularity image | |
CN106529578A (en) | Vehicle brand model fine identification method and system based on depth learning | |
CN106548160A (en) | A kind of face smile detection method | |
CN110766020A (en) | System and method for detecting and identifying multi-language natural scene text | |
CN106846339A (en) | Image detection method and device | |
CN111368830A (en) | License plate detection and identification method based on multi-video frame information and nuclear phase light filtering algorithm | |
CN107644221A (en) | Convolutional neural networks traffic sign recognition method based on compression of parameters | |
CN105868708A (en) | Image object identifying method and apparatus | |
CN111461039A (en) | Landmark identification method based on multi-scale feature fusion | |
CN110458201A (en) | A kind of remote sensing image object-oriented classification method and sorter | |
CN108597003A (en) | A kind of article cover generation method, device, processing server and storage medium | |
CN105825216A (en) | Method of locating text in complex background image | |
CN101615244A (en) | Handwritten plate blank numbers automatic identifying method and recognition device | |
CN115565071A (en) | Hyperspectral image transform network training and classifying method | |
CN107657251A (en) | Determine the device and method of identity document display surface, image-recognizing method | |
CN107145888A (en) | Video caption real time translating method | |
CN117522735A (en) | Multi-scale-based dense-flow sensing rain-removing image enhancement method | |
CN110363198B (en) | Neural network weight matrix splitting and combining method | |
CN106228553A (en) | High-resolution remote sensing image shadow Detection apparatus and method |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170426 |
|
RJ01 | Rejection of invention patent application after publication |