CN109376788A - A kind of image analysis method based on the high discrimination of deep learning - Google Patents
A kind of image analysis method based on the high discrimination of deep learning Download PDFInfo
- Publication number
- CN109376788A CN109376788A CN201811292095.3A CN201811292095A CN109376788A CN 109376788 A CN109376788 A CN 109376788A CN 201811292095 A CN201811292095 A CN 201811292095A CN 109376788 A CN109376788 A CN 109376788A
- Authority
- CN
- China
- Prior art keywords
- specific
- sample
- subsample
- data set
- classification
- 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
- 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
- G06F18/29—Graphical models, e.g. Bayesian networks
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of image analysis methods based on the high discrimination of deep learning, for generating image analysis system, the following steps are included: collecting data, establish familiar object data set, familiar object data set is categorized into the common sample of different classification according to object category, different specific common samples is specifically divided into again to each common sample of classification；Object mark is carried out to the subsample of specific common sample；For specific object, certain objects data set is established, certain objects data set is categorized into different classification specific samples according to object category, different specific specific samples is specifically divided into again to each classification specific sample；Object mark is carried out to the subsample of specific specific sample；Model training is carried out to each specific sample combining target detection algorithm and obtains object special purpose model.The system that this method generates is suitable for the special dimensions such as public security, the administration of justice, supports the customized extension of user, user can establish object special purpose model according to their own needs, and image recognition rate is high.
Description
Technical field
The present invention relates to field of image recognition, more particularly to a kind of image analysis side based on the high discrimination of deep learning
Method.
Background technique
In the electronic data evidence obtainings such as public security, administration of justice field, there are a large amount of picture recognition demands, and mostly at present is by artificial side
Formula is classified, is screened, often single evidence source sample just need a few hours even a couple of days, spend a large amount of manpower and time at
This；And with internet development, data volume substantially expands, and the mode of artificial treatment does not catch up with case rhythm increasingly.It is existing
Intelligent recognition classification schemes be mostly certain familiar object classifications identification, and due to the particularity in the fields such as public security, the administration of justice,
It need to identify a large amount of sensitive content, such as: lethal weapon, tobacco.Evidence obtaining field is related to many complex scenes and composograph, current
Intelligent recognition model training program only simply presses tagsort mostly, very low for more complex image recognition rate.And
Currently without the function of the customized extension of intelligent identifying system user of support.
Summary of the invention
In view of the above shortcomings of the prior art, the present invention provides a kind of image analyses based on the high discrimination of deep learning
Method can identify sensitive content in evidence obtaining field, and discrimination is higher, while supporting the customized extension of user.
In order to solve the above-mentioned technical problem, present invention employs the following technical solutions:
A kind of image analysis method based on the high discrimination of deep learning of the present invention, for generating image analysis system
System, specific step is as follows for this method:
S1: data are collected, familiar object data set is established, familiar object data set is categorized into difference according to object category
The common sample of classification, different specific common samples is specifically divided into again to the common sample of each classification；
S2: object mark is carried out to the subsample of specific common sample；
S3: it is directed to specific object, certain objects data set is established, certain objects data set is classified according to object category
At different classification specific samples, different specific specific samples is specifically divided into again to each classification specific sample；
S4: object mark is carried out to the subsample of specific specific sample；
S5: union is taken to the familiar object data set and the certain objects data set: by the common sample of classification
Classification samples are merged into the classification specific sample, specific common sample and specific specific sample in the classification samples
Merge into specific sample, if the specific common sample and the specific specific sample are repeated sample, retain described in
Specific common sample is specific sample；
S6: model training is carried out to each specific sample combining target detection algorithm and obtains object special purpose model；
S7: the object special purpose model is encapsulated into api interface and is called for front end.
As optimization, described image analysis system includes the A object special purpose models, and the object special purpose model is by normal
See object data set and certain objects data set composition, the familiar object data set includes the M common sample of classifying, and described point
The common sample of class is made of N number of specific common sample, and the certain objects data set includes X classification specific sample and Y tool
Body specific sample, wherein A, M, N, X, Y are the positive integer more than or equal to 1；The specific common sample and described specific specific
Sample includes Z subsample, wherein Z is the positive integer more than or equal to 1.
As optimization, the object mask method in the step S2 and S4 is rectangle frame mark and special-shaped collimation mark note.
As optimization, the algorithm of target detection in the step S6 is Faster-RCNN algorithm.
As optimization, the specific implementation step of the step S6 is as follows:
6.1) subsample for extracting fixed proportion at random to each specific sample is labeled as detection model subsample, remaining
Subsample be training pattern subsample；
6.2) from the region that is marked in training pattern subsample, extraction represents region at random from top to bottom, according to the generation
Table section, which generates, suggests window, and joint training class probability and frame regression algorithm obtain object special purpose model；
6.3) the detection model subsample is matched with the object special purpose model, if there is at least 40% detection
The matching degree of model subsample reaches predetermined probabilities, then approves the object special purpose model, and the prediction frame for recording subsample is sat
It marks and automatic marking is carried out to subsample；Otherwise return step 6.2) it reselects and represents region.
As optimization, the predetermined probabilities are 99.9%.
As optimization, needed before being trained to the subsample to sub- sample preprocessing into unified format.
The beneficial effects of the present invention are:
The system that this method generates is suitable for the special dimensions such as public security, the administration of justice, supports the customized extension of user, and user can be with
Object special purpose model is established according to their own needs, and image recognition rate is high.
Detailed description of the invention
Fig. 1 is a kind of flow chart of the image analysis method based on the high discrimination of deep learning of the present invention；
Fig. 2 is that a kind of system of image analysis method based on the high discrimination of deep learning of the present invention forms figure；
Fig. 3 is a kind of example flow diagram of the image analysis method based on the high discrimination of deep learning of the present invention.
Specific embodiment
The present invention is described in further detail with reference to the accompanying drawing.
As shown in Figure 1, Figure 3, a kind of image analysis method based on the high discrimination of deep learning, for generating image analysis
System, comprising the following steps:
S1: data are collected, familiar object data set is established, familiar object data set is categorized into difference according to object category
The common sample of classification, different specific common samples is specifically divided into again to the common sample of each classification.It is received in the present embodiment
Collecting data is collected in coco data set, can also be collected in other existing data sets.
S2: object mark is carried out to the subsample of specific common sample；
S3: it is directed to specific object, certain objects data set is established, certain objects data set is classified according to object category
At different classification specific samples, different specific specific samples is specifically divided into again to each classification specific sample；
S4: object mark is carried out to the subsample of specific specific sample.
In the present embodiment, step S2 and step S4 can be used rectangle frame according to the concrete shape of subsample and mark and special-shaped frame
Mark.If there is multiple objects in subsample, multiple objects are individually marked, and callout box is all surrounded object and got over apart from object
It is short better, give up the object that fuzzy and human eye can not recognize.For Special-shaped object, rectangle frame mark can include interference on a large scale
When object, is infused using multilateral abnormity collimation mark, the quality of feature extraction can be improved.Be utilized respectively LabelImg software or
Labelme software plays the frame segment that object needs to mark, and software, which generates detailed coordinate to the point for forming callout box automatically, to be believed
Breath.
S5: take union to familiar object data set and certain objects data set: will classify common sample and specific sample of classifying
Originally classification samples are merged into, the specific common sample and specific specific sample in classification samples merge into specific sample,
If specific common sample and specific specific sample are repeated sample, retaining specific common sample is specific sample.According to specific
The title of common sample and specific specific sample is to determine whether be repeated sample.
S6: model training is carried out to each specific sample combining target detection algorithm and obtains object special purpose model.This reality
It applies in example using Faster-RCNN algorithm, or the algorithm of other training patterns.
S7: object special purpose model is encapsulated into api interface and is called for front end.
As shown in Fig. 2, image analysis system includes A object special purpose model, and object special purpose model is by normal in the present embodiment
See object data set and certain objects data set composition, familiar object data set includes the M common sample of classification, common sample of classifying
This is made of N number of specific common sample, and certain objects data set includes X classification specific sample and a specific specific samples of Y,
In, A, M, N, X, Y are the positive integer more than or equal to 1；Specific common sample and specific specific sample include Z subsample,
In, Z is the positive integer more than or equal to 1.
In the present embodiment, the specific implementation step of step S6 is as follows:
6.1) subsample for extracting fixed proportion at random to each specific sample is labeled as detection model subsample, remaining
Subsample be training pattern subsample；
6.2) from the region that is marked in training pattern subsample, extraction represents region at random from top to bottom, according to Representative Region
Domain, which generates, suggests window, and joint training class probability and frame regression algorithm obtain object special purpose model.
Training pattern subsample is inputted into CNN, the region that is marked in training pattern subsample is mentioned at random from top to bottom
Replace table section, generated with RPN and suggest window (proposals), every picture generates 300 suggestion windows, suggestion window
It is mapped on the last layer convolution feature map of CNN, generates each RoI by pooling layers of RoI fixed-size
Feature map, using Softmax Loss (detection class probability) and Smooth L1Loss (detection frame returns) to classification
Probability and frame return (Bounding box regression) joint training and obtain object special purpose model.
6.3) it will test model subsample to be matched with object special purpose model, if there is at least 40% detection model increment
This matching degree reaches predetermined probabilities, then approves object special purpose model, record the prediction frame coordinate of subsample and to subsample
Carry out automatic marking；Otherwise return step 6.2) it reselects and represents region.
In the present embodiment, predetermined probabilities 99.9%.
In the present embodiment, needed before being trained to subsample to sub- sample preprocessing into unified format.The present embodiment
Subsample format be JPG format, subsample size is in 300*300-500*500.
The working principle of the invention: object special purpose model, such as lethal weapon special purpose model are established, in certain objects data set
Picture relevant to lethal weapon is collected, then takes union with familiar object data set, binding object marks method and model training method is raw
At object special purpose model.The picture of suspect is input in object special purpose model by law enforcement agency, and object special purpose model is by suspicion
It is all found out in the picture of people about the picture of lethal weapon.
Finally, it should be noted that those skilled in the art various changes and modifications can be made to the invention without departing from
The spirit and scope of the present invention.In this way, if these modifications and changes of the present invention belongs to the claims in the present invention and its waits system
Within the scope of counting, then the present invention is also intended to encompass these modification and variations.
Claims (7)
1. a kind of image analysis method based on the high discrimination of deep learning, for generating image analysis system, which is characterized in that
The following steps are included:
S1: collecting data, establish familiar object data set, and familiar object data set is categorized into different points according to object category
The common sample of class is specifically divided into different specific common samples to each common sample of classification again；
S2: object mark is carried out to the subsample of specific common sample；
S3: it is directed to specific object, certain objects data set is established, certain objects data set is categorized into not according to object category
Same classification specific sample, is specifically divided into different specific specific samples to each classification specific sample again；
S4: object mark is carried out to the subsample of specific specific sample；
S5: union is taken to the familiar object data set and the certain objects data set: by the common sample of the classification and institute
It states classification specific sample and is merged into classification samples, the specific common sample and specific specific sample in the classification samples carry out
It is merged into specific sample, if the specific common sample and the specific specific sample are repeated sample, is retained described specific
Common sample is specific sample；
S6: model training is carried out to each specific sample combining target detection algorithm and obtains object special purpose model；
S7: the object special purpose model is encapsulated into api interface and is called for front end.
2. a kind of image analysis method based on the high discrimination of deep learning according to claim 1, which is characterized in that institute
Stating image analysis system includes A object special purpose models, and the object special purpose model is by familiar object data set and specific
Object data set composition, the familiar object data set include the M common sample of classification, and the common sample of classification is by N number of tool
The common sample composition of body, the certain objects data set include X classification specific sample and Y specific specific samples, wherein A,
M, N, X, Y are the positive integer more than or equal to 1；The specific common sample and the specific specific sample include Z subsample,
Wherein, Z is the positive integer more than or equal to 1.
3. a kind of image analysis method based on the high discrimination of deep learning according to claim 1, which is characterized in that institute
Stating the object mask method in step S2 and S4 is rectangle frame mark and special-shaped collimation mark note.
4. a kind of image analysis method based on the high discrimination of deep learning according to claim 1, which is characterized in that institute
Stating the algorithm of target detection in step S6 is Faster-RCNN algorithm.
5. a kind of image analysis method based on the high discrimination of deep learning according to claim 1, which is characterized in that institute
The specific implementation step for stating step S6 is as follows:
6.1) subsample of fixed proportion is extracted at random to each specific sample labeled as detection model subsample, remaining son
Sample is training pattern subsample；
6.2) from the region that is marked in training pattern subsample, extraction represents region at random from top to bottom, according to the Representative Region
Domain, which generates, suggests window, and joint training class probability and frame regression algorithm obtain object special purpose model；
6.3) the detection model subsample is matched with the object special purpose model, if there is at least 40% detection model
The matching degree of subsample reaches predetermined probabilities, then approves the object special purpose model, records the prediction frame coordinate of subsample simultaneously
Automatic marking is carried out to subsample；Otherwise return step 6.2) it reselects and represents region.
6. a kind of image analysis method based on the high discrimination of deep learning according to claim 5, which is characterized in that institute
Stating predetermined probabilities is 99.9%.
7. a kind of image analysis method based on the high discrimination of deep learning according to claim 5, which is characterized in that
It is needed before being trained to the subsample to sub- sample preprocessing into unified format.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811292095.3A CN109376788A (en) | 2018-10-31 | 2018-10-31 | A kind of image analysis method based on the high discrimination of deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811292095.3A CN109376788A (en) | 2018-10-31 | 2018-10-31 | A kind of image analysis method based on the high discrimination of deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109376788A true CN109376788A (en) | 2019-02-22 |
Family
ID=65396799
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811292095.3A Pending CN109376788A (en) | 2018-10-31 | 2018-10-31 | A kind of image analysis method based on the high discrimination of deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109376788A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111523610A (en) * | 2020-05-06 | 2020-08-11 | 青岛联合创智科技有限公司 | Article identification method for efficient sample marking |
| CN111651629A (en) * | 2019-03-27 | 2020-09-11 | 上海铼锶信息技术有限公司 | Method and system for constructing full sample data |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104217216A (en) * | 2014-09-01 | 2014-12-17 | 华为技术有限公司 | Method and device for generating detection model, method and device for detecting target |
| CN106203498A (en) * | 2016-07-07 | 2016-12-07 | 中国科学院深圳先进技术研究院 | A kind of City scenarios rubbish detection method and system |
| CN107368787A (en) * | 2017-06-16 | 2017-11-21 | 长安大学 | A kind of Traffic Sign Recognition algorithm that application is driven towards depth intelligence |
| US9984283B2 (en) * | 2015-02-14 | 2018-05-29 | The Trustees Of The University Of Pennsylvania | Methods, systems, and computer readable media for automated detection of abnormalities in medical images |
| CN108154118A (en) * | 2017-12-25 | 2018-06-12 | 北京航空航天大学 | A kind of target detection system and method based on adaptive combined filter with multistage detection |
| CN108182456A (en) * | 2018-01-23 | 2018-06-19 | 哈工大机器人（合肥）国际创新研究院 | A kind of target detection model and its training method based on deep learning |
| CN108511058A (en) * | 2018-03-01 | 2018-09-07 | 海南星德智慧科技有限公司 | A kind of sick detection device of sugar net |
| CN108520518A (en) * | 2018-04-10 | 2018-09-11 | 复旦大学附属肿瘤医院 | A kind of thyroid tumors Ultrasound Image Recognition Method and its device |
-
2018
- 2018-10-31 CN CN201811292095.3A patent/CN109376788A/en active Pending
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104217216A (en) * | 2014-09-01 | 2014-12-17 | 华为技术有限公司 | Method and device for generating detection model, method and device for detecting target |
| US9984283B2 (en) * | 2015-02-14 | 2018-05-29 | The Trustees Of The University Of Pennsylvania | Methods, systems, and computer readable media for automated detection of abnormalities in medical images |
| CN106203498A (en) * | 2016-07-07 | 2016-12-07 | 中国科学院深圳先进技术研究院 | A kind of City scenarios rubbish detection method and system |
| CN107368787A (en) * | 2017-06-16 | 2017-11-21 | 长安大学 | A kind of Traffic Sign Recognition algorithm that application is driven towards depth intelligence |
| CN108154118A (en) * | 2017-12-25 | 2018-06-12 | 北京航空航天大学 | A kind of target detection system and method based on adaptive combined filter with multistage detection |
| CN108182456A (en) * | 2018-01-23 | 2018-06-19 | 哈工大机器人（合肥）国际创新研究院 | A kind of target detection model and its training method based on deep learning |
| CN108511058A (en) * | 2018-03-01 | 2018-09-07 | 海南星德智慧科技有限公司 | A kind of sick detection device of sugar net |
| CN108520518A (en) * | 2018-04-10 | 2018-09-11 | 复旦大学附属肿瘤医院 | A kind of thyroid tumors Ultrasound Image Recognition Method and its device |
Non-Patent Citations (2)
| Title |
|---|
| YIN CUI 等: "Large Scale Fine-Grained Categorization and Domain-Specific Transfer Learning", 《ARXIV:1806.06193V1》 * |
| 李碧雯: "基于迁移学习的跨项目软件缺陷预测", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111651629A (en) * | 2019-03-27 | 2020-09-11 | 上海铼锶信息技术有限公司 | Method and system for constructing full sample data |
| CN111651629B (en) * | 2019-03-27 | 2023-08-18 | 上海铼锶信息技术有限公司 | Method and system for constructing full sample data |
| CN111523610A (en) * | 2020-05-06 | 2020-08-11 | 青岛联合创智科技有限公司 | Article identification method for efficient sample marking |
| CN111523610B (en) * | 2020-05-06 | 2023-04-21 | 青岛联合创智科技有限公司 | Article identification method for efficient labeling of samples |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106529499A (en) | Fourier descriptor and gait energy image fusion feature-based gait identification method | |
| CN104143079B (en) | The method and system of face character identification | |
| CN105160310A (en) | 3D (three-dimensional) convolutional neural network based human body behavior recognition method | |
| CN105184160B (en) | A kind of method of the Android phone platform application program malicious act detection based on API object reference relational graphs | |
| CN111178197A (en) | Mass R-CNN and Soft-NMS fusion based group-fed adherent pig example segmentation method | |
| CN107909572A (en) | Pulmonary nodule detection method and system based on image enhancement | |
| CN108228114A (en) | Control method and storage medium | |
| CN106295532B (en) | A kind of human motion recognition method in video image | |
| CN106408030A (en) | SAR image classification method based on middle lamella semantic attribute and convolution neural network | |
| CN105608454A (en) | Text structure part detection neural network based text detection method and system | |
| CN108764084A (en) | Video classification methods based on spatial domain sorter network and the time domain network integration | |
| CN107945875A (en) | Pulmonary nodule detection method and system based on data enhancing | |
| CN109993201A (en) | A kind of image processing method, device and readable storage medium storing program for executing | |
| CN102915432A (en) | Method and device for extracting vehicle-bone microcomputer image video data | |
| CN110245697A (en) | A kind of dirty detection method in surface, terminal device and storage medium | |
| CN109284760A (en) | A kind of furniture detection method and device based on depth convolutional neural networks | |
| WO2020071558A1 (en) | Business form layout analysis device, and analysis program and analysis method therefor | |
| CN109376788A (en) | A kind of image analysis method based on the high discrimination of deep learning | |
| CN107578424A (en) | A kind of dynamic background difference detecting method, system and device based on space-time classification | |
| CN112052730A (en) | 3D dynamic portrait recognition monitoring device and method | |
| CN115393666A (en) | Small sample expansion method and system based on prototype completion in image classification | |
| CN113076860B (en) | Bird detection system under field scene | |
| CN113673618A (en) | Tobacco insect target detection method fused with attention model | |
| CN113076917B (en) | Pedestrian quality evaluation method and system | |
| CN111199050A (en) | System for automatically desensitizing medical records and application |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190222 |