CN110008927A - One kind automating determination method based on the improved alert of deep learning model Fast-Rcnn - Google Patents
One kind automating determination method based on the improved alert of deep learning model Fast-Rcnn Download PDFInfo
- Publication number
- CN110008927A CN110008927A CN201910301206.0A CN201910301206A CN110008927A CN 110008927 A CN110008927 A CN 110008927A CN 201910301206 A CN201910301206 A CN 201910301206A CN 110008927 A CN110008927 A CN 110008927A
- Authority
- CN
- China
- Prior art keywords
- rcnn
- pond
- learning model
- deep learning
- size
- 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
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (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 discloses one kind to automate determination method based on the improved alert of deep learning model Fast-Rcnn, the following steps are included: S1: Image Acquisition, the analog video signal of CCD or cmos camera is stored after A/D using image pick-up card or Video Adapter, then computer is sent to be handled；S2: candidate frame determines；S3: depth network extracts feature, CNN network can connect several full articulamentums after convolutional layer, the characteristic pattern that convolutional layer generates is mapped to the feature vector of a regular length, uses around the pixel image block as the input of CNN for training and predicting；S4: image classification can be responsible for creating and returning keymake appropriate by adapter for the practical class of object specific to target device；S5: picture size adjustment；S6: prediction of result, most latter two loss layers will be changed to one softmax layers, using this algorithm only detect human body it is mobile when, can have good prediction to true alert degree.
Description
Technical field
The present invention relates to computer software technical fields, specially a kind of to be improved based on deep learning model Fast-Rcnn
Alert automate determination method.
Background technique
There are thousands of family and trade company to be mounted with that security protection camera, most of camera have such function at present
Can, i.e., when security protection alarm is triggered, camera can capture photo site automatically and pass the security protection control centre at rear back.Mesh
The preceding processing for these images is essentially artificial treatment, i.e., has staff's point to open picture viewing to determine whether really from the background
There is alert, and whether needs responding.Not only speed is slow for such processing mode, but also does not have uniformity.Different personnel's
Judgment criteria is different, is difficult to unification for the judgement of alert, and the same staff is in the judgment criteria of same time
It may also be different.It gives one example, when staff's fatigue is, the accuracy of judgement may decline.
A problem is also that the size issue of data volume.The user of annual security protection is increasing, at present may manpower
What is also handled comes, but when number of users more to a certain degree when, required manpower is too many, and cost is also just on
Rise, so there are many disadvantages with the method that manpower judges at present.
Summary of the invention
The purpose of the present invention is to provide one kind to be determined based on the improved alert automation of deep learning model Fast-Rcnn
Method, not instead of by the movement of detection object, only detected using algorithm human body it is mobile when, to true alert degree
There can be good prediction, to solve the problems mentioned in the above background technology.
To achieve the above object, the invention provides the following technical scheme: a kind of changed based on deep learning model Fast-Rcnn
Into alert automate determination method, the alert automate determination method specific step is as follows:
S1: Image Acquisition, using image pick-up card or Video Adapter by CCD (Charge Coupled Device, charge coupling
Clutch part) or the analog video signal of CMOS (complementary metal oxide semiconductor) video camera stored after A/D, then send meter
Calculation machine is handled；
S2: candidate frame determines that image normalization is sent directly into network for 224 × 224, first five stage is the conv+ on basis
Relu+pooling form ends up in the 5th stage, inputs P candidate region (picture numbers × geometric position × 4 1+, serial number
For training), each candidate region is uniformly divided into M × N block by roi_pool layers, to every piece of progress max pooling.By feature
Candidate region not of uniform size is changed into the unified data of size on figure, is sent into next layer；
S3: depth network extracts feature, and CNN network can connect several full articulamentums after convolutional layer, by convolutional layer
The characteristic pattern of generation is mapped to the feature vector of a regular length, uses an image block around the pixel as CNN's
Input for training and predicting, in program the 6th and 32 it is corresponding be pool1 and pool5 layers in former network.Whole network output
As a result, you can correspond to former network size finally by output result, to save the picture feature result of other layers；
S4: image classification, for object practical class specific to target device, it is appropriate to be responsible for creating and returning by adapter
Keymake, CIFAR102 and CIFAR100 database is two subsets of Tiny, has separately included 10 classes and 100 class objects
Images body classification；
S5: picture size adjustment, the pond oI layer are used to different size of input tensor pond turning to fixed size, the pond RoI
Change layer specify pond window quantity for W times HW × H, the size of each pond window be according to pond regional change,
S6: prediction of result, most latter two loss layers will be changed to one softmax layers, and input is the score of classification, output
Probability is finally used NMS to each classification, is classified using softmax, is obtained frame using recurrence and is returned, the main body of CNN
Structure may come from AlexNet, can be from VGGNet.
Preferably, the CMOS in the step S1 is complementary metal oxide semiconductor, it is originally one in computer system
The important chip of kind, saves the most basic data of System guides, and CMOS passes through the image sensing being processed into digital photography
Device.
Preferably, the node that is set as input layer for common max pooling layers in the step S2 and be the section of output layer
Point and then decision function indicate whether i-node is selected as maximum value by j node and is exported, and not selected there are two types of possible: not in range
It is interior, or be not maximum value, for roi max pooling, an input node may be connected with multiple output nodes, be set as
The node of input layer is the output node of the candidate region.
Preferably, the picture size of two databases in the step S4 is all 32 × 32, and is color image.
Preferably, CIFAR10 includes 60,000 image, wherein 5 general-purposes are in model training, 1 general-purpose is in test, every one kind object
Body has 5000 images for training, and 1000 images are similar with CIFAR-10 composition for testing CFAR-100.
Preferably, the pond layer in the step S5 includes super parameter, and the super parameter in pond includes filter size f
With stride s, common parameter value is f=2, s=2.
Preferably, applying frequency is very high, and effect is equivalent to height and width reduction half, also has using f=3, s=
2 the case where.
Preferably, Softmax function in the step S6, or normalization exponential function, are that one kind of logical function pushes away
Extensively, Softmax function is actually the log of gradient normalization of finite term discrete probability distribution.
Technical effect and advantage of the invention: proposed by the present invention a kind of based on deep learning model Fast-Rcnn improvement
Alert automate determination method, compared with prior art, can only detect people using the improved calculation method of Fast-Rcnn
Body it is mobile when, can have good prediction to true alert degree, avoid leading because of toy in image or natural cause
The mobile technical problem for causing computer higher to the rate of false alarm of alert of the object of cause.
Detailed description of the invention
Fig. 1, which is that the present invention is a kind of, automates determination method frame based on the improved alert of deep learning model Fast-Rcnn
Schematic diagram.
Fig. 2 is a kind of stream that determination method is automated based on the improved alert of deep learning model Fast-Rcnn of the present invention
Cheng Tu.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.Herein
Described specific embodiment is only used to explain the present invention, is not intended to limit the present invention.Based on the embodiments of the present invention,
Every other embodiment obtained by those of ordinary skill in the art without making creative efforts, belongs to this hair
The range of bright protection.
Sentenced referring to Fig. 1, the present invention provides one kind based on the improved alert automation of deep learning model Fast-Rcnn
Determine method, comprising the following steps:
S1: Image Acquisition, using image pick-up card or Video Adapter by CCD (Charge Coupled Device, charge coupling
Clutch part) or the analog video signal of CMOS (complementary metal oxide semiconductor) video camera stored after A/D, then send meter
Calculation machine is handled, and CCD refers to charge-coupled device, be it is a kind of indicate signal magnitude with charge amount, with coupled modes transmit believe
Number detecting element；
S2:S2: candidate frame determines that image normalization is sent directly into network for 224 × 224, first five stage is basic
Conv+relu+pooling form ends up in the 5th stage, P candidate region of input (picture numbers × geometric position × 4 1+,
Serial number is for training), each candidate region is uniformly divided into M × N block by roi_pool layers, to every piece of progress max pooling.It will
Candidate region not of uniform size is changed into the unified data of size on characteristic pattern, is sent into next layer, and RoI refers to
Mapping of " candidate frame " obtained after the completion of SelectiveSearch on characteristic pattern, candidate frame is generated by RPN, so
Each " candidate frame " is mapped on characteristic pattern again afterwards, obtains RoIs；
S3:S3: depth network extracts feature, and CNN network can connect several full articulamentums after convolutional layer, by convolution
The characteristic pattern (feature map) that layer generates is mapped to the feature vector of a regular length, uses one around the pixel
Image block as CNN input for training and predict, in program the 6th and 32 it is corresponding be pool1 and pool5 in former network
Layer.Whole network output as a result, you can finally by output result correspond to former network size, to save other layers
Picture feature result；
S4: image classification, for object practical class specific to target device, it is appropriate to be responsible for creating and returning by adapter
Keymake, CIFAR102 and CIFAR100 database is two subsets of Tiny, has separately included 10 classes and 100 class objects
Images body classification；
S5: picture size adjustment, the pond oI layer are used to different size of input tensor pond turning to fixed size, the pond RoI
Change layer specify pond window quantity for W times HW × H, the size of each pond window is according to pond regional change；
S6: prediction of result, most latter two loss layers will be changed to one softmax layers, and input is the score of classification, output
Probability is finally used NMS to each classification, is classified using softmax, is obtained frame using recurrence and is returned, the main body of CNN
Structure may come from AlexNet, can be from VGGNet.
CMOS in the step S1 is complementary metal oxide semiconductor, it is originally a kind of important in computer system
Chip, saves the most basic data of System guides, and CMOS passes through the imaging sensor being processed into digital photography.
The node of input layer is set as by common max pooling layers in the step S2 and is then sentenced for the node of output layer
Certainly whether function representation i-node by j node is selected as maximum value output, it is not selected there are two types of may: not in range, or not
It is maximum value, for roi max pooling, an input node may be connected with multiple output nodes, be set as input layer
Node is the output node of the candidate region.
The picture size of two databases in the step S4 is all 32 × 32, and is color image, CIFAR10 packet
Containing 60,000 image, wherein 5 general-purposes, in model training, for 1 general-purpose in test, each type objects have 5000 images for training,
1000 images are similar with CIFAR-10 composition for testing CFAR-100.
Pond layer in the step S5 includes super parameter, and the super parameter in pond includes filter size f and stride
S, common parameter value are f=2, and s=2, applying frequency is very high, and effect is equivalent to height and width reduction half, are also had
Using f=3, the case where s=2.
Softmax function in the step S6, or normalization exponential function, are a kind of popularizations of logical function,
Softmax function is actually the log of gradient normalization of finite term discrete probability distribution.
In summary: preferential selection is passed through using the analog video signal of CMOS (complementary metal oxide semiconductor) video camera
It is stored after A/D, then computer is sent to be handled, image normalization is sent directly into network for 224 × 224, first five stage is base
The conv+relu+pooling form of plinth, after every piece of progress max pooling, by candidate regions not of uniform size on characteristic pattern
Domain is changed into the unified data of size, is sent into next layer.
CNN network can connect several full articulamentums after convolutional layer, and the characteristic pattern that convolutional layer generates is mapped to one
The feature vector of a regular length uses around the pixel image block as the input of CNN for training and predicting, right
In object practical class specific to target device, be responsible for creating and returning keymake appropriate by adapter, will be different size of
Input tensor pond turns to fixed size, and it is W times HW × H that the pond RoI layer, which specifies the quantity of pond window, most latter two
Loss layers will be changed to one softmax layers, and input is the score of classification, and output probability uses NMS to each classification, uses
Softmax is classified.
Finally, it should be noted that the foregoing is only a preferred embodiment of the present invention, it is not intended to restrict the invention,
Although the present invention is described in detail referring to the foregoing embodiments, for those skilled in the art, still may be used
To modify the technical solutions described in the foregoing embodiments or equivalent replacement of some of the technical features,
All within the spirits and principles of the present invention, any modification, equivalent replacement, improvement and so on should be included in of the invention
Within protection scope.
Claims (6)
1. one kind automates determination method based on the improved alert of deep learning model Fast-Rcnn, it is characterised in that: the alert
Automating determination method, specific step is as follows:
S1: Image Acquisition, using image pick-up card or Video Adapter by the analog video signal of CCD or cmos camera through A/D
After store, then computer is sent to be handled；
S2: candidate frame determines that image normalization is sent directly into network for 224 × 224, first five stage is the conv+relu+ on basis
Pooling form ends up in the 5th stage, inputs P candidate region, each candidate region is uniformly divided into M by roi_pool layers
Candidate region not of uniform size on characteristic pattern is changed into the unified data of size to every piece of progress max pooling by × N block,
It is sent into next layer；
S3: depth network extracts feature, and CNN network can connect several full articulamentums after convolutional layer, convolutional layer is generated
Characteristic pattern be mapped to the feature vector of a regular length, use an image block around the pixel as the input of CNN
For training and predicting, in program the 6th and 32 it is corresponding be pool1 and pool5 layers in former network；The knot of whole network output
Fruit can correspond to former network size finally by output result, to save the picture feature result of other layers；
S4: image classification, for object practical class specific to target device, be responsible for creating and returning volume appropriate by adapter
Device is write, CIFAR102 and CIFAR100 database is two subsets of Tiny, has separately included 10 classes and 100 class object Images bodies
Classification；
S5: different size of input tensor pond is turned to fixed size by picture size adjustment, the pond oI layer, and the pond RoI layer is specified
The quantity of pond window be W times HW × H, the size of each pond window is according to pond regional change；
S6: prediction of result, most latter two loss layers will be changed to one softmax layers, and input is the score of classification, output probability,
NMS is used to each classification, is classified using softmax, frame is obtained using recurrence and returns.
2. according to claim 1 a kind of based on deep learning model Fast-Rcnn improved alert automation judgement side
Method, it is characterised in that: the CMOS in the step S1 is complementary metal oxide semiconductor, it is originally a kind of in computer system
Important chip, saves the most basic data of System guides, and CMOS passes through the imaging sensor being processed into digital photography.
3. according to claim 1 or 2 a kind of based on the improved alert automation judgement of deep learning model Fast-Rcnn
Method, it is characterised in that: the node that is set as input layer for common max pooling layers in the step S2 and be the section of output layer
Then decision function indicates whether i-node is selected as maximum value by j node and is exported to point.
4. according to claim 1-3 a kind of automatic based on the improved alert of deep learning model Fast-Rcnn
Change determination method, it is characterised in that: the picture size of two databases in the step S4 is all 32 × 32, and is colored
Image.
5. according to claim 1 a kind of based on deep learning model Fast-Rcnn improved alert automation judgement side
Method, it is characterised in that: the pond layer in the step S5 includes super parameter, and the super parameter in pond includes filter size f
With stride s, common parameter value is f=2, s=2.
6. according to claim 1 a kind of based on deep learning model Fast-Rcnn improved alert automation judgement side
Method, it is characterised in that: Softmax function in the step S6, Softmax function are actually finite term discrete probability distribution
Log of gradient normalization.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910301206.0A CN110008927A (en) | 2019-04-15 | 2019-04-15 | One kind automating determination method based on the improved alert of deep learning model Fast-Rcnn |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910301206.0A CN110008927A (en) | 2019-04-15 | 2019-04-15 | One kind automating determination method based on the improved alert of deep learning model Fast-Rcnn |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110008927A true CN110008927A (en) | 2019-07-12 |
Family
ID=67172003
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910301206.0A Pending CN110008927A (en) | 2019-04-15 | 2019-04-15 | One kind automating determination method based on the improved alert of deep learning model Fast-Rcnn |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110008927A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112455676A (en) * | 2019-09-09 | 2021-03-09 | 中国电力科学研究院有限公司 | Intelligent monitoring and analyzing system and method for health state of photovoltaic panel |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170124409A1 (en) * | 2015-11-04 | 2017-05-04 | Nec Laboratories America, Inc. | Cascaded neural network with scale dependent pooling for object detection |
| CN106855944A (en) * | 2016-12-22 | 2017-06-16 | 浙江宇视科技有限公司 | Pedestrian's Marker Identity method and device |
| CN107729799A (en) * | 2017-06-13 | 2018-02-23 | 银江股份有限公司 | Crowd's abnormal behaviour vision-based detection and analyzing and alarming system based on depth convolutional neural networks |
| CN109190668A (en) * | 2018-08-01 | 2019-01-11 | 福州大学 | The detection of multiclass certificate and classification method based on Faster-RCNN |
| US10198671B1 (en) * | 2016-11-10 | 2019-02-05 | Snap Inc. | Dense captioning with joint interference and visual context |
| CN109492534A (en) * | 2018-10-12 | 2019-03-19 | 高新兴科技集团股份有限公司 | A kind of pedestrian detection method across scene multi-pose based on Faster RCNN |
| CN109525787A (en) * | 2018-12-13 | 2019-03-26 | 南京邮电大学 | Real-time caption translating and network system realization towards live scene |
-
2019
- 2019-04-15 CN CN201910301206.0A patent/CN110008927A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170124409A1 (en) * | 2015-11-04 | 2017-05-04 | Nec Laboratories America, Inc. | Cascaded neural network with scale dependent pooling for object detection |
| US10198671B1 (en) * | 2016-11-10 | 2019-02-05 | Snap Inc. | Dense captioning with joint interference and visual context |
| CN106855944A (en) * | 2016-12-22 | 2017-06-16 | 浙江宇视科技有限公司 | Pedestrian's Marker Identity method and device |
| CN107729799A (en) * | 2017-06-13 | 2018-02-23 | 银江股份有限公司 | Crowd's abnormal behaviour vision-based detection and analyzing and alarming system based on depth convolutional neural networks |
| CN109190668A (en) * | 2018-08-01 | 2019-01-11 | 福州大学 | The detection of multiclass certificate and classification method based on Faster-RCNN |
| CN109492534A (en) * | 2018-10-12 | 2019-03-19 | 高新兴科技集团股份有限公司 | A kind of pedestrian detection method across scene multi-pose based on Faster RCNN |
| CN109525787A (en) * | 2018-12-13 | 2019-03-26 | 南京邮电大学 | Real-time caption translating and network system realization towards live scene |
Non-Patent Citations (5)
| Title |
|---|
| AI之路: "Fast RCNN算法详解", 《HTTPS://BLOG.CSDN.NET/U014380165/ARTICLE/DETAILS/72851319 2/5》 * |
| ROSS GIRSHICK等: "Fast R-CNN", 《COMPUTER VISION AND PATTERN RECOGNITION》 * |
| SHENXIAOLU1984: "【目标检测】Fast RCNN算法详解", 《 HTTPS://BLOG.CSDN.NET/SHENXIAOLU1984/ARTICLE/DETAILS/51036677》 * |
| 月见樽: "Fast-RCNN阅读笔记", 《HTTPS://WWW.JIANSHU.COM/P/802BBCE17F74》 * |
| 阎相伊: "基于深度卷积神经网络的企业生产安全智能监控系统", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112455676A (en) * | 2019-09-09 | 2021-03-09 | 中国电力科学研究院有限公司 | Intelligent monitoring and analyzing system and method for health state of photovoltaic panel |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108764372B (en) | Construction method and device, mobile terminal, the readable storage medium storing program for executing of data set | |
| CN109670452A (en) | Method for detecting human face, device, electronic equipment and Face datection model | |
| CN102789672B (en) | A kind of baby's sleeping position intelligent identification Method and device | |
| CN108323204A (en) | A kind of method and intelligent terminal of detection face flaw point | |
| CN107871134A (en) | A kind of method for detecting human face and device | |
| CN110378945A (en) | Depth map processing method, device and electronic equipment | |
| CN109886153B (en) | Real-time face detection method based on deep convolutional neural network | |
| CN106951870B (en) | Intelligent detection and early warning method for active visual attention of significant events of surveillance video | |
| CN108549891A (en) | Multi-scale diffusion well-marked target detection method based on background Yu target priori | |
| CN112699808B (en) | Dense target detection method, electronic equipment and related products | |
| CN106469302A (en) | A kind of face skin quality detection method based on artificial neural network | |
| CN108986075A (en) | A kind of judgment method and device of preferred image | |
| CN108921830A (en) | A kind of demographic method based on image retrieval | |
| CN109345553A (en) | A kind of palm and its critical point detection method, apparatus and terminal device | |
| CN109816745A (en) | Human body thermodynamic chart methods of exhibiting and Related product | |
| CN112347887A (en) | Object detection method, object detection device and electronic equipment | |
| CN110400338A (en) | Depth map processing method, device and electronic equipment | |
| CN109711241A (en) | Object detecting method, device and electronic equipment | |
| CN108764371A (en) | Image processing method, device, computer readable storage medium and electronic equipment | |
| CN106027902A (en) | Photographing method and mobile terminal | |
| CN111339902A (en) | Liquid crystal display number identification method and device of digital display instrument | |
| CN116385430A (en) | Machine vision flaw detection method, device, medium and equipment | |
| CN111310531B (en) | Image classification method, device, computer equipment and storage medium | |
| CN109934072A (en) | Personnel statistical method and device | |
| CN111639668A (en) | Crowd density detection method based on deep learning |
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: 20190712 |