CN109460820A - A kind of neural network training method, device, computer equipment and storage medium - Google Patents
A kind of neural network training method, device, computer equipment and storage medium Download PDFInfo
- Publication number
- CN109460820A CN109460820A CN201811137897.7A CN201811137897A CN109460820A CN 109460820 A CN109460820 A CN 109460820A CN 201811137897 A CN201811137897 A CN 201811137897A CN 109460820 A CN109460820 A CN 109460820A
- Authority
- CN
- China
- Prior art keywords
- candidate region
- registration
- true value
- neural network
- sample image
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
The invention discloses a kind of neural network training method, device, computer equipment and storage mediums, using the solution of the present invention, in the training process of neural network, can obtain the corresponding prospect probability in each candidate region on sample image;Calculate the registration for the true value candidate region demarcated in candidate region and sample image;Candidate region ready for use is selected from candidate region according to prospect probability and registration, other networks after current location in neural network are trained according to candidate region ready for use, prospect probability shows a possibility that there are prospects in candidate region, registration show in candidate region comprising true value candidate region part number, so the present invention passes through the restriction to candidate region and the registration of true value candidate region while prospect probability requires, improve a possibility that there are test objects in candidate region ready for use, improve the data validity of candidate region ready for use.
Description
Technical field
The present invention relates to field of computer technology more particularly to a kind of neural network training method, device, computer equipments
And storage medium.
Background technique
Currently, in neural network such as Faster-Rcnn (Faster-Regions with Convolutional Neural
Network Features) network study during, can be in the sample image for having demarcated true value candidate region
Hundreds of candidate region is selected, then selects the area of certain amount (such as 300) from the candidate region of these substantial amounts
Domain as sample data in Faster-Rcnn region Recurrent networks and Classification Neural be trained, generally, if
It is the region selected more include the background area in sample image, and foreground area is less, then these areas selected
The data validity in domain is lower, and the speed of neural network learning is slower, is not able to satisfy user demand.
Summary of the invention
The main purpose of the embodiment of the present invention be to provide a kind of neural network training method, device, computer equipment and
Storage medium promotes the validity of candidate region data in neural network training process, promotes the pace of learning of neural network.
To achieve the above object, first aspect of the embodiment of the present invention provides a kind of neural network training method, the nerve net
Network training method includes:
In the training process of neural network, the corresponding prospect probability in each candidate region on sample image is obtained;
The registration for the true value candidate region demarcated in the candidate region and the sample image is calculated, the true value is waited
Favored area is the region on the sample image where test object;
Candidate region ready for use is selected from the candidate region according to the prospect probability and the registration;
It is trained according to other networks after current location in neural network described in the candidate region ready for use.
To achieve the above object, second aspect of the embodiment of the present invention provides a kind of neural metwork training device, the nerve net
Network training device includes:
Module is obtained, in the training process of neural network, each candidate region obtained on sample image to be corresponding
Prospect probability;
Computing module, for calculating being overlapped for the true value candidate region demarcated in the candidate region and the sample image
Degree, the true value candidate region are the region on the sample image where test object;
Selecting module, it is to be used for being selected from the candidate region according to the prospect probability and the registration
Candidate region;
Training module, for according to the candidate region ready for use to its after current location in the neural network
Its network is trained.
To achieve the above object, the third aspect of the embodiment of the present invention provides a kind of computer equipment, the computer equipment packet
Include processor, memory and communication bus;
The communication bus is for realizing the connection communication between the processor and the memory;
For storing one or more programs, the processor is used to executing to be stored the memory in the memory
One or more program, to realize such as the step of above-mentioned neural network training method.
To achieve the above object, fourth aspect of the embodiment of the present invention provides a kind of storage medium, which is stored with
One or more program, one or more of programs can be executed by one or more processor, to realize as above-mentioned
Neural network training method the step of.
The embodiment of the invention provides a kind of neural network training method, device, computer equipment and storage mediums, use
The embodiment of the present invention, in the training process of neural network, the corresponding prospect in each candidate region that can be obtained on sample image is general
Rate;Calculate the registration for the true value candidate region demarcated in candidate region and sample image;According to prospect probability and registration
Candidate region ready for use is selected from candidate region, finally according to candidate region ready for use to current location in neural network
Network later is trained.Prospect probability shows a possibility that there are prospects in candidate region, and registration shows candidate
In region comprising true value candidate region part number, so the application is on the basis of prospect probability requires by time
The restriction of the registration for the true value candidate region demarcated in favored area and sample image, further improves candidate regions ready for use
A possibility that there are test objects in domain improves the data validity for the candidate region that trained neural network uses, improves
The pace of learning of neural network.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention for those skilled in the art without creative efforts, can also basis
These attached drawings obtain other attached drawings.
Fig. 1 is a kind of flow diagram of neural network training method in the embodiment of the present invention;
Fig. 2 is a kind of structural schematic diagram of neural metwork training device in the embodiment of the present invention.
Specific embodiment
In order to make the invention's purpose, features and advantages of the invention more obvious and easy to understand, below in conjunction with the present invention
Attached drawing in embodiment, technical scheme in the embodiment of the invention is clearly and completely described, it is clear that described reality
Applying example is only a part of the embodiment of the present invention, and not all embodiments.Based on the embodiments of the present invention, those skilled in the art
Member's every other embodiment obtained without making creative work, shall fall within the protection scope of the present invention.
In the prior art, during training Faster-Rcnn network, the data validity of the candidate region used is not
Foot is unfavorable for promoting the speed of neural network learning, and the present embodiment proposes a kind of neural network training method, this method thus
The data validity of the candidate region of middle trained neural network compared with the existing technology in the data validity of candidate region obtain
Promotion is arrived, the speed of neural metwork training is faster.
Fig. 1 is combined to illustrate to how to carry out neural metwork training by taking Faster-Rcnn network as an example below, it can
It,, can be with if it needs that candidate region ready for use is selected to be trained for other types of neural network with understanding
The screening of candidate region ready for use is carried out using the scheme in the application in step 101- step 103.Referring to Fig. 1, this implementation
Example propose neural network training method include:
Step 101, in the training process of neural network, obtain sample image on the corresponding prospect in each candidate region it is general
Rate;
For the training of neural network in the present embodiment, the sample image of label information can be had to neural network inputs,
The label information shows the region known to the position of sample image, and there are a known object, (i.e. what the object is specifically
It is known, e.g. people or cat or dog etc.).
Sample image can carry out sample image by a series of convolutional layer after inputting Faster-Rcnn network
The extraction of feature maps (i.e. characteristic pattern), the number of plies of convolutional layer are arranged according to the actual needs, and the present embodiment does not have this
It limits.Such as the present embodiment can extract the feature of sample image by the conv+relu+pooling layer on one group of basis
Maps (i.e. characteristic pattern).The feature maps is shared that (Region Proposal Network, region mentions for subsequent RPN
Name network) layer and full articulamentum.
After feature maps inputs RPN network, RPN network can generate multiple candidate window anchors by sliding window method
Multiple candidate regions are obtained, and each anchors can be given a mark to obtain the prospect probability of each anchors, prospect probability exists
It being shown in the present embodiment in the corresponding candidate region the anchors there are the probability of prospect foreground, score is higher,
The probability that then there is prospect is bigger.
The present embodiment is generated candidate region and can be realized using scheme in the prior art to candidate region marking.Example
Such as in order to promote the probability that candidate window chooses prospect, the present embodiment can be introduced commonly used in detection by anchors
Multi-scale method generates the size multiple candidate regions different with length-width ratio.
Step 102, the registration for calculating the true value candidate region demarcated in candidate region and sample image, true value candidate regions
Domain is the region on sample image where test object;
In sample image, the position of the test object in true value candidate region and classification be it is known, in order to further
A possibility that there are test objects in the candidate region of training for promotion neural network also calculates candidate region and sample in the present embodiment
The registration for the true value candidate region demarcated in this image, it is to be appreciated that registration is higher, then there is inspection in candidate region
A possibility that surveying object is higher.
It, can be according to candidate regions when the registration for the true value candidate region demarcated in calculating candidate region and sample image
The coordinate of domain and true value candidate region realizes, i.e., according to the coordinate of true value candidate region and the coordinate of each candidate region come
The overlapping area for determining true value candidate region and candidate region, then obtains registration.
In one example, the registration of the present embodiment may be considered the coincidence face of true value candidate region and candidate region
Product accounts for the ratio of the sum of area of true value candidate region and candidate region, and ratio is higher, then registration is higher.
In the present embodiment, in order to reduce the training time of neural network, it may be considered that reduce the candidate regions for calculating registration
The quantity in domain.In one example, neural network that can only to the preset quantity of the condition that meets (or being not less than preset quantity)
The calculating of registration is carried out, such as selects the candidate region of preset quantity in the candidate region that prospect probability is more than some threshold value
Calculate registration.In another example, registration only can be carried out to the candidate region that prospect probability is more than some threshold value
It calculates.
In order to guarantee the data validity of candidate regions, needing from candidate region while reject too small and beyond boundary time
The process of favored area, rejecting can be after the step of calculating registration, can also be before the step of calculating registration, this reality
Example is applied to be not limited in this respect.
Step 103 selects candidate region ready for use according to prospect probability and registration from candidate region;
The candidate region that prospect probability is higher or registration is higher, it includes the probability of complete prospect is bigger, so
It, can be by limiting that there are the biggish times of the probability of complete prospect to select for prospect probability and registration in the present embodiment
Favored area.
Optionally, candidate region ready for use is selected from candidate region according to prospect probability and registration includes: root
The candidate region is carried out according to the prospect probability and the registration integrated ordered;According to ranking results from candidate region
Select the candidate region of preset quantity as candidate region ready for use.
Wherein, the quantity of the candidate region ready for use of selection is arranged according to actual needs, and the present embodiment does not limit this
It is fixed, such as selection 300,500 etc..The prospect probability and registration of comprehensive each candidate region to all candidate regions into
It, can be by the sum of the prospect probability of each candidate region and registration as the foundation to sort, or prospect is general when row sequence
Rate and registration set corresponding weight, are arranged according to the weight of prospect probability, registration and the two candidate region
Sequence, then selects candidate region ready for use, and the present embodiment does not limit this.
In another example, candidate region ready for use is selected from candidate region according to prospect probability and registration
When, candidate region can be ranked up respectively according to prospect probability and registration, select prospect probability full from candidate region
Foot and registration are all satisfied the candidate region of certain condition, such as select prospect probability not less than second default point from candidate region
Number threshold value, and registration is not less than the region of registration threshold value as candidate region ready for use etc..
Step 104 instructs other networks after current location in neural network according to candidate region ready for use
Practice.
The purpose of the present embodiment is to select the relatively high candidate region of prospect probability and registration to nerve net
Other networks after network current location are trained, and other networks here can be understood as after neural network current location
The network for needing to be trained using candidate region ready for use, including but not limited to current location in neural network after
Region Recurrent networks and Classification Neural.Wherein to neural network current location, it should be understood that selected in neural network
The network of candidate region ready for use, in Faster-Rcnn network, the current location of neural network is PPN network.
When being trained in the present embodiment according to candidate region ready for use, it can choose and utilize whole times ready for use
Favored area training, can also only use a part therein, the present embodiment does not limit this.
In one example of the present embodiment, whole candidate regions can be carried out with the calculating of registration, then according to each
The prospect probability and registration of a candidate region carry out integrated ordered (specific ordering rule can be found in above-mentioned related content),
It (such as selects a certain number of candidate regions ready for use for being selected to be arranged in front from all candidate regions according to ranking results
Select 300 be arranged in front or 500 candidate regions).
It is contemplated that the calculating meeting of registration is all carried out when the quantity of candidate region is more to all candidate regions
A large amount of calculation amount is brought, so, in order to reduce the data volume of neural computing, can only calculate and all wait in the present embodiment
The registration of a part in favored area.
Optionally, in one example, being overlapped for the true value candidate region demarcated in candidate region and sample image is calculated
Degree includes:
The candidate region of preset quantity is selected according to the sequence of the prospect probability of each candidate region;
Calculate the weight for the true value candidate region demarcated in the candidate region candidate region Zhong Ge and sample image of preset quantity
It is right.
If then selecting preceding 1000 candidate regions of prospect probabilistic for example, the quantity of candidate region is 2,000
Then domain carries out the calculating of registration with true value candidate region respectively to this 1000 candidate regions.
In one example, when calculating registration, the coordinate that can use candidate region and true value candidate region is real
It is existing.Optionally, the true value candidate region demarcated in the candidate region candidate region Zhong Ge and sample image of preset quantity is calculated
Registration includes:
Obtain each candidate in the coordinate of true value candidate region demarcated in sample image and the candidate region of preset quantity
The coordinate in region;
According to the coordinate of true value candidate region and the coordinate of the candidate region candidate region Zhong Ge of preset quantity, calculate pre-
If the registration of the candidate region candidate region Zhong Ge and true value candidate region of quantity.
For example, the candidate region to true value candidate region and respectively selected, obtain its to the coordinate of angular vertex [x1, y1,
X2, y2], wherein (x1, y1) and (x2, y2) respectively represents rectangle upper left and bottom right angular coordinate, according to true value candidate region and
Each candidate region selected determines the overlapping area of true value candidate region and candidate region, calculates overlapping area and accounts for true value candidate
The ratio of the sum of the area in region and candidate region is as registration.
Optionally, in another example, the weight for the true value candidate region demarcated in candidate region and sample image is calculated
It is right to include:
Selection prospect probability is more than the candidate region of the first preset fraction threshold value;
Calculate being overlapped for the candidate region candidate region Zhong Ge that is selected and the true value candidate region demarcated in sample image
Degree.
If then selecting prospect probability is more than 0.8 candidate region, then for example, the quantity of candidate region is 2,000
These candidate regions are carried out with true value candidate region with the calculating of registration respectively.
In one example, when calculating registration, the coordinate that can use candidate region and true value candidate region is real
It is existing.Optionally, the weight of the candidate region candidate region Zhong Ge selected and the true value candidate region demarcated in sample image is calculated
It is right to include:
Each candidate regions in the coordinate for the true value candidate region demarcated in acquisition sample image and the candidate region selected
The coordinate in domain;
According to the coordinate of true value candidate region and the coordinate of the candidate region candidate region Zhong Ge selected, calculate selected
The registration of the candidate region candidate region Zhong Ge and true value candidate region selected.
For example, the candidate region to true value candidate region and respectively selected can join in such a way that coordinate calculates registration
See above-mentioned example, details are not described herein.
Referring to fig. 2, in order to solve the technical problems existing in the prior art, the present embodiment also proposes a kind of neural network instruction
Practice device, which includes:
Module 21 is obtained, it is corresponding in the training process of neural network, obtaining each candidate region on sample image
Prospect probability;
Computing module 22, for calculating the registration for the true value candidate region demarcated in candidate region and sample image, very
Being worth candidate region is the region on sample image where test object;
Selecting module 23, for selecting candidate regions ready for use from candidate region according to prospect probability and registration
Domain;
Training module 24, for according to candidate region ready for use to other networks after current location in neural network
It is trained.
In order to reduce the quantity for the candidate region for calculating registration, optionally, in one example, computing module 22 is used
In the candidate region according to the sequence selection preset quantity of the prospect probability of each candidate region;Calculate the candidate of preset quantity
The registration for the true value candidate region demarcated in the region candidate region Zhong Ge and sample image.
Optionally, computing module 22 select to preset specifically for the sequence of the prospect probability according to each candidate region
The candidate region of quantity;It obtains in the coordinate of true value candidate region demarcated in sample image and the candidate region of preset quantity
The coordinate of each candidate region;According to the coordinate of true value candidate region and the seat of the candidate region candidate region Zhong Ge of preset quantity
Mark, calculates the registration of the candidate region candidate region Zhong Ge and true value candidate region of preset quantity.
In order to reduce the quantity for the candidate region for calculating registration, optionally, in another example, computing module 22,
It is more than the candidate region of the first preset fraction threshold value for selecting prospect probability;Calculate each candidate regions in the candidate region selected
The registration for the true value candidate region demarcated in domain and sample image.
Optionally, computing module 22 are more than the candidate region of the first preset fraction threshold value specifically for selection prospect probability;
Obtain the coordinate for the true value candidate region demarcated in sample image and the coordinate of the candidate region candidate region Zhong Ge selected;
According to the coordinate of true value candidate region and the coordinate of the candidate region candidate region Zhong Ge selected, the candidate selected is calculated
The registration of the region candidate region Zhong Ge and true value candidate region.
Optionally, selecting module 23 are specifically used for according to the prospect probability and the registration to the candidate region
It carries out integrated ordered;Select the candidate region of preset quantity as time ready for use from the candidate region according to ranking results
Favored area.
Further, the present embodiment also provides a kind of computer equipment, the computer equipment include processor, memory and
Communication bus;
Communication bus is for realizing the connection communication between processor and memory;
Memory is for storing one or more programs, and processor is for executing one or more stored in memory
Program, to realize such as the step of neural network training method in above-mentioned example.
Further, the present embodiment also provides a kind of storage medium, which is stored with one or more program,
One or more program can be executed by one or more processor, to realize such as the neural metwork training side in above-mentioned example
The step of method.
Using the scheme of the present embodiment, be finally selected from candidate region according to prospect probability and registration it is to be used
Candidate region, other networks after current location in neural network are instructed with the candidate region ready for use selected
Practice, prospect probability shows a possibility that there are prospects in candidate region, and registration shows in candidate region and waits comprising true value
The number of the part of favored area, so the present embodiment is on the basis of prospect probability requires by candidate region and sample graph
The restriction of the registration for the true value candidate region demarcated as in further improves and there is detection pair in candidate region ready for use
As a possibility that, improve the data validity for the candidate region that trained neural network uses.
In several embodiments provided herein, it should be understood that disclosed device and method can pass through it
Its mode is realized.For example, the apparatus embodiments described above are merely exemplary, for example, the division of module, only
A kind of logical function partition, there may be another division manner in actual implementation, for example, multiple module or components can combine or
Person is desirably integrated into another system, or some features can be ignored or not executed.Another point, shown or discussed is mutual
Between coupling, direct-coupling or communication connection can be through some interfaces, the INDIRECT COUPLING or communication link of device or module
It connects, can be electrical, mechanical or other forms.
Module may or may not be physically separated as illustrated by the separation member, show as module
Component may or may not be physical module, it can and it is in one place, or may be distributed over multiple networks
In module.Some or all of the modules therein can be selected to achieve the purpose of the solution of this embodiment according to the actual needs.
It should be noted that for the various method embodiments described above, describing for simplicity, therefore, it is stated as a series of
Combination of actions, but those skilled in the art should understand that, the present invention is not limited by the sequence of acts described because
According to the present invention, certain steps can use other sequences or carry out simultaneously.Secondly, those skilled in the art should also know
It knows, the embodiments described in the specification are all preferred embodiments, and related actions and modules might not all be this hair
Necessary to bright.
In the above-described embodiments, it all emphasizes particularly on different fields to the description of each embodiment, there is no the portion being described in detail in some embodiment
Point, it may refer to the associated description of other embodiments.
The above are to a kind of neural network training method provided by the present invention, device, computer equipment and storage medium
Description, for those skilled in the art, thought according to an embodiment of the present invention, in specific embodiments and applications
There will be changes, and to sum up, the contents of this specification are not to be construed as limiting the invention.
Claims (9)
1. a kind of neural network training method characterized by comprising
In the training process of neural network, the corresponding prospect probability in each candidate region on sample image is obtained;
Calculate the registration for the true value candidate region demarcated in the candidate region and the sample image, the true value candidate regions
Domain is the region on the sample image where test object;
Candidate region ready for use is selected from the candidate region according to the prospect probability and the registration;
Other networks after current location in the neural network are trained according to the candidate region ready for use.
2. neural network training method as described in claim 1, which is characterized in that it is described calculate the candidate region with it is described
The registration for the true value candidate region demarcated in sample image includes:
The candidate region of preset quantity is selected according to the sequence of the prospect probability of each candidate region;
Calculate the true value candidate region demarcated in the candidate region candidate region Zhong Ge and the sample image of the preset quantity
Registration.
3. neural network training method as claimed in claim 2, which is characterized in that the candidate for calculating the preset quantity
The registration for the true value candidate region demarcated in the region candidate region Zhong Ge and the sample image includes:
It obtains each in the coordinate for the true value candidate region demarcated in the sample image and the candidate region of the preset quantity
The coordinate of candidate region;
According to the coordinate of the true value candidate region and the coordinate of the candidate region candidate region Zhong Ge of the preset quantity, meter
Calculate the candidate region candidate region Zhong Ge of the preset quantity and the registration of the true value candidate region.
4. neural network training method as described in claim 1, which is characterized in that it is described calculate the candidate region with it is described
The registration for the true value candidate region demarcated in sample image includes:
Selecting the prospect probability is more than the candidate region of the first preset fraction threshold value;
Calculate being overlapped for the candidate region candidate region Zhong Ge that is selected and the true value candidate region demarcated in the sample image
Degree.
5. neural network training method as claimed in claim 4, which is characterized in that described to calculate in the candidate region selected
The registration for the true value candidate region demarcated in each candidate region and the sample image includes:
It obtains and is respectively waited in the coordinate and the candidate region selected for the true value candidate region demarcated in the sample image
The coordinate of favored area;
According to the coordinate of the true value candidate region and the coordinate of the candidate region candidate region Zhong Ge selected, calculate
The registration of the candidate region candidate region Zhong Ge selected and the true value candidate region.
6. neural network training method as described in any one in claim 1-5, which is characterized in that described general according to the prospect
Rate and the registration select the candidate region ready for use to include: from the candidate region
The candidate region is carried out according to the prospect probability and the registration integrated ordered;
Select the candidate region of preset quantity as candidate region ready for use from the candidate region according to ranking results.
7. a kind of neural metwork training device characterized by comprising
Module is obtained, for obtaining the corresponding prospect in each candidate region on sample image in the training process of neural network
Probability;
Computing module, for calculating the registration for the true value candidate region demarcated in the candidate region and the sample image,
The true value candidate region is the region on the sample image where test object;
Selecting module, for selecting time ready for use from the candidate region according to the prospect probability and the registration
Favored area;
Training module, for according to the candidate region ready for use to other nets after current location in the neural network
Network is trained.
8. a kind of computer equipment, which is characterized in that including processor, memory and communication bus;
The communication bus is for realizing the connection communication between the processor and the memory;
The memory is for storing one or more programs, and the processor is for executing one stored in the memory
Or multiple programs, to realize such as the step of neural network training method of any of claims 1-6.
9. a kind of storage medium, which is characterized in that the storage medium is stored with one or more program, it is one or
Multiple programs can be executed by one or more processor, to realize such as neural network of any of claims 1-6
The step of training method.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811137897.7A CN109460820A (en) | 2018-09-28 | 2018-09-28 | A kind of neural network training method, device, computer equipment and storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811137897.7A CN109460820A (en) | 2018-09-28 | 2018-09-28 | A kind of neural network training method, device, computer equipment and storage medium |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109460820A true CN109460820A (en) | 2019-03-12 |
Family
ID=65607084
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811137897.7A Pending CN109460820A (en) | 2018-09-28 | 2018-09-28 | A kind of neural network training method, device, computer equipment and storage medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109460820A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110009090A (en) * | 2019-04-02 | 2019-07-12 | 北京市商汤科技开发有限公司 | Neural metwork training and image processing method and device |
CN110263852A (en) * | 2019-06-20 | 2019-09-20 | 北京字节跳动网络技术有限公司 | Data processing method, device and electronic equipment |
CN110599503A (en) * | 2019-06-18 | 2019-12-20 | 腾讯科技(深圳)有限公司 | Detection model training method and device, computer equipment and storage medium |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106529565A (en) * | 2016-09-23 | 2017-03-22 | 北京市商汤科技开发有限公司 | Target identification model training and target identification method and device, and computing equipment |
US9858496B2 (en) * | 2016-01-20 | 2018-01-02 | Microsoft Technology Licensing, Llc | Object detection and classification in images |
US9947228B1 (en) * | 2017-10-05 | 2018-04-17 | StradVision, Inc. | Method for monitoring blind spot of vehicle and blind spot monitor using the same |
CN108416287A (en) * | 2018-03-04 | 2018-08-17 | 南京理工大学 | A kind of pedestrian detection method excavated based on omission negative sample |
CN108470172A (en) * | 2017-02-23 | 2018-08-31 | 阿里巴巴集团控股有限公司 | A kind of text information identification method and device |
-
2018
- 2018-09-28 CN CN201811137897.7A patent/CN109460820A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9858496B2 (en) * | 2016-01-20 | 2018-01-02 | Microsoft Technology Licensing, Llc | Object detection and classification in images |
CN106529565A (en) * | 2016-09-23 | 2017-03-22 | 北京市商汤科技开发有限公司 | Target identification model training and target identification method and device, and computing equipment |
CN108470172A (en) * | 2017-02-23 | 2018-08-31 | 阿里巴巴集团控股有限公司 | A kind of text information identification method and device |
US9947228B1 (en) * | 2017-10-05 | 2018-04-17 | StradVision, Inc. | Method for monitoring blind spot of vehicle and blind spot monitor using the same |
CN108416287A (en) * | 2018-03-04 | 2018-08-17 | 南京理工大学 | A kind of pedestrian detection method excavated based on omission negative sample |
Non-Patent Citations (2)
Title |
---|
王林等: ""Faster R-CNN模型在车辆检测中的应用"", 《计算机应用》 * |
程嘉晖: ""基于深度卷积神经网络的飞行器图像识别算法研究"", 《中国优秀硕士学位论文全文数据库·工程科技Ⅱ辑》 * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110009090A (en) * | 2019-04-02 | 2019-07-12 | 北京市商汤科技开发有限公司 | Neural metwork training and image processing method and device |
CN110599503A (en) * | 2019-06-18 | 2019-12-20 | 腾讯科技(深圳)有限公司 | Detection model training method and device, computer equipment and storage medium |
CN110599503B (en) * | 2019-06-18 | 2021-05-28 | 腾讯科技(深圳)有限公司 | Detection model training method and device, computer equipment and storage medium |
CN110263852A (en) * | 2019-06-20 | 2019-09-20 | 北京字节跳动网络技术有限公司 | Data processing method, device and electronic equipment |
CN110263852B (en) * | 2019-06-20 | 2021-10-08 | 北京字节跳动网络技术有限公司 | Data processing method and device and electronic equipment |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Liu et al. | Matching-cnn meets knn: Quasi-parametric human parsing | |
CN110874594B (en) | Human body appearance damage detection method and related equipment based on semantic segmentation network | |
CN109765462A (en) | Fault detection method, device and the terminal device of transmission line of electricity | |
JP2018195293A (en) | Image processing system, method for performing multi-label meaning edge detection in image, and non-transitory computer-readable storage medium | |
CN109460820A (en) | A kind of neural network training method, device, computer equipment and storage medium | |
CN106910217A (en) | Vision map method for building up, computing device, computer-readable storage medium and intelligent vehicle | |
CN111291739A (en) | Face detection and image detection neural network training method, device and equipment | |
CN110378235A (en) | A kind of fuzzy facial image recognition method, device and terminal device | |
CN105117739A (en) | Clothes classifying method based on convolutional neural network | |
CN110378297A (en) | A kind of Remote Sensing Target detection method based on deep learning | |
CN109558902A (en) | A kind of fast target detection method | |
CN104751147A (en) | Image recognition method | |
CN109117760A (en) | Image processing method, device, electronic equipment and computer-readable medium | |
JP2020503923A5 (en) | ||
CN109740668A (en) | Depth model training method and device, electronic equipment and storage medium | |
CN104992454B (en) | A kind of compartmentalization becomes the image partition method of class automatically | |
CN107944578A (en) | Seat method, apparatus, electronic equipment, storage medium are selected in online check-in with tourist site | |
CN109740752A (en) | Depth model training method and device, electronic equipment and storage medium | |
CN111027403A (en) | Gesture estimation method, device, equipment and computer readable storage medium | |
CN102147921A (en) | Graph theory-based Chinese medicinal tongue nature and tongue coat separation algorithm | |
CN105976395A (en) | Video target tracking method based on sparse representation | |
CN110263874A (en) | A kind of image classification method and device based on the study of attention relational graph | |
CN109614990A (en) | A kind of object detecting device | |
CN108052894A (en) | More attribute recognition approaches, equipment, medium and the neutral net of a kind of target object | |
CN108921162A (en) | Licence plate recognition method and Related product 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: 20190312 |