CN110276397A - A kind of image processing method based on door machine, device and system - Google Patents
A kind of image processing method based on door machine, device and system Download PDFInfo
- Publication number
- CN110276397A CN110276397A CN201910547952.8A CN201910547952A CN110276397A CN 110276397 A CN110276397 A CN 110276397A CN 201910547952 A CN201910547952 A CN 201910547952A CN 110276397 A CN110276397 A CN 110276397A
- Authority
- CN
- China
- Prior art keywords
- feature
- layer
- layers
- neural network
- information
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
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
A kind of methods, devices and systems of the image procossing based on door machine, wherein this method at least includes the following steps: S10, obtains the corresponding door G of each layer of feature X of neural network；S20 enhances its useful information using door G to each layer of neural network of feature X, while calculating the region that this feature X has garbage；S30 supplements the region in each layer of feature X with garbage using the useful information of other layers of feature；S40 is connected all layers of neural network of feature entirely.The information that useful information inhibits useless can be screened by door machine system, the mode connected entirely can make all features carry out information interchange between any two to make the information of all features all tool different levels.Therefore enough useful informations can be obtained from the feature of different levels when doing last Fusion Features and introduce invalid even harmful information without worrying.
Description
Technical field
The present invention relates to technical field of machine vision, more particularly to a kind of image procossing based on door machine method,
Device and system.
Background technique
In deep learning, the target of image segmentation is classified for each pixel in the picture, this is just needed
The feature that there is high-resolution, high-level semantic information.Existing convolutional neural networks are obtained when doing image segmentation
High-level semantic feature needs sufficiently large receptive field, and the simplest method for obtaining big receptive field is exactly continuous
Down-sampling is carried out, this, which has resulted in its resolution ratio when the semantic information that feature has higher level, can become very low, with this
There is high-resolution low level feature in the relatively shallow-layer of convolutional neural networks simultaneously, but its semantic information is weaker.Cause
This is needed generally for high-resolution, the feature of high-level semantic information is obtained by the Fusion Features of different levels.
The common method of Fusion Features is successively to merge feature to low level from high-level, is then directly given a forecast
Or by the Fusion Features of all different levels to giving a forecast again later together, corresponding representative work is U-
Net and FPN (feature pyramid network).Since the focus of the feature of different levels is different, compare the feature concern of shallow hierarchy
Detailed information and high-level feature pays close attention to semantic information, have bigger difference between different information.U-Net and FPN this two
Kind of technology is fused together feature is indiscriminate, causes the introducing of garbage to lose original useful letter
Breath.Therefore, all information is subject to being introduced into all features for selection can be highly beneficial for last prediction.
Summary of the invention
Present invention mainly solves be exactly feature selecting and Fusion Features problem in deep learning.Existing technology is all
It is by feature from high-level to the indiscriminate direct fusion of low level, and be first will be special with door machine system for improvement of the invention
Useful information extracts in sign, is then again carried out useful information between different features in a manner of connecting entirely
Distribution supplement is to achieve the purpose that Fusion Features, finally again by all Fusion Features to giving a forecast together.Door machine system can be born
Duty screens useful information and inhibits useless information, and the mode connected entirely can make the letter of all feature progress between any two
Breath exchange has the information of different levels to make all features all.It therefore can be from when doing last Fusion Features
The feature of different levels obtains enough useful informations and introduces invalid or even harmful information without worrying.
The purpose of the present invention is to provide a kind of methods, devices and systems of image procossing based on door machine, specific skills
Art scheme is as follows:
In a first aspect, the embodiment of the invention provides a kind of methods of image procossing based on door machine, comprising:
S10 obtains the corresponding door of feature of each layer of neural network；
S20 enhances its useful information using door, while calculating this feature to have to each layer of feature of neural network
The region of garbage；
S30 mends the region in each layer of feature with garbage using the useful information of other layers of feature
It fills；
S40 is connected all layers of neural network of feature entirely.
Second aspect, the embodiment of the invention provides a kind of image processing apparatus based on door machine, comprising:
Door obtains module, for obtaining each layer of the corresponding door of feature of neural network；
Enhancing module enhances its useful information using door, and based on for each layer of the feature to neural network
Calculate the region that this feature has garbage；
Complementary module, for in each layer of feature with garbage region, using the useful letter of other layers of feature
Breath is supplemented；
Full link block, for being connected all layers of neural network of feature entirely.
The third aspect, the present invention also provides a kind of image processing systems based on door machine, including memory and processing
Device, memory store instruction；Processor unit is used to execute following steps according to instruction stored in memory:
S10 obtains the corresponding door of feature of each layer of neural network；
S20 enhances its useful information using door, while calculating this feature to have to each layer of feature of neural network
The region of garbage；
S30 mends the region in each layer of feature with garbage using the useful information of other layers of feature
It fills；
S40 is connected all layers of neural network of feature entirely.
Technical solution based on the application first passes through and is extracted information useful in feature with door machine system, then again
Useful information is distributed between different features in a manner of connecting entirely and is supplemented to achieve the purpose that Fusion Features,
Finally again by all Fusion Features to giving a forecast together.Door machine system can be responsible for screening the letter that useful information inhibits useless
Breath, the mode connected entirely can make all features carry out information interchange between any two to make all features all have use
The information of different levels.Therefore it can be obtained from the feature of different levels when doing last Fusion Features enough
Useful information introduces invalid even harmful information without worrying.
Detailed description of the invention
In order to illustrate the technical solution of the embodiments of the present invention more clearly, required use in being described below to embodiment
Attached drawing be briefly described, it should be apparent that, drawings in the following description are only some embodiments of the invention, for
For those of ordinary skill in the art, without creative efforts, it can also be obtained according to these attached drawings other
Attached drawing.
Fig. 1 is the flow chart of the image processing method of the embodiment of the present invention；
Fig. 2 is the specific algorithm process of the useful information of every layer of feature of acquisition of the embodiment of the present invention.
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.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts it is all its
His embodiment, shall fall within the protection scope of the present invention.
To solve the problems, such as that the feature selecting and Fusion Features in deep learning, the embodiment of the present application provide one kind and be based on
The image processing method and device of door machine.Based on door machine system, the information that useful information inhibits useless can be screened；And it is complete
The mode of connection can make all features carry out information interchange between any two to make all features all tool different layers
Secondary information.Therefore, the embodiment of the present invention first passes through door machine system and extracts useful information, then again to connect entirely
Useful information is distributed supplement to achieve the purpose that Fusion Features, finally again by institute by mode between different features
Some Fusion Features to being predicted together.
A kind of image processing method based on door machine is provided for the embodiments of the invention first below to be introduced.
It should be noted that provided by the embodiment of the present invention it is a kind of based on the image processing method of door machine by a kind of figure
As performed by processing unit, wherein the image processing apparatus can be independent image processing software in the related technology, can also
Think the feature card in image processing software；In addition, the image processing apparatus can be applied in electronic equipment, the electronics
Equipment is terminal device and/or server.
As shown in Figure 1, the embodiment of the invention provides a kind of image processing methods based on door machine, including walk as follows
It is rapid:
S10 obtains the corresponding door of feature of each layer of neural network.
In one embodiment, it is corresponding that each layer of neural network of feature X can be obtained by gate function Sigmoid
Gating feature G.Firstly, introducing gate function Sigmoid.Gate function Sigmoid is the real-number function that a bounded can be micro-, definition
It is as follows:
The output area of Sigmoid function is (0,1), and output can indicate probability, confidence level etc..In order to reduce useless letter
The influence to Fusion Features is ceased, to reduce garbage as far as possible when carrying out Fusion Features, it is necessary to before Fusion Features
Whether the information in first judging characteristic useful, in the embodiment of the present application Selection utilization gate function Sigmoid as feature whether
Useful discriminant function.Specific implementation is using feature X as the input of function Sigmoid, the output of the function, i.e. door G
It is gating feature of the codomain between (0,1), gating feature G is the confidence level for illustrating input feature vector X, works as confidence
When spending higher, it is believed that the information in this feature X is more useful.
In this way, for L feature X of different levels in neural network1, X2... XL, gate function can be passed through
Sigmoid obtains its corresponding L gating feature G1, G2... GL。
S20 enhances its useful information using door, while calculating this feature to have to each layer of feature of neural network
The region of garbage；
G is that each layer of feature X is inputted to the gating feature obtained after gate function Sigmoid, while G ∈ (0,1), this Shen
Please in useful information in feature, and the garbage in inhibitory character are extracted by X*G.Wherein, in each layer of feature
Useful information can be used to supplement the feature of other layers.
Specifically, X*G is to be weighted using gating feature G to feature X.Since gating feature G and feature X are in space
On be one-to-one, the lower point of numeric ratio in gating feature G, in feature X the confidence level of corresponding points also just compared with
It is low, i.e. the information of corresponding points be it is useless, the numerical value of the point can be made to further decrease using X*G, i.e., inhibition garbage；
And the higher point of numeric ratio in gating feature G, the confidence level of corresponding points is also just higher in feature X, i.e. the letter of corresponding points
Breath be it is useful, using X*G then the numerical value of the point will not reduce excessive.Therefore, having in feature can be extracted using X*G
With the garbage in information inhibitory character.
For n-th (n is the positive integer less than or equal to L) layer feature XnFor, it is necessary first to reinforce the useful of oneself itself
Information uses Xn+GnWeighting scheme calculated.It is lost due to may cause information after being weighted, in the present embodiment
Carry out compensated information using the mode that residual error connects to lose, that is, uses Xn+(1+Gn) weighting scheme calculated.That is,
For each feature Xn, utilize Xn+(1+Gn) enhance its useful information.
Meanwhile every layer of feature XnAccording to its GnUseful information is sent to other layers, every layer of feature all with other layers of spy
Sign is connected, with the useful information that complements each other between the layers.#
For n-th (n is the positive integer less than or equal to L) layer feature XnFor, after the useful information for reinforcing oneself itself, also
Its useless information can be supplemented.And want complementary features XnIn useless information it is necessary to first obtaining its area with garbage
Domain.In the present embodiment, X can be passed throughn+(1-Gn) obtain feature XnRegion with garbage.
S30 mends the region in each layer of feature with garbage using the useful information of other layers of feature
It fills；
It can be using the useful information in all other layer of feature come to this layer of feature X in the present embodimentnIn have nothing
It is supplemented with the region of information, to obtain the enhancing of the useful information in the region with garbage.
Firstly, also they are carried out in order not to interfere the garbage in other layers of feature to the feature of this layer
Weighting uses Xn*GnMode be weighted.And you need to add is that XnThe information in middle garbage region, so its
The information that its layer feature passes over all have passed through 1-G againnWeighting so that information supplement is to the region required supplementation with.It is i.e. special
Levy XnEnhanced information is passed through in middle garbage region are as follows:
(X1*G1+Xz*Gz+Xn-1*Gn-1+Xn+1*Gn+1…XL*GL)*(1-Gn)
That is, the feature X extracted for every layernFor, mode through this embodiment is obtained using door machine system
Useful information may be expressed as:
Xn*(1+Gn)+(X1*G1+X2*G2+Xn1*Gn1+Xn|1*Gn|1…+XL*GL)*(1-Gn)
Fig. 2 shows the specific algorithm processes according to an embodiment of the present invention that useful information is obtained based on door machine system.
By the above-mentioned means, to each layer of feature Xn, its useful information can be enhanced, and inhibit useless information.
S40 is connected all layers of neural network of feature entirely.
All layers of feature is carried out to the operation of step S20-S30, that is, by all layers of feature all with from it
He supplements at the useful information of layer, i.e., is connected all layers of feature entirely.By useful letter in a manner of connecting entirely
Breath is distributed supplement to achieve the purpose that Fusion Features, finally again by all Fusion Features between the feature of different layers
To giving a forecast together.
For one layer of feature XnFor, the useful information in other layers of feature is collected to oneself, while also by oneself
Useful information pass to other layers of feature so that useful information is sufficiently exchanged between the layers.
The mode connected entirely can make all layers of feature carry out information interchange between any two to make all layers of spy
Sign all tools information of different levels.Therefore it can be obtained from the feature of different levels when doing last Fusion Features
Enough useful informations introduce invalid even harmful information without worrying.
Therefore, the application utilizes door machine system that can retain useful information, inhibit useless information, while obtaining and needing to mend
The region of information is filled, and information is subjected to exchange between any two by way of connecting entirely, it can be from the feature of different levels
It obtains enough useful informations and introduces invalid even harmful information without worrying, to utilize each feature itself
Useful information goes to supplement other features, so that each feature tool information of different levels, reaches the mesh of Fusion Features
's.
Corresponding to above method embodiment, the embodiment of the invention also provides a kind of image processing apparatus, comprising:
Door obtains module, for obtaining each layer of the corresponding door of feature of neural network；
Enhancing module enhances its useful information using door, and based on for each layer of the feature to neural network
Calculate the region that this feature has garbage；
Complementary module, for in each layer of feature with garbage region, using the useful letter of other layers of feature
Breath is supplemented；
Full link block, for being connected all layers of neural network of feature entirely.
The present invention also provides a kind of image processing systems based on door machine, including memory and processor.
Wherein, memory is applied for storage, is instructed, module and data, processing unit are stored in by operation and store list
Application, instruction, module and data in member, thereby executing various function application (such as the image segmentation of the invention of client
Device) and data processing.Storage unit mainly includes application memory area and data storage area, wherein the storage of application memory area
Operating system, application software (such as sound playout software, image playout software) etc.；Data storage area storage client makes
With the data (such as audio data, video data, phone directory) etc. created.Memory includes high-speed random access memory,
It can also include nonvolatile memory, for example, at least a disk memory, flush memory device or other volatile solid-states
Memory device.
Processing unit is the control centre of client, for executing the application software being stored in storage unit and/or mould
Block, and the data being stored in memory are called, execute the various functions and processing data of client.
In addition, client can also include camera, microphone, bluetooth module, sensor, power supply etc., no longer go to live in the household of one's in-laws on getting married herein
It states.
In embodiments of the present invention, memory store instruction；Processor unit is used for according to finger stored in memory
It enables, executes following steps:
S10 obtains the corresponding door of feature of each layer of neural network；
S20 enhances its useful information using door, while calculating this feature to have to each layer of feature of neural network
The region of garbage；
S30 mends the region in each layer of feature with garbage using the useful information of other layers of feature
It fills；
S40 is connected all layers of neural network of feature entirely.
In conclusion technical scheme can retain useful information using door machine system, inhibit useless information, together
When obtain and require supplementation with the region of information, and information is subjected to exchange between any two by way of connecting entirely, can be never
The feature of same level obtains enough useful informations and introduces invalid even harmful information without worrying, thus using every
The useful information of a feature itself goes to supplement other features, so that each feature tool information of different levels, reaches special
Levy the purpose of fusion.
It is apparent to those skilled in the art that for convenience and simplicity of description, each mould of foregoing description
Block, the specific work process respectively instructed, can refer to corresponding processes in the foregoing method embodiment, details are not described herein.
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 the module,
Only a kind of logical function partition, there may be another division manner in actual implementation, such as multiple units or components can be with
In conjunction with or be desirably integrated into another system, or some features can be ignored or not executed.Another point, it is shown or discussed
Mutual coupling, direct-coupling or communication connection can be through some interfaces, the INDIRECT COUPLING of device or unit
Or communication connection, it can be electrical property, mechanical or other forms.
It, can also be in addition, the functional units in various embodiments of the present invention may be integrated into one processing unit
It is that each unit physically exists alone, can also be integrated in one unit with two or more units.Above-mentioned integrated list
Member both can take the form of hardware realization, can also realize in the form of software functional units.
If the integrated unit is realized in the form of SFU software functional unit and sells or use as independent product
When, it can store in a computer readable storage medium.Based on this understanding, technical solution of the present invention essence
On all or part of the part that contributes to existing technology or the technical solution can be with the shape of software product in other words
Formula embodies, which is stored in a storage medium, including some instructions are used so that a calculating
Machine equipment (can be personal computer, server or the network equipment etc.) executes each embodiment the method for the present invention
All or part of the steps.And storage medium above-mentioned includes: USB flash disk, mobile hard disk, read-only memory (ROM, Read-Only
Memory), random access memory (RAM, Random Access Memory), magnetic or disk etc. are various can store journey
The medium of sequence code.
The above, the above embodiments are merely illustrative of the technical solutions of the present invention, rather than its limitations；Although referring to before
Stating embodiment, invention is explained in detail, those skilled in the art should understand that: it still can be to preceding
Technical solution documented by each embodiment is stated to modify or equivalent replacement of some of the technical features；And these
It modifies or replaces, the spirit and scope for technical solution of various embodiments of the present invention that it does not separate the essence of the corresponding technical solution.
Claims (10)
1. a kind of method of the image procossing based on door machine, at least includes the following steps:
S10 obtains the corresponding door G of each layer of feature X of neural network；
S20 enhances its useful information using door G to each layer of neural network of feature X, while calculating this feature X with nothing
With the region of information；
S30 supplements the region in each layer of feature X with garbage using the useful information of other layers of feature；
S40 is connected all layers of neural network of feature entirely.
2. step S10 includes according to the method described in claim 1, wherein,
The corresponding door G of each layer of feature X of neural network is obtained by gate function Sigmoid.
3. method according to claim 1 or 2, wherein step S20 includes,
The useful information in feature is extracted by X*G, and the garbage in inhibitory character.
4. method according to any one of claim 1-3, wherein step S20 includes,
For each layer in L layers of neural network of feature Xn, utilize Xn*(1+Gn) enhancing its useful information, wherein n is small
In the positive integer for being equal to L.
5. method according to any of claims 1-4, wherein step S20 includes,
For each layer in L layers of neural network of feature Xn, pass through Xn*(1-Gn) obtain feature XnArea with garbage
Domain, wherein n is the positive integer less than or equal to L.
6. method according to any one of claims 1-5, wherein step 30 includes,
For each layer in L layers of neural network of feature XnIn with garbage region, using the useful letter of other layers of feature
Cease the information after being supplemented are as follows:
(X1*G1+X2*G2…+Xn-1*Gn-1+Xn+1*Gn+1…+XL*GL)*(1-Gn), wherein n is the positive integer less than or equal to L.
7. method according to claim 1 to 6, wherein step 30 includes,
For each layer in L layers of neural network of feature Xn, it is expressed as using the useful information that door machine system obtains:
Xn*(1+Gn)+(X1*G1+X2*G2+Xn-1*Gn-1+Xn+1*Gn+1…+XL*GL)*(1-Gn), wherein n is just less than or equal to L
Integer.
8. -7 method according to claim 1, wherein step S40 includes,
After all layers in L layers of neural network of feature is carried out the operation of step S20-S30, all layers of feature is all used to
It is supplemented from the useful information of other layers.
9. a kind of image processing apparatus based on door machine, comprising:
Door obtains module, for obtaining each layer of the corresponding door of feature of neural network；
Enhance module, for each layer of the feature to neural network, enhances its useful information using door, and for calculating this
Feature has the region of garbage；
Complementary module, for in each layer of feature with garbage region, using other layers of feature useful information into
Row supplement；
Full link block, for being connected all layers of neural network of feature entirely.
10. a kind of image processing system based on door machine, including memory and processor, memory store instruction；Processor
Unit is used to execute following steps according to instruction stored in memory:
S10 obtains the corresponding door of feature of each layer of neural network；
S20 enhances its useful information using door to each layer of feature of neural network, while calculating this feature with useless
The region of information；
S30 supplements the region in each layer of feature with garbage using the useful information of other layers of feature；
S40 is connected all layers of neural network of feature entirely.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910547952.8A CN110276397B (en) | 2019-06-24 | 2019-06-24 | Door mechanism-based image feature extraction method, device and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910547952.8A CN110276397B (en) | 2019-06-24 | 2019-06-24 | Door mechanism-based image feature extraction method, device and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110276397A true CN110276397A (en) | 2019-09-24 |
| CN110276397B CN110276397B (en) | 2021-03-09 |
Family
ID=67961512
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910547952.8A Active CN110276397B (en) | 2019-06-24 | 2019-06-24 | Door mechanism-based image feature extraction method, device and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110276397B (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH04184686A (en) * | 1990-11-20 | 1992-07-01 | Canon Inc | Pattern recognizing device |
| CN102957929A (en) * | 2011-08-22 | 2013-03-06 | 索尼公司 | Video signal processing apparatus, video signal processing method, and computer program |
| CN106060372A (en) * | 2015-04-02 | 2016-10-26 | 安讯士有限公司 | Method and system for image stabilization |
| CN107341462A (en) * | 2017-06-28 | 2017-11-10 | 电子科技大学 | A kind of video classification methods based on notice mechanism |
| CN108389224A (en) * | 2018-02-26 | 2018-08-10 | 北京市商汤科技开发有限公司 | Image processing method and device, electronic equipment and storage medium |
| CN108710902A (en) * | 2018-05-08 | 2018-10-26 | 江苏云立物联科技有限公司 | A kind of sorting technique towards high-resolution remote sensing image based on artificial intelligence |
-
2019
- 2019-06-24 CN CN201910547952.8A patent/CN110276397B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH04184686A (en) * | 1990-11-20 | 1992-07-01 | Canon Inc | Pattern recognizing device |
| CN102957929A (en) * | 2011-08-22 | 2013-03-06 | 索尼公司 | Video signal processing apparatus, video signal processing method, and computer program |
| CN106060372A (en) * | 2015-04-02 | 2016-10-26 | 安讯士有限公司 | Method and system for image stabilization |
| CN107341462A (en) * | 2017-06-28 | 2017-11-10 | 电子科技大学 | A kind of video classification methods based on notice mechanism |
| CN108389224A (en) * | 2018-02-26 | 2018-08-10 | 北京市商汤科技开发有限公司 | Image processing method and device, electronic equipment and storage medium |
| CN108710902A (en) * | 2018-05-08 | 2018-10-26 | 江苏云立物联科技有限公司 | A kind of sorting technique towards high-resolution remote sensing image based on artificial intelligence |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110276397B (en) | 2021-03-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US10733431B2 (en) | Systems and methods for optimizing pose estimation | |
| US11966839B2 (en) | Auto-regressive neural network systems with a soft attention mechanism using support data patches | |
| KR101803471B1 (en) | Deep learning system and learning method using of convolutional neural network based image patterning | |
| CN110023964B (en) | Training and/or using neural network models to generate intermediate outputs of spectral images | |
| Doulamis et al. | FAST-MDL: Fast Adaptive Supervised Training of multi-layered deep learning models for consistent object tracking and classification | |
| CN109313490A (en) | It is tracked using the eye gaze of neural network | |
| CN109964237A (en) | Picture depth prediction neural network | |
| WO2019108251A1 (en) | Optimizations for dynamic object instance detection, segmentation, and structure mapping | |
| CN112149637B (en) | Method and device for generating a target re-recognition model and for target re-recognition | |
| GB2596370A (en) | Model training method and apparatus, and prediction method and apparatus | |
| DE102022119386A1 (en) | METHOD AND APPARATUS FOR PERFORMING DENSE PREDICTION USING TRANSFORMER BLOCKS | |
| CN112232165B (en) | Data processing method, device, computer and readable storage medium | |
| CN112215171A (en) | Target detection method, device, equipment and computer readable storage medium | |
| WO2024041108A1 (en) | Image correction model training method and apparatus, image correction method and apparatus, and computer device | |
| US20240104403A1 (en) | Method for training click rate prediction model | |
| JP2023533907A (en) | Image processing using self-attention-based neural networks | |
| CN115169548A (en) | Tensor-based continuous learning method and device | |
| CN112116700B (en) | Monocular view-based three-dimensional reconstruction method and device | |
| CN117611697A (en) | Image processing method, apparatus, device, storage medium, and program product | |
| CN117671254A (en) | Image segmentation method and device | |
| CN112131429A (en) | Video classification method and system based on depth prediction coding network | |
| CN110276397A (en) | A kind of image processing method based on door machine, device and system | |
| CN113902649B (en) | Image restoration method and device, equipment, medium and product thereof | |
| Liu et al. | Knowledge-guided multi-perception attention network for image dehazing | |
| CN113362409B (en) | Image coloring and model training method and device, electronic equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| TR01 | Transfer of patent right | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20220729 Address after: Room 618, 6 / F, building 5, courtyard 15, Kechuang 10th Street, Beijing Economic and Technological Development Zone, Daxing District, Beijing 100176 Patentee after: Xiaomi Automobile Technology Co.,Ltd. Address before: 100080 soho1219, Zhongguancun, 8 Haidian North 2nd Street, Haidian District, Beijing Patentee before: SHENDONG TECHNOLOGY (BEIJING) Co.,Ltd. |