CN110189330A - A method of the background removal based on deep learning - Google Patents
A method of the background removal based on deep learning Download PDFInfo
- Publication number
- CN110189330A CN110189330A CN201910276000.7A CN201910276000A CN110189330A CN 110189330 A CN110189330 A CN 110189330A CN 201910276000 A CN201910276000 A CN 201910276000A CN 110189330 A CN110189330 A CN 110189330A
- Authority
- CN
- China
- Prior art keywords
- input
- layer
- initial stage
- network
- background removal
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/194—Segmentation; Edge detection involving foreground-background segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of methods of background removal based on deep learning.The present invention includes the following steps: step 1, the foundation to initial stage database, by using the arrangement to COCO data set and adds required for modification obtains to initial stage database;And find the label information corresponding to picture raw information matching in initial stage database;Label information to initial stage database and corresponding to it carries out pretreatment operation, including image regulation and image format conversion;The initial stage database that step 1 obtains is input in lifting rice Soviet Union Tiramisu network by step 2 as original training data, obtains the model of basic background removal;Picture to be identified is input in above-mentioned model as input data and identifies by step 3, obtains recognition result.The present invention realizes full-automatic operation, and the required picture operated need to be only input in model to the purpose that can be automatically performed background removal.
Description
Technical field
The present invention relates to image procossing, background removal field more particularly to a kind of background removals based on deep learning
Method.
Background technique
Currently, the background removal based on conventional method is widely studied both at home and abroad, but due to conventional method sheet
The limitation that body has, thus IoU (accuracy rate) reach 85 or so accuracy rate had reached saturation.Based on nerve net
The characteristic of the background removal of network application of itself in terms of AI (artificial intelligence) age rapidly developed and background removal, always
It is the topic for enduring concern to the fullest extent.In conventional direction, background removal is generally divided into several steps:
1. background modeling direction:
In terms of for picture even video background removal, the quality of background modeling has been largely fixed background removal
The quality of effect, background modeling can be analyzed from both direction: one is in the picture inputted based on picture or video itself
It is studied in prime information;Second is that being based on area information.
2. background subtraction direction:
Background subtraction is usually that the background modeling for carrying out picture to input information carries out again information used by later
The operation of processing.The problem of this link most critical, is the separation for carrying out foreground and background according to input object i.e. according to determination
Whether there is or not be greater than a scheduled threshold value to be divided for difference between present image and background image.Based on to different backgrounds
The considerations of modeling effect and different input objects, background subtraction also has different corresponding methods.Due in input information
Influence of the variation of intensity of illumination for the color of image, brightness and other information is very big, therefore for carrying out to image
For the task of background removal, can cope under varying environment intensity of illumination variation letter be the key that background removal algorithm it
Place.
Background information caused by intensity of illumination variation, which generates variation, can generally range two aspects, be on the one hand back
The part of scape changes, i.e. localized variation, is on the other hand all changed for entire background information, i.e. global change.Mainly have
Gradually changing for sunlight in such as one day is gradually changed, the cataclysm of suddenly change such as weather becomes cloudy day etc. from sunny moment
And shade generated shadow for example under illumination condition such as is shown at the three big variations in background information, is come at present
It says, gradually changes due to being that there is no illumination suddenly changes to generate strongly anti-to background information like that for generation variation step by step
Answer, therefore during background removal, the influence gradually changed is easier to be overcome, and in terms of shadow Detection this
A therefore do not generated in terms of background removal great dry with regard to handle in the separation of background and prospect
It disturbs, so in general, the sudden variation for illumination is more scabrous at present.
Summary of the invention
The object of the present invention is to provide a kind of method of background removal based on deep learning, this method can be effectively to figure
Piece achievees the purpose that full-automatic background removal, and due to the optimization characteristics of deep learning itself, for picture background removal
Precision also can guarantee that IoU is 84.6 at present, and the best IoU of classical background removal is 85, and this is most of not in loss
In the case where necessary resource, and for the method used by us, the resource of computer can not only be saved well,
Requirement can also be reached in precision simultaneously, and can the replacing background for after more effective convenient and fast solution is provided;
Meanwhile this method uses the basic framework of deep learning, to guarantee required by capable of reaching us for the background removal of picture
Precision.
In order to achieve the above objectives, the technical solution adopted by the present invention are as follows:
A method of the background removal based on deep learning, step include:
Step 1, the foundation to initial stage database, by using to COCO data set arrangement and addition modification obtain institute
Need to initial stage database;And find the label information (API) corresponding to picture raw information matching in initial stage database;
Label information to initial stage database and corresponding to it carries out pretreatment operation, including image regulation and picture format turn
It changes;
The initial stage database that step 1 obtains is input to lifting rice Soviet Union Tiramisu net as original training data by step 2
In network, the model of basic background removal is obtained, specific:
Using initial stage database as training network parameter in initial data input Tiramisu network;It is dense in the network
Block is mainly according to XL=HL([x0,x1,x2..xL-1]);Wherein HLOperation include include Batch Normalization,
ReLU and two convolution, respectively 3*3 and 1*1, x0,x1,x2..xL-1For initial data, XLFor by HLAfterwards defeated
Information out;
The network structure, which is accomplished by, is obtaining characteristic pattern output x by initial convolutional layer0, using jump connection point
For two paths, wherein main path first passes around Dense Block i.e. dense piece, and includes H in each dense pieceL;Another
Any operation is not used on path, ensure that the integrality of input information, while by x0With the output result phase in main path
Merge, original input information is optimized, removes the part of wherein redundancy, and result is used for by Transition Layer
Dimensionality reduction, simultaneously because the presence of jump connection, enables gradient more rapidly to carry out the iteration of weight;
Wherein Batch Normalization makes the characteristic mean 0 after extracting, variance 1, convenient for calculating later;
The convolution operation of 3*3 and 1*1 is also referred to as bottleneck layer, i.e., it is special to reduce input using convolution operation
Levy the quantity of figure;ReLU is for extracting sample characteristics;
Lifting rice Soviet Union network training uses supervision from below mode of learning, including an initial convolutional layer, multiple
Dense Block, that is, dense piece, multiple conversion layers, multiple jumps connection;
The initial convolutional layer is used to carry out convolution to the initial data of input, obtains original characteristic pattern;It is described dense
Block is used for feature reuse, promotes the efficiency of transmission of information and gradient in a network;The conversion layer includes reduction, is used to
Number of channels is reduced or increased, alleviates gradient and disappears;
Described each dense piece includes batch normalization (Batch Normalization), amendment linear unit (ReLU)
And convolution (Convolution) operation, the feature for dimensionality reduction and each channel of fusion;Pass through transition between each dense piece
Layer (Transition Layer) is connected, including convolution (convolution) and average pond (mean
Pooling operation);
Picture to be identified is input in above-mentioned model as input data and identifies by step 3, obtains identification knot
Fruit.
The foundation of the initial stage database refers to COCO data set, VOC pascal data set and CamVid data
Collection investigation on the basis of, according to require selection COCO data set;And the picture in the COCO data set of selection is diluted, people
To select acquisition initial stage database after the required image strong with task dependencies.
The internal structure of the Dense Block is accomplished by
It is x in output0In the case where, x first0By Batch Normalization, the normalized operation of batch is carried out,
Sample characteristics are extracted using ReLU later, the convolution operation for finally carrying out 3*3 and 1*1 obtains the input of lower layer, x0Passed through
Operation be referred to as H operation, then to x1Need first before H operation by upper one layer of output x0It is overlapped to obtain new x1
H operation is carried out again, and each layer later requires for the output of front layer to be overlapped, and carries out H operation after this;Each layer
Input from all layers of front output.
Compared with prior art, the invention has the characteristics that:
First, traditional background removal is usually semi-automatic or stingy figure manually operation, and the present invention realizes full-automatic
The required picture operated need to be only input in model the purpose that can be automatically performed background removal by operation.
Second, more generally, for the neural network network number of plies, network depth is wider, the essence of obtained model
It spends higher, but at the same time, can be transmitted due to input information and gradient information crossing between multilayer, generate gradient and disappear, ladder
The situation that degree explodes and computing resource is inadequate.Meanwhile in the increase with the network number of plies and network-wide, network can because
Precision prematurely reaches saturation and generates very fast downslide.In addition to this, since current network is all layer-by-layer transmitting information,
When the number of plies reaches certain amount, weight will appear the case where can not correcting, therefore neural network accuracy is also difficult to improve, network-wide and
The increase of depth is also just without meaning.The present invention uses Tiramisu network, is connected using Dense Block and jump
Operation, being equivalent to each layer can directly be connected with input information and gradient information, maximum journey between layers in guarantee network
Under the premise of the information transmission of degree, directly all layers are connected.Gradient disappearance is not only alleviated in this way, simultaneously because jump
The presence of connection strengthens the transmitting of profile information, and input information is more effectively utilized, reduces number of parameters.Meanwhile
It is more efficient that this connection method transmits characteristic information and gradient information, and network is also just more easier to train, computer money
Source also can rationally be utilized.In addition to this, intensively the presence of link (Dense Connection) this connection also has
The effect of regularization, in addition number of parameters mentioned above is reduced, this, which plays the phenomenon that over-fitting, inhibits to make well
With.
Third, in the direction of current convolutional neural networks improvement effect or the depth such as residual error network of intensification network
(ResNet), or widen network width such as GoogleNet Inception, Tiramisu net of the present invention
Network has then focused in the feature of input information, by construct with input information be directly connected to reach the pole to feature
It causes to utilize, decreases parameter while our institute's ideal effects to reach, parameter reduction can not only save memory, make net
Network training more rapidly, moreover it is possible to the appearance of over-fitting.
Detailed description of the invention
Fig. 1 is that the present invention is based on the flow charts of the background removal of Tiramisu network.
Fig. 2 is the overall construction drawing of Tiramisu network of the present invention.
Fig. 3 is the schematic diagram of internal structure of the Dense Block of one embodiment of the invention.Wherein number in convolutional layer bracket
Word indicates convolution kernel size
Specific embodiment
To enable features described above and advantage of the invention to be clearer and more comprehensible, special embodiment below, and institute's attached drawing is cooperated to make
Detailed description are as follows.
The present invention provides a kind of method of background removal based on deep learning, as shown in Figure 1, this method includes training rank
Section and cognitive phase;The training stage includes the following steps:
The first step builds required data set.Wherein the present invention obtains 10000 pictures as background removal task
Initial stage database;And find the corresponding label information (API) of above-mentioned raw information matching.
Second step, the label information to initial stage database and corresponding to it carry out pretreatment operation, including image canonical
Change, image format conversion etc., furthermore constrained using some priori conditions into data, with the Tiramisu after promotion
Neural network accuracy;It further illustrates and needs to pre-process test picture, i.e., (256* is normalized in image using matlab
256), to reduce due to the inconsistent brought some unnecessary errors of picture.
Third step, using 7000 pictures as training network parameter in initial data input Tiramisu network.The net
Dense piece is mainly according to X in networkL=HL([x0,x1,x2..xL-1]).Wherein HLOperation include include Batch
Normalization, ReLU and two convolution, respectively 3*3 and 1*1, x0,x1,x2..xL-1For initial data,
XLFor by HLOutput information afterwards, for the network as shown in figure 3, whole network structure is 100 layers, Fig. 2 only depicts one
Point.Characteristic pattern output x is being obtained by initial convolutional layer0, two paths are divided into using jump connection, wherein main diameter first passes around
Dense Block, that is, dense piece further includes H in each dense pieceL,;Any operation is not used on another paths, is protected
The integrality of input information is demonstrate,proved, while by x0Mutually merge with the output result in another paths, original input information is carried out
Optimization, removes the part of wherein redundancy, so that characteristic pattern is more representative, and result is passed through Transition
Layer, for dimensionality reduction and reduce parameter amount, simultaneously because jump connection presence, enable gradient more rapidly into
The iteration of row weight.Wherein Batch Normalization makes the characteristic mean 0 after extracting, variance 1, convenient for later
It calculates;The convolution operation of 3*3 and 1*1 is also referred to as bottleneck layer, i.e., is reduced using convolution operation
The quantity of input feature vector figure, thus can dimensionality reduction reduce calculation amount and merge each channel extracted feature;ReLU is used
In extraction sample characteristics.
Referring to FIG. 3, the figure is the Dense Block schematic diagram of internal structure of one embodiment of the invention.It is x in output0
In the case where, x first0By Batch Normalization, the normalized operation of batch is carried out, is extracted later using ReLU
Sample characteristics, the convolution operation for finally carrying out 3*3 and 1*1 obtain the input of lower layer, x0The operation passed through is referred to as H behaviour
Make, then to x1Need first before H operation by upper one layer of output x0It is overlapped to obtain new x1H operation is carried out again, it
Each layer afterwards requires for the output of front layer to be overlapped, and carries out H operation after this.In simple terms, each layer defeated
Enter the output from all layers of front.
Using the supervised learning mode from lower rising when training Tiramisu network, intersect entropy loss, study using standard
The RMSProp optimizer that rate is 0.001.Pretreated picture and data are first used as the first hidden layer of training is inputted (i.e. will
It is input to initial convolutional layer), first learn the parameter of the first Dense Block when training;Further, due to the limitation of network,
The constraint of sparsity constraints and priori conditions makes network structure obtain the feature for more having characterization ability than data itself;It is learning
Acquistion is to after (n-1)th layer, input by n-1 layers of output as n-th layer, thus trained n-th layer respectively obtains the ginseng of each layer
Number;100 layers of Tiramisu network is used for the present embodiment, i.e., the output of the previous hidden layer obtained study is as under
The input of one hidden layer is sequentially completed 100 layers of training, to respectively obtain the relevant parameter of each hidden layer.In addition, Tiramisu
Growth rate is 16, so we can give 16 filters of each layer of addition, is up-sampled since Tiramisu network exists, meeting
Give up some filters, so final layer is 1072 filters.
5th step saves parameter adjusted, obtains background removal model.
With continued reference to FIG. 2, the cognitive phase includes the following steps:
Test data (3000 picture) is tested the background that the above-mentioned training stage obtains as testing by the first step
It removes the accuracy rate of model and result is saved as into JPG format;
Second step carries out pretreatment operation, including image to above-mentioned primary data information (pdi) and corresponding label information
Furthermore regularization, image format conversion etc. are constrained using some priori conditions into data, after being promoted
Tiramisu neural network accuracy;It further illustrates and needs to pre-process test picture, image is subjected to normalizing using matlab
Change (256*256), to reduce due to the inconsistent brought some unnecessary errors of picture.
Above-mentioned pretreated data or test data are input in background removal model obtained above by third step
It is identified, is obtained for recognition result corresponding to required picture.
It is above to implement to be merely illustrative of the technical solution of the present invention rather than be limited, the ordinary skill people of this field
Member can be with modification or equivalent replacement of the technical solution of the present invention are made, without departing from the spirit and scope of the present invention, this hair
Bright protection scope should be subject to described in claims.
Claims (3)
1. a kind of method of the background removal based on deep learning, it is characterised in that include the following steps:
Step 1, the foundation to initial stage database, by using to COCO data set arrangement and addition modification obtain required for
To initial stage database;And find the label information (API) corresponding to picture raw information matching in initial stage database;To first
Phase database and the label information corresponding to it carry out pretreatment operation, including image regulation and image format conversion;
The initial stage database that step 1 obtains is input to lifting rice Soviet Union Tiramisu network as original training data by step 2
In, the model of basic background removal is obtained, specific:
Using initial stage database as training network parameter in initial data input Tiramisu network;Dense piece of master in the network
If according to XL=HL([x0,x1,x2..xL-1]);Wherein HLOperation include include Batch Normalization, ReLU and
Two convolution, respectively 3*3 and 1*1, x0,x1,x2..xL-1For initial data, XLFor by HLOutput letter afterwards
Breath;
The network structure, which is accomplished by, is obtaining characteristic pattern output x by initial convolutional layer0, it is divided into two using jump connection
Path, wherein main path first passes around Dense Block i.e. dense piece, and includes H in each dense pieceL;On another paths
Any operation is not used, ensure that the integrality of input information, while by x0Mutually merge with the output result in main path, it will
Original input information optimizes, and removes the part of wherein redundancy, and result is used for dimensionality reduction by Transition Layer, together
When due to jump connection presence, enable gradient more rapidly to carry out the iteration of weight;
Wherein Batch Normalization makes the characteristic mean 0 after extracting, variance 1, convenient for calculating later;3*3 with
And the convolution operation of 1*1 is also referred to as bottleneck layer, i.e., reduces input feature vector figure using convolution operation
Quantity;ReLU is for extracting sample characteristics;
Lifting rice Soviet Union network training uses supervision from below mode of learning, including an initial convolutional layer, multiple Dense
Block, that is, dense piece, multiple conversion layers, multiple jumps connection;
The initial convolutional layer is used to carry out convolution to the initial data of input, obtains original characteristic pattern;Described dense piece with
In feature reuse, the efficiency of transmission of information and gradient in a network is promoted;The conversion layer includes reduction, for reducing
Or increase number of channels, alleviate gradient and disappears;
Described each dense piece includes batch normalization (BatchNormalization), amendment linear unit (ReLU) and volume
Product (Convolution) operation, the feature for dimensionality reduction and each channel of fusion;Pass through transition zone between each dense piece
(Transition Layer) is connected, including convolution (convolution) and averagely pond (mean pooling)
Operation;
Picture to be identified is input in above-mentioned model as input data and identifies by step 3, obtains recognition result.
2. a kind of method of background removal based on deep learning according to claim 1, it is characterised in that the initial stage
The foundation of database refer to COCO data set, VOC pascal data set and CamVid data set investigate on the basis of,
According to require selection COCO data set;And the picture in the COCO data set of selection is diluted, artificially select it is required with
Initial stage database is obtained after the strong image of task dependencies.
3. a kind of method of background removal based on deep learning according to claim 1 or 2, it is characterised in that described
The internal structure of Dense Block is accomplished by
It is x in output0In the case where, x first0By Batch Normalization, the normalized operation of batch is carried out, later
Sample characteristics are extracted using ReLU, the convolution operation for finally carrying out 3*3 and 1*1 obtains the input of lower layer, x0The behaviour passed through
Make to be referred to as H operation, then to x1Need first before H operation by upper one layer of output x0It is overlapped to obtain new x1Again into
Row H operation, each layer later require for the output of front layer to be overlapped, and carry out H operation after this;Each layer defeated
Enter the output from all layers of front.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910276000.7A CN110189330A (en) | 2019-04-08 | 2019-04-08 | A method of the background removal based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910276000.7A CN110189330A (en) | 2019-04-08 | 2019-04-08 | A method of the background removal based on deep learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110189330A true CN110189330A (en) | 2019-08-30 |
Family
ID=67713778
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910276000.7A Pending CN110189330A (en) | 2019-04-08 | 2019-04-08 | A method of the background removal based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110189330A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111368802A (en) * | 2020-03-28 | 2020-07-03 | 河南工业职业技术学院 | Material shape recognition method based on binocular image vision |
CN113011556A (en) * | 2021-02-20 | 2021-06-22 | 安徽大学 | Method for establishing network identification model based on INC-DenseUnet |
CN113221814A (en) * | 2021-05-26 | 2021-08-06 | 华瑞新智科技(北京)有限公司 | Road traffic sign identification method, equipment and storage medium |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106650744A (en) * | 2016-09-16 | 2017-05-10 | 北京航空航天大学 | Image object co-segmentation method guided by local shape migration |
CN106991386A (en) * | 2017-02-27 | 2017-07-28 | 杭州电子科技大学 | A kind of gesture identification method based on depth residual error network |
-
2019
- 2019-04-08 CN CN201910276000.7A patent/CN110189330A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106650744A (en) * | 2016-09-16 | 2017-05-10 | 北京航空航天大学 | Image object co-segmentation method guided by local shape migration |
CN106991386A (en) * | 2017-02-27 | 2017-07-28 | 杭州电子科技大学 | A kind of gesture identification method based on depth residual error network |
Non-Patent Citations (2)
Title |
---|
GAO HUANG ET AL: "DENSELY CONNECTED CONVOLUTIONAL NETWORKS", 《ARXIV:1608.06993V1》 * |
SIMON JE´GOU ET AL: "The One Hundred Layers Tiramisu:Fully Convolutional DenseNets for Semantic Segmentation", 《ARXIV:1611.09326V3》 * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111368802A (en) * | 2020-03-28 | 2020-07-03 | 河南工业职业技术学院 | Material shape recognition method based on binocular image vision |
CN113011556A (en) * | 2021-02-20 | 2021-06-22 | 安徽大学 | Method for establishing network identification model based on INC-DenseUnet |
CN113011556B (en) * | 2021-02-20 | 2022-10-11 | 安徽大学 | Method for establishing network identification model based on INC-DenseUnet |
CN113221814A (en) * | 2021-05-26 | 2021-08-06 | 华瑞新智科技(北京)有限公司 | Road traffic sign identification method, equipment and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111882002B (en) | MSF-AM-based low-illumination target detection method | |
CN108804397B (en) | Chinese character font conversion generation method based on small amount of target fonts | |
CN110298321B (en) | Road blocking information extraction method based on deep learning image classification | |
CN110189330A (en) | A method of the background removal based on deep learning | |
CN113313657A (en) | Unsupervised learning method and system for low-illumination image enhancement | |
CN107808132A (en) | A kind of scene image classification method for merging topic model | |
CN106250931A (en) | A kind of high-definition picture scene classification method based on random convolutional neural networks | |
CN108711141A (en) | The motion blur image blind restoration method of network is fought using improved production | |
CN108847223A (en) | A kind of audio recognition method based on depth residual error neural network | |
CN110991362A (en) | Pedestrian detection model based on attention mechanism | |
CN105574827A (en) | Image defogging method and device | |
CN108256482A (en) | A kind of face age estimation method that Distributed learning is carried out based on convolutional neural networks | |
CN111861906B (en) | Pavement crack image virtual augmentation model establishment and image virtual augmentation method | |
CN106528826A (en) | Deep learning-based multi-view appearance patent image retrieval method | |
CN108364278B (en) | Rock core fracture extraction method and system | |
CN111627080B (en) | Gray level image coloring method based on convolution nerve and condition generation antagonistic network | |
CN110245620B (en) | Non-maximization inhibition method based on attention | |
CN109740451A (en) | Road scene image semantic segmentation method based on importance weighting | |
CN108985457A (en) | A kind of deep neural network construction design method inspired by optimization algorithm | |
CN113569881A (en) | Self-adaptive semantic segmentation method based on chain residual error and attention mechanism | |
CN110503149A (en) | Local feature classification method and system in a kind of image | |
CN110135501A (en) | High dynamic range images evidence collecting method based on neural network framework | |
US20200349416A1 (en) | Determining computer-executed ensemble model | |
CN110807369A (en) | Efficient short video content intelligent classification method based on deep learning and attention mechanism | |
CN112836755B (en) | Sample image generation method and system 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: 20190830 |