CN110378398A - A kind of deep learning network improvement method based on the jump fusion of Analysis On Multi-scale Features figure - Google Patents
A kind of deep learning network improvement method based on the jump fusion of Analysis On Multi-scale Features figure Download PDFInfo
- Publication number
- CN110378398A CN110378398A CN201910566224.1A CN201910566224A CN110378398A CN 110378398 A CN110378398 A CN 110378398A CN 201910566224 A CN201910566224 A CN 201910566224A CN 110378398 A CN110378398 A CN 110378398A
- Authority
- CN
- China
- Prior art keywords
- fusion
- jump
- analysis
- convolutional layer
- layer
- 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
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of deep learning network improvement methods based on the jump fusion of Analysis On Multi-scale Features figure, Fusion Features are carried out by the jump connection between Analysis On Multi-scale Features figure layer, it enables the network to make full use of high low-level feature by fusion high-level semantic and low layer location information, model is improved to the sensibility and perceptibility of Small object, while improving model totality detection performance.Secondly by the more classification policies of multi-angle of view, the accurate detection of target category under high dynamic scene is realized.The present invention considers from speed, practical, robustness etc., proposes a kind of deep learning network improvement method based on the jump fusion of Analysis On Multi-scale Features figure, improves detection performance of the SSD algorithm under high dynamic scene.
Description
Technical field
The present invention relates to a kind of deep learning network improvement methods based on the jump fusion of Analysis On Multi-scale Features figure, belong to target
Detection technique field.
Background technique
Deep neural network structure includes multiple feature extraction operation, and every to pass through one layer of convolution operation, network layer is deeper,
The profile and detailed information of characteristic pattern are fewer, and semantic information just becomes richer, and the receptive field of model is also bigger.Model
Understand and has removed concern biggish object in the picture, and for Small object, the identification capability of model is with regard to poor.In target detection
One technological difficulties is small target deteection, and former SSD algorithm (the more frame detectors of single-shot) is performed poor.
The quick visual angle change of vehicle-carried mobile platform causes target detection to be easy to appear missing inspection erroneous detection simultaneously.
Summary of the invention
Goal of the invention: the present invention proposes a kind of deep learning network improvement side based on the jump fusion of Analysis On Multi-scale Features figure
Method improves detection performance of the SSD algorithm under small target deteection and high dynamic scene.
Technical solution: the technical solution adopted by the present invention is a kind of deep learning based on the jump fusion of Analysis On Multi-scale Features figure
Network improvement method, comprising the following steps:
Construct the Fusion Features network based on convolutional layer;
Design feature fusion connection module;
Convergence strategy and up-sampling mode are selected, is obtained based on SSD Analysis On Multi-scale Features figure layer jump fusion structure;
Incorporate the above-mentioned scale feature figure layer jump fusion structure of multi-angle of view Strategies Training.
The Fusion Features network is characterized the connection of figure layer jump.
The characteristic pattern layer jump connection successively includes the four or three fusion convolutional layer, the full articulamentum of the 7th fusion, the 6th
Two fusion convolutional layers, the seven or two fusion convolutional layer, the eight or two convolutional layer and the 9th 2 convolutional layer.
The Fusion Features link block first up-samples high-level characteristic figure, the high-level characteristic after being up-sampled
Figure.Low-level feature after obtaining dimensionality reduction after 1 × 1 convolution kernel dimensionality reduction and the activation of line rectification function using low-level feature figure
Figure.Then carry out Fusion Features operation, i.e. splicing or element summation, the high low layer characteristic pattern after obtaining splicing/element summation, most
Afterwards using 3 × 3 convolution kernel convolution algorithms to reduce aliasing effect, then line rectification function is activated, after obtaining fusion completely
High low layer characteristic pattern.
The convergence strategy is first element summation, then carries out batch normalization.
The up-sampling mode is bilinear interpolation.
The utility model has the advantages that the present invention carries out Fusion Features by the jump connection between Analysis On Multi-scale Features figure layer, it is high by fusion
Layer is semantic and low layer location information enables the network to make full use of high low-level feature, improves model to the sensibility and sense of Small object
Degree of knowing, while improving model totality detection performance.Secondly by the more classification policies of multi-angle of view, mesh under high dynamic scene is realized
Mark the accurate detection of classification.The present invention considers from speed, practical, robustness etc., proposes a kind of based on Analysis On Multi-scale Features
The deep learning network improvement method of figure jump fusion, improves detection performance of the SSD algorithm under high dynamic scene.
Detailed description of the invention
Fig. 1 be system flow schematic diagram of the invention;
Fig. 2 a is that multi-scale prediction characteristic pattern layer jump connects graph of forming relations;
Fig. 2 b is the structural schematic diagram of multi-scale prediction characteristic pattern layer jump connection;
Fig. 3 Fusion Module a flow chart;
Fig. 4 Fusion Module b flow chart;
Fig. 5 multi-angle of view characteristic pattern layer jump connecting detection model framework.
Specific embodiment
In the following with reference to the drawings and specific embodiments, the present invention is furture elucidated, it should be understood that these embodiments are merely to illustrate
It the present invention rather than limits the scope of the invention, after the present invention has been read, those skilled in the art are to of the invention each
The modification of kind equivalent form falls within the application range as defined in the appended claims.
As shown in Figure 1, the present embodiment to the improved method of multiple dimensioned network the following steps are included:
1) construction feature converged network.
Feature pyramid is fused in SSD algorithm, forms multi-scale prediction feature pyramid network FPNSSD.First will
The Fusion Features of the eight or two convolutional layer Conv8_2 and the 9th 2 convolutional layer Conv9_2 in SSD algorithm obtain the eight or two fusion volume
Lamination Conv8_2_ff, the eight or two fusion convolutional layer Conv8_2_ff using up-sampling and with the seven or two convolutional layer Conv7_2
Get back the seven or two fusion convolutional layer Conv7_2_ff after progress Fusion Features.Seven or two fusion convolutional layer Conv7_2_ff passes through
The six or two is obtained with the six or two convolutional layer Conv6_2 Fusion Features after up-sampling merges convolutional layer Conv6_2_ff, and the six or two
Fusion convolutional layer Conv6_2_ff obtains the 7th and merges full connection after up-sampling with the 7th full articulamentum fc7 Fusion Features again
Layer fc7_ff.The 7th full articulamentum fc7_ff of fusion is obtained after up-sampling with the four or three convolutional layer Conv4_3 Fusion Features
Four or three fusion convolutional layer Conv4_3_ff.Finally using the four or three fusion convolutional layer Conv4_3_ff, the full connection of the 7th fusion
Layer fc7_ff, the six or two fusion convolutional layer Conv6_2_ff, the seven or two fusion convolutional layer Conv7_2_ff, the eight or two fusion convolution
Layer Conv8_2_ff and the 9th 2 convolutional layer Conv9_2 is as multi-scale prediction characteristic pattern.
Then the adjacent connection AdjacentSSD of design multi-scale prediction characteristic pattern.The 7th full articulamentum fc7 warp in SSD
It crosses after up-sampling and obtains the four or three with the four or three convolutional layer Conv4_3 Fusion Features and merge convolutional layer Conv4_3_ff, the six or two
Convolutional layer Conv6_2 obtains the 7th with the 7th full articulamentum fc7 Fusion Features after up-sampling and merges full articulamentum fc7_ff.
Seven or two convolutional layer Conv7_2 obtains the six or two with the six or two convolutional layer Conv6_2 Fusion Features after up-sampling and merges volume
Lamination Conv6_2_ff, the eight or two convolutional layer Conv8_2 are obtained by up-sampling with the seven or two convolutional layer Conv7_2 Fusion Features
Seven or two fusion convolutional layer Conv7_2_ff.9th 2 convolutional layer Conv9_2 by up-sampling after with the eight or two convolutional layer
Conv8_2 Fusion Features obtain the eight or two fusion convolutional layer Conv8_2_ff.Finally using the four or three fusion convolutional layer Conv4_
3_ff, the full articulamentum fc7_ff of the 7th fusion, the six or two fusion convolutional layer Conv6_2_ff, the seven or two fusion convolutional layer Conv7_
2_ff, the eight or two fusion convolutional layer Conv8_2_ff and the 9th 2 convolutional layer Conv9_2 are as multi-scale prediction characteristic pattern.
As shown in figures 2 a and 2b, the characteristic pattern layer jump for then designing multi-scale prediction connects SKIPSSD.In SSD
9th 2 convolutional layer Conv9_2 merges to obtain the seven or two fusion convolutional layer with the seven or two convolutional layer Conv7_2 after up-sampling
Conv7_2_ff.Eight or two convolutional layer Conv8_2 merges to obtain the six or two after up-sampling with the six or two convolutional layer Conv6_2
Merge convolutional layer Conv6_2_ff.Seven or two convolutional layer Conv7_2 merges to obtain after being up-sampled with the 7th full articulamentum fc7
The 7th full articulamentum fc7_ff of fusion.Six or two convolutional layer Conv6_2 carry out again up-sampling and with the four or three convolutional layer Conv4_3
Fusion obtains the four or three fusion convolutional layer Conv4_3_ff.The four or the three fusion convolutional layer Conv4_3_ formed after above-mentioned fusion
Ff, the full articulamentum fc7_ff of the 7th fusion, the six or two fusion convolutional layer Conv6_2_ff and the seven or two fusion convolutional layer Conv7_
2_ff is used as multi-scale prediction special together with the eight or two convolutional layer Conv8_2 and the 9th 2 convolutional layer Conv9_2 in former SSD
Sign figure.
The part connection Part-SKIPSSD for continuing to design multi-scale prediction is partially connected compared with aforementioned jump connection
The Fusion Features for reducing layer, only with the four or three fusion convolutional layer Conv4_3_ff, the 7th full articulamentum fc7_ff of fusion, the
Six or two fusion the seven or the two convolutional layer Conv7_2 of convolutional layer Conv6_2_ff and original SSD, the eight or two convolutional layer Conv8_2, the
92 convolutional layer Conv9_2.
Two-way jump connection Bi-SKIPSSD is redesigned, compared with aforementioned jump connection, two-way jump, which connects, increases the
The Fusion Features of eight or two convolutional layer Conv8_2 and the 9th 2 convolutional layer Conv9_2, the six or two convolutional layer Conv6_2 pass through convolution
After pondization operation with obtain the eight or two after the eight or two convolutional layer Conv8_2 Fusion Features and merge convolutional layer Conv8_2_ff, the 7th
Two convolutional layer Conv7_2 obtain the 9th 2 with the 9th 2 convolutional layer Conv9_2 Fusion Features and melt after also passing through the operation of convolution pondization
Close convolutional layer Conv9_2_ff.Two-way jump connection is using the four or three fusion convolutional layer Conv4_3_ff, the full connection of the 7th fusion
Layer fc7_ff, the six or two fusion convolutional layer Conv6_2_ff, the seven or two fusion convolutional layer Conv7_2_ff, the eight or two fusion convolution
Convolutional layer Conv9_2_ff is as multi-scale prediction characteristic pattern for the fusion of layer Conv8_2_ff and the 9th 2.
The jump of design fusion part basis network characterization figure connects Base-SKIPSSD, and jump connection is not using original
The strategy merged between six layers of predicted characteristics figure of SSD network, but jump connection is carried out on entire basic network.Such as
4th 1 convolutional layer Conv4_1 is merged to obtain after the operation of convolution pondization with the four or three convolutional layer Conv4_3 characteristic pattern
Four or three fusion convolutional layer Conv4_3_ff, subsequent 7th full articulamentum fc7, the six or two convolutional layer Conv6_2, volume seven or two
Lamination Conv7_2, the eight or two convolutional layer Conv8_2 and the 9th 2 convolutional layer Conv9_2 are all made of similar mode and they are right
The foundation characteristic layer answered carries out Fusion Features, and convolutional layer Conv4_3_ff, the 7th fusion are merged entirely in the four or three obtained after fusion
Articulamentum fc7_ff, the six or two fusion convolutional layer Conv6_2_ff, the seven or two fusion convolutional layer Conv7_2_ff, the eight or two fusion
Convolutional layer Conv9_2_ff is as multi-scale prediction characteristic pattern for the fusion of convolutional layer Conv8_2_ff and the 9th 2.
More than test six kinds of Fusion Features networks on VOC2007 data set.Wherein the jump of interlayer is connected accuracy
It is promoted to 79.0% from 77.2%, there is best detection performance, therefore the present embodiment selects layer jump connection shown in Fig. 2
As Fusion Features network structure.
2) design feature fusion connection module.
Design feature Fusion Module a first.As shown in figure 3, Fusion Module a (includes first advanced spy to high-level characteristic figure
The figure layer of sign) up-sampled after high-level characteristic figure, then through 3 × 3 convolution kernel convolution algorithms, then line
Property rectification function (relu) activation obtain high-level characteristic figure to be fused.Then by low-level feature figure through 3 × 3 convolution kernel convolution algorithms
It activates to obtain low-level feature figure to be fused with relu activation primitive.Fusion Features operation splicing or element summation are carried out again, are obtained
High low layer characteristic pattern after splicing/element summation obtains finally by the dimensionality reduction of 1 × 1 convolution kernel and the activation of relu activation primitive
To the high low layer characteristic pattern merged completely.
Then Fusion Module b is designed.As shown in figure 4, Fusion Module b first up-samples high-level characteristic figure, obtain
High-level characteristic figure after up-sampling.Using low-level feature figure after 1 × 1 convolution kernel dimensionality reduction and the activation of line rectification function
Low-level feature figure after to dimensionality reduction.Then Fusion Features operation is carried out, i.e. splicing or element summation obtains splicing/element summation
High low layer characteristic pattern afterwards, finally using 3 × 3 convolution kernel convolution algorithms to reduce aliasing effect, then line rectification function
Activation obtains complete fused high low layer characteristic pattern.
Fusion Module a and b are tested in VOC2007 test data set.Fusion Module b connects as the Fusion Features in network
Connection module, accuracy improve 0.3% than Fusion Module a, so that network has better performance, therefore the present embodiment selects
Fusion Module b is as Fusion Features link block.
3) convergence strategy is selected, is included the following steps:
Using the amalgamation mode of splicing or element summation when S3.1, Fusion Features, element summation in this patent can make
Network has better performance;
S3.2, using batch normalization to operate Fusion Features after splicing/element summation more abundant.
4) up-sampling mode is selected, is included the following steps:
S4.1, select deconvolution+cavity convolution or bilinear interpolation as up-sampling mode;
S4.2, test above two up-sampling mode in VOC2007 test data set, bilinear interpolation than deconvolution+
Empty convolution accuracy improves 0.6%, therefore bilinear interpolation is selected so that network is had better performance as up-sampling mode.
5) it is obtained by preceding 4 steps based on SSD Analysis On Multi-scale Features figure layer jump fusion structure.
6) model for incorporating multi-angle of view Strategies Training step 5) obtains as shown in Figure 5 based on the jump of Analysis On Multi-scale Features figure
The multi-angle of view SSD improved model structure of fusion, includes the following steps:
S6.1, using target all angles sample as different classifications, choose three Typical angles, respectively positive, side
Face, the back side, while background classification is added, achieving the effect that, which reduces false detection rate, increases model robustness;
S6.2, the training that the more classification based training samples of multi-angle of view are used for SKIPSSD obtain jumping based on Analysis On Multi-scale Features figure
The multi-angle of view SSD improved model of fusion.
To sum up, the present invention carries out Fusion Features by the jump connection between Analysis On Multi-scale Features figure layer, by merging high-rise language
Justice and low layer location information enable the network to make full use of high low-level feature, improve model to the sensibility and perception of Small object
Degree, while improving model totality detection performance.Secondly by the more classification policies of multi-angle of view, target under high dynamic scene is realized
The accurate detection of classification.The present invention considers from speed, practical, robustness etc., proposes a kind of based on Analysis On Multi-scale Features figure
The deep learning network improvement method of jump fusion, improves detection performance of the SSD algorithm under high dynamic scene, same this paper
Improved method be also applied for other deep learning network such as YOLO, practical value is high, has broad application prospects.
Claims (6)
1. a kind of deep learning network improvement method based on the jump fusion of Analysis On Multi-scale Features figure, which is characterized in that including following
Step:
Construct the Fusion Features network based on convolutional layer;
Design feature fusion connection module;
Convergence strategy and up-sampling mode are selected, is obtained based on SSD Analysis On Multi-scale Features figure layer jump fusion structure;
Incorporate the above-mentioned scale feature figure layer jump fusion structure of multi-angle of view Strategies Training.
2. the deep learning network improvement method according to claim 1 based on the jump fusion of Analysis On Multi-scale Features figure, special
Sign is that the Fusion Features network is characterized the connection of figure layer jump.
3. the deep learning network improvement method according to claim 2 based on the jump fusion of Analysis On Multi-scale Features figure, special
Sign is that the characteristic pattern layer jump connection is successively complete including the four or three fusion convolutional layer (Conv4_3_ff), the 7th fusion
Articulamentum (fc7_ff), the six or two fusion convolutional layer (Conv6_2_ff), the seven or two fusion convolutional layer (Conv7_2_ff), the 8th
Two convolutional layers (Conv8_2) and the 9th 2 convolutional layer (Conv9_2).
4. the deep learning network improvement method according to claim 1 based on the jump fusion of Analysis On Multi-scale Features figure, special
Sign is that the Fusion Features link block first up-samples high-level characteristic figure, the high-level characteristic after being up-sampled
Figure.Low-level feature after obtaining dimensionality reduction after 1 × 1 convolution kernel dimensionality reduction and the activation of line rectification function using low-level feature figure
Figure.Then carry out Fusion Features operation, i.e. splicing or element summation, the high low layer characteristic pattern after obtaining splicing/element summation, most
Afterwards using 3 × 3 convolution kernel convolution algorithms to reduce aliasing effect, then line rectification function is activated, after obtaining fusion completely
High low layer characteristic pattern.
5. the deep learning network improvement method according to claim 1 based on the jump fusion of Analysis On Multi-scale Features figure, special
Sign is that the convergence strategy is first element summation, then carries out batch normalization.
6. the deep learning network improvement method according to claim 1 based on the jump fusion of Analysis On Multi-scale Features figure, special
Sign is that the up-sampling mode is bilinear interpolation.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910566224.1A CN110378398B (en) | 2019-06-27 | 2019-06-27 | Deep learning network improvement method based on multi-scale feature map jump fusion |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910566224.1A CN110378398B (en) | 2019-06-27 | 2019-06-27 | Deep learning network improvement method based on multi-scale feature map jump fusion |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110378398A true CN110378398A (en) | 2019-10-25 |
CN110378398B CN110378398B (en) | 2023-08-25 |
Family
ID=68250974
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910566224.1A Active CN110378398B (en) | 2019-06-27 | 2019-06-27 | Deep learning network improvement method based on multi-scale feature map jump fusion |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110378398B (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111222534A (en) * | 2019-11-15 | 2020-06-02 | 重庆邮电大学 | Single-shot multi-frame detector optimization method based on bidirectional feature fusion and more balanced L1 loss |
CN111476249A (en) * | 2020-03-20 | 2020-07-31 | 华东师范大学 | Construction method of multi-scale large-receptive-field convolutional neural network |
CN112070070A (en) * | 2020-11-10 | 2020-12-11 | 南京信息工程大学 | LW-CNN method and system for urban remote sensing scene recognition |
WO2021129691A1 (en) * | 2019-12-23 | 2021-07-01 | 长沙智能驾驶研究院有限公司 | Target detection method and corresponding device |
CN113496158A (en) * | 2020-03-20 | 2021-10-12 | 中移(上海)信息通信科技有限公司 | Object detection model optimization method, device, equipment and storage medium |
CN117828407A (en) * | 2024-03-04 | 2024-04-05 | 江西师范大学 | Double-stage gating attention time sequence classification method and system for bidirectional jump storage |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109118539A (en) * | 2018-07-16 | 2019-01-01 | 深圳辰视智能科技有限公司 | Point cloud and picture fusion method, device and its equipment based on Analysis On Multi-scale Features |
US20190156144A1 (en) * | 2017-02-23 | 2019-05-23 | Beijing Sensetime Technology Development Co., Ltd | Method and apparatus for detecting object, method and apparatus for training neural network, and electronic device |
CN109800628A (en) * | 2018-12-04 | 2019-05-24 | 华南理工大学 | A kind of network structure and detection method for reinforcing SSD Small object pedestrian detection performance |
-
2019
- 2019-06-27 CN CN201910566224.1A patent/CN110378398B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190156144A1 (en) * | 2017-02-23 | 2019-05-23 | Beijing Sensetime Technology Development Co., Ltd | Method and apparatus for detecting object, method and apparatus for training neural network, and electronic device |
CN109118539A (en) * | 2018-07-16 | 2019-01-01 | 深圳辰视智能科技有限公司 | Point cloud and picture fusion method, device and its equipment based on Analysis On Multi-scale Features |
CN109800628A (en) * | 2018-12-04 | 2019-05-24 | 华南理工大学 | A kind of network structure and detection method for reinforcing SSD Small object pedestrian detection performance |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111222534A (en) * | 2019-11-15 | 2020-06-02 | 重庆邮电大学 | Single-shot multi-frame detector optimization method based on bidirectional feature fusion and more balanced L1 loss |
WO2021129691A1 (en) * | 2019-12-23 | 2021-07-01 | 长沙智能驾驶研究院有限公司 | Target detection method and corresponding device |
CN111476249A (en) * | 2020-03-20 | 2020-07-31 | 华东师范大学 | Construction method of multi-scale large-receptive-field convolutional neural network |
CN113496158A (en) * | 2020-03-20 | 2021-10-12 | 中移(上海)信息通信科技有限公司 | Object detection model optimization method, device, equipment and storage medium |
CN112070070A (en) * | 2020-11-10 | 2020-12-11 | 南京信息工程大学 | LW-CNN method and system for urban remote sensing scene recognition |
CN117828407A (en) * | 2024-03-04 | 2024-04-05 | 江西师范大学 | Double-stage gating attention time sequence classification method and system for bidirectional jump storage |
CN117828407B (en) * | 2024-03-04 | 2024-05-14 | 江西师范大学 | Double-stage gating attention time sequence classification method and system for bidirectional jump storage |
Also Published As
Publication number | Publication date |
---|---|
CN110378398B (en) | 2023-08-25 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110378398A (en) | A kind of deep learning network improvement method based on the jump fusion of Analysis On Multi-scale Features figure | |
CN111126202A (en) | Optical remote sensing image target detection method based on void feature pyramid network | |
CN110163246A (en) | The unsupervised depth estimation method of monocular light field image based on convolutional neural networks | |
CN113392960B (en) | Target detection network and method based on mixed hole convolution pyramid | |
CN112052763B (en) | Video abnormal event detection method based on two-way review generation countermeasure network | |
CN112150450B (en) | Image tampering detection method and device based on dual-channel U-Net model | |
CN111291779A (en) | Vehicle information identification method and system, memory and processor | |
CN112699889A (en) | Unmanned real-time road scene semantic segmentation method based on multi-task supervision | |
CN116597326A (en) | Unmanned aerial vehicle aerial photography small target detection method based on improved YOLOv7 algorithm | |
CN114596316A (en) | Road image detail capturing method based on semantic segmentation | |
CN112819063A (en) | Image identification method based on improved Focal loss function | |
CN116309485A (en) | Pavement crack detection method for improving UNet network structure | |
CN112037225A (en) | Marine ship image segmentation method based on convolutional nerves | |
CN117422696A (en) | Belt wear state detection method based on improved YOLOv8-Efficient Net | |
CN116452950A (en) | Multi-target garbage detection method based on improved YOLOv5 model | |
CN114821316A (en) | Three-dimensional ground penetrating radar crack disease identification method and system | |
CN117350635A (en) | Building material warehouse-in and warehouse-out management system and method | |
CN117173526A (en) | Insulator defect detection method adopting self-adaptive feature fusion and light-weight YOLOv5s | |
CN116912130A (en) | Image defogging method based on multi-receptive field feature fusion and mixed attention | |
CN115393733B (en) | Automatic water body identification method and system based on deep learning | |
CN111178275A (en) | Fire detection method based on convolutional neural network | |
CN111047569A (en) | Image processing method and device | |
CN116229104A (en) | Saliency target detection method based on edge feature guidance | |
CN116416534A (en) | Unmanned aerial vehicle spare area identification method facing protection target | |
CN112131925B (en) | Construction method of multichannel feature space pyramid |
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 |