CN108764063A - A kind of pyramidal remote sensing image time critical target identifying system of feature based and method - Google Patents
A kind of pyramidal remote sensing image time critical target identifying system of feature based and method Download PDFInfo
- Publication number
- CN108764063A CN108764063A CN201810427107.2A CN201810427107A CN108764063A CN 108764063 A CN108764063 A CN 108764063A CN 201810427107 A CN201810427107 A CN 201810427107A CN 108764063 A CN108764063 A CN 108764063A
- Authority
- CN
- China
- Prior art keywords
- layer
- characteristic layer
- characteristic
- network
- feature
- 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
- 238000000034 method Methods 0.000 title claims abstract description 50
- 230000004927 fusion Effects 0.000 claims abstract description 105
- 238000013507 mapping Methods 0.000 claims abstract description 34
- 238000012545 processing Methods 0.000 claims abstract description 30
- 230000008569 process Effects 0.000 claims abstract description 18
- 238000000605 extraction Methods 0.000 claims abstract description 13
- 238000011282 treatment Methods 0.000 claims description 13
- 238000005070 sampling Methods 0.000 claims description 12
- 239000000284 extract Substances 0.000 claims description 9
- 230000000694 effects Effects 0.000 claims description 8
- 238000002203 pretreatment Methods 0.000 claims 1
- 238000001514 detection method Methods 0.000 description 26
- 108091006146 Channels Proteins 0.000 description 16
- 238000012360 testing method Methods 0.000 description 16
- 238000012549 training Methods 0.000 description 11
- 238000005516 engineering process Methods 0.000 description 6
- 238000013135 deep learning Methods 0.000 description 4
- 238000013461 design Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 3
- 230000005284 excitation Effects 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 238000013480 data collection Methods 0.000 description 2
- PCHJSUWPFVWCPO-UHFFFAOYSA-N gold Chemical compound [Au] PCHJSUWPFVWCPO-UHFFFAOYSA-N 0.000 description 2
- 239000010931 gold Substances 0.000 description 2
- 229910052737 gold Inorganic materials 0.000 description 2
- 238000010191 image analysis Methods 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000002776 aggregation Effects 0.000 description 1
- 238000004220 aggregation Methods 0.000 description 1
- 238000005267 amalgamation Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000001125 extrusion Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000007500 overflow downdraw method Methods 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 230000009897 systematic effect Effects 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of pyramidal remote sensing image time critical target identifying system of feature based and method, which includes target's feature-extraction sub-network, characteristic layer sub-network, candidate region generates sub-network and classification returns sub-network.Target's feature-extraction sub-network is used to carry out multi-layer process of convolution to pending image, and is exported each level convolution processing result as a characteristic layer.A upper characteristic layer and current signature layer overlap-add procedure are obtained present fusion characteristic layer by characteristic layer sub-network, and top fusion feature layer is top characteristic layer.Candidate region generates sub-network and is used to extract candidate region from different levels fusion feature layer, classification, which returns sub-network candidate region is mapped to different levels fusion feature layer, obtains mapping treated multiple level fusion feature layers, and treated that multiple level fusion feature layers carry out target discriminations exports results to mapping.Utilize feature pyramid hierarchical structure so that the feature of all scales all has abundant semantic information.
Description
Technical field
The invention belongs to field of remote sensing image processing, more particularly, to a kind of pyramidal remote sensing image of feature based
Time critical target identifying system and method.
Background technology
Large format remote sensing images Moving target detection be an important composition in remote Sensing Image Analysis field and because its tool
Have a characteristics such as multiple dimensioned, breadth is wide, visual angle is special and extremely challenging, basic task be exactly for a given width aviation or
The essence comprising one or more classification target object and each class target is determined whether in person's satellite image
True position.The detection of large format remote sensing images time critical target is as the important of computer vision field and remote Sensing Image Analysis field
Component part has very high research significance.With the rapid hair of contemporary remote sensing technology, sensor technology and Internet technology
Exhibition, people can utilize remote sensing technology to realize that the ground region of large format or sea area detect, so that remote sensing obtains
Aggregation of data it is more and more, wherein also including all various information such as geography, humanity.Remote sensing technology is in ground plane mesh
It is marked with and ShipTargets is found and relief, illegal immigrant, guard territory, environmental monitoring, military information collection, ground resource
Numerous aspects such as investigation all have a wide range of applications.
For target detection there are two basic task, one is to determine in image that two are to determine the position of target with the presence or absence of target
It sets.Core research field of the target detection as vision and Digital Image Processing in natural scene was always to learn in recent years
Art circle is studied and the hot spot direction of industrial quarters application.Since 2012, convolutional neural networks obtain in image classification field
Important breakthrough, the powerful ability in feature extraction of deep learning is equally also that high-precision target detection brings possibility, target detection
Problem is converted into picture classification problem, and deep learning is applied to target detection, has pushed the at full speed of target detection technique field
Development, the object detection method based on deep learning also make rapid progress.
Existing deep learning target detection model usually only utilizes last layer single scale feature, although last layer Feature Semantics
Information is very strong, but location information is very weak, poor to multiscale target detection result, meanwhile, the big parameter of existing network scale is more, right
Computational resource requirements are high.
Invention content
The present invention detects problem for large format remote sensing images multiscale target, proposes that a kind of feature based is pyramidal distant
Feel image time critical target identifying system and method.Feature pyramid network SEM-FPN based on channel weighting is applied to remote sensing
Image time critical target detects, and is improved according to the problem of network model practical application, and algorithm realizes multiple dimensioned mesh
Target detects, while having the characteristics that lightweight.
As an aspect of of the present present invention, the present invention provides a kind of pyramidal remote sensing image time critical target identification of feature based
System, including：
Target's feature-extraction sub-network is equipped with many levels output end, for carrying out multi-layer convolution to pending image
Processing, and exported each level convolution processing result as a characteristic layer；
Characteristic layer sub-network is equipped with many levels input terminal and many levels output end, a same target of level input terminal
One level output end of feature extraction sub-network connects, for being worked as a upper characteristic layer and current signature layer overlap-add procedure
Preceding fusion feature layer, top fusion feature layer are top characteristic layer；
Candidate region generates sub-network, is equipped with many levels input terminal and many levels output end, a level input terminal
A level output end with characteristic layer sub-network connects, for extracting candidate region from different levels fusion feature layer；
Classification returns sub-network, is equipped with many levels input terminal and RPN input terminals, a level input terminal is the same as feature straton
One level output end connection of network, RPN input terminals generate the output end connection of sub-network with candidate region, are used for candidate regions
Domain mapping to different levels fusion feature layer obtains mapping treated multiple level fusion feature layers, and treated to mapping
Multiple level fusion feature layers carry out target discrimination and export result.
Preferably, characteristic layer sub-network includes multiple characteristic layer submodules, be denoted as first characteristic layer submodule, second
Characteristic layer submodule ..., ith feature straton module ..., n-th characteristic layer submodule；1≤i≤N-1；
A level input terminal of one input terminal of ith feature straton module as this feature straton network, i-th
Another input terminal of characteristic layer submodule is connected with the output end of i+1 characteristic layer submodule, n-th characteristic layer submodule
A level input terminal of the input terminal as this feature straton network；
Preceding N-1 characteristic layer submodule to last layer characteristic layer and current signature layer overlap-add procedure for obtaining present fusion
Characteristic layer；N-th characteristic layer submodule is used to export current signature layer as present fusion characteristic layer, wherein last layer is special
Sign layer is by being obtained after carrying out convolution to current signature layer.
Preferably, any one characteristic layer submodule includes in preceding N-1 characteristic layer submodule：
Last layer handles subelement, currently processed subelement and superpositing unit, and the output end that last layer handles subelement is same
The first input end of superpositing unit connects, and the output end of currently processed subelement is connected with the second input terminal of superpositing unit；
Last layer handles subelement and is used for last layer feature after the progress up-sampling treatment output processing of last layer characteristic layer
Layer, currently processed subelement are used to use current signature layer, superpositing unit after the progress process of convolution output processing of current signature layer
Last layer characteristic layer and current signature layer after processing are overlapped processing after to processing, export present fusion characteristic layer.
Preferably, characteristic layer submodule further includes aliasing effect processing unit, output end of the input terminal with superpositing unit
Connection, aliasing effect processing unit are used to carry out process of convolution to present fusion characteristic layer, export final present fusion characteristic layer.
Preferably, characteristic layer sub-network further includes the N+1 characteristic layer submodule, and candidate region generates sub-network equipped with attached
Add level input terminal；The N+1 characteristic layer submodule input terminal is connected with the output end of n-th characteristic layer submodule, N+1
Characteristic layer submodule output end generates the extra play secondary input end connection of sub-network with candidate region；
The fusion feature layer that the N+1 characteristic layer submodule is used to export n-th characteristic layer submodule up-samples
The N+1 characteristic layer is exported, candidate region generates sub-network and extracts candidate region from the N+1 characteristic layer simultaneously.
Preferably, it includes sequentially connected mapping submodule, fusion submodule and target discrimination that classification, which returns sub-network,
Module, mapping submodule, which is used to candidate region mapping to different levels fusion feature layer, obtains mapping treated multiple levels
Fusion feature layer, fusion submodule are used for that treated that multiple level fusion feature layers carry out fusion treatments obtains targets to mapping
Judge that characteristic layer, target discrimination submodule are used to carry out target discrimination to target discrimination characteristic layer to export result.
As another aspect of the present invention, the present invention provides a kind of identification based on remote sensing image time critical target identifying system
Method includes the following steps：
S110 carries out multi-layer process of convolution to pending image, and using each level convolution processing result as pyramid
The hierarchy characteristic layer in characteristic layer；
I-th of hierarchy characteristic layer in pyramid characteristic layer and i+1 hierarchy characteristic layer overlap-add procedure are obtained i-th by S120
A level fusion feature layer；And i is traversed into preceding N-1 level in pyramid characteristic layer and obtains N-1 level fusion feature layer；It will
N-th hierarchy characteristic layer is as n-th level fusion feature layer in pyramid characteristic layer；
S130 extracts candidate region from N number of level fusion feature layer；
Candidate region is mapped to N number of level fusion feature layer and obtains mapping treated multiple level fusion features by S140
Layer, and treated that multiple level fusion feature layers carry out target discriminations exports results to mapping.
Preferably, further include following steps between step S110 and step S120：
Carrying out process of convolution to the i-th hierarchy characteristic layer makes the i-th hierarchy characteristic layer channel with logical to i+1 hierarchy characteristic layer
Road is identical；Carrying out up-sampling treatment to i+1 hierarchy characteristic layer makes the i-th hierarchy characteristic layer size with to i+1 hierarchy characteristic
Layer size is identical；1≤i≤N-1.
Preferably, step S140 includes following sub-step：
Step S141：Candidate region is mapped into N number of level fusion feature layer and obtains mapping treated N number of level fusion
Characteristic layer；
Step S142：And treated that N number of level fusion feature layer carries out that fusion treatment obtains target discrimination is special to mapping
Levy layer；
Step S143：Target discrimination characteristic layer carries out target discrimination and exports result.
On the whole, the present invention has the advantage that compared with existing remote sensing images multiscale target detection technique：
1, the present invention provides a kind of remote sensing image time critical target recognition methods of feature based pyramid network, by feature
Pyramid structure thought is introduced into Remote Sensing Target detection, using the feature pyramid hierarchical structure of convolutional neural networks, and
Information between each level is subjected to fusion utilization, is connected using top-down side, high-rise semantic information is passed downwards
It passs so that the feature of all scales all has abundant semantic information, and this method obtains excellent detection result, and network has
Good adaptability；
2, the present invention provides a kind of remote sensing image time critical target recognition methods of feature based pyramid network, propose more
The pond level candidate region (RoI) characteristic pattern fusion method, by the multi-level ponds RoI Fusion Features, this method is by network parameter
Amount is greatly reduced, and realizes network lightweight, improves training and test speed.
3, the present invention provides a kind of remote sensing image time critical target recognition methods of feature based pyramid network, SE is tied
Structure incorporates among target detection network structure elements, explicitly models the relation of interdependence between channel, adaptively again
The characteristic response of calibrated channel, lifting feature extractability.
Description of the drawings
Fig. 1 is the structural representation of the pyramidal remote sensing image multiscale target identifying system of feature based provided by the invention
Figure；
Fig. 2 is the flow chart of the pyramidal remote sensing image multiscale target recognition methods of feature based provided by the invention；
Fig. 3 is SE structural schematic diagrams in remote sensing image multiscale target identifying system provided by the invention；
Fig. 4 is Analysis On Multi-scale Features information fusion signal in remote sensing image multiscale target identifying system provided by the invention
Figure；
Fig. 5 is that classification returns sub-network construction feature gold in remote sensing image multiscale target identifying system provided by the invention
Word tower structure schematic diagram；
Fig. 6 is the middle-level pool area Fusion Features signal of remote sensing image multiscale target identifying system provided by the invention
Figure；
Fig. 7 is data set analysis target scale distribution schematic diagram of the present invention；
Fig. 8 is remote sensing images time critical target testing result in case study on implementation of the present invention；
It is to identify scene once that Fig. 9 (a1), which is identifying that scene once uses Fast RCNN testing results, Fig. 9 (a2),
Using R-FCN testing results, Fig. 9 (a3) is once to use FPN testing results in identification scene, and Fig. 9 (b1) is in identification scene
It is to identify scene two times using R-FCN testing results that two times, which use Fast RCNN testing results, Fig. 9 (b2), and Fig. 9 (b3) is
Scene is being identified two times using FPN testing results, and Fig. 9 (c1) is to identify scene three times using Fast RCNN testing results, scheming
9 (c2) use R-FCN testing results, Fig. 9 (c3) to use FPN detection knots for three times in identification scene for three times in identification scene
Fruit.
Specific implementation mode
In order to make the purpose , technical scheme and advantage of the present invention be clearer, case with reference to the accompanying drawings and embodiments,
The present invention will be described in further detail.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention,
It is not intended to limit the present invention.In addition, technical characteristic involved in the various embodiments of the present invention described below is only
It does not constitute a conflict with each other and can be combined with each other.
Embodiment one
The pyramidal remote sensing image time critical target identifying system of a kind of feature based provided by the invention, the image recognition system
System includes target's feature-extraction sub-network, characteristic layer sub-network, candidate region generates sub-network and classification returns sub-network.Its
In, target's feature-extraction sub-network is equipped with many levels output end, and classification returns sub-network and is equipped with many levels input terminal and RPN
Input terminal, characteristic layer sub-network and candidate region generate sub-network and are equipped with many levels input terminal and many levels output end,
One level output end of target's feature-extraction sub-network is connected with a level input terminal of characteristic layer sub-network, feature straton
One level output end of network generates the level input terminal connection of sub-network, characteristic layer sub-network one with candidate region
Level output end returns one level input terminal connection of sub-network with classifying, and candidate region generates the output end of sub-network with classification
Return the RPN input terminals connection of sub-network.
Target's feature-extraction sub-network is used to carry out multi-layer process of convolution to pending image, and by each level convolution
Handling result is exported as a characteristic layer.Characteristic layer sub-network is worked as a upper characteristic layer and current signature layer overlap-add procedure
Preceding fusion feature layer, top fusion feature layer are top characteristic layer.Candidate region generates sub-network and is used for from different levels
Fusion feature layer extracts candidate region, and classification returns sub-network and candidate region is mapped to the acquisition of different levels fusion feature layer
Mapping treated multiple level fusion feature layers, and treated that multiple level fusion feature layers carry out target discriminations to mapping
Export result.
Embodiment two
On the basis of embodiment one, characteristic layer sub-network includes multiple characteristic layer submodules, is denoted as first feature
Straton module, second characteristic layer submodule ..., ith feature straton module ..., n-th characteristic layer submodule.Its
In, the level input terminal of an input terminal of ith feature straton module as this feature straton network, ith feature
Another input terminal of straton module with i+1 characteristic layer submodule output end connect, n-th characteristic layer submodule it is defeated
Enter a level input terminal of the end as this feature straton network, 1≤i≤N-1, preceding N-1 characteristic layer submodule for
One layer of characteristic layer and current signature layer overlap-add procedure obtain present fusion characteristic layer；N-th characteristic layer submodule is used for will be current
Characteristic layer is exported as present fusion characteristic layer, and last layer characteristic layer is by being obtained after carrying out convolution to current signature layer.
Embodiment three
As shown in Fig. 2, on the basis of embodiment two, any one feature straton in preceding N-1 characteristic layer submodule
Module includes that last layer processing subelement, currently processed subelement and superpositing unit, the output end that last layer handles subelement are same
The first input end of superpositing unit connects, and the output end of currently processed subelement is connected with the second input terminal of superpositing unit, on
One layer of processing subelement is used for last layer characteristic layer T2 after the progress up-sampling treatment output processing of last layer characteristic layer, current to locate
Manage subelement be used for current signature layer carry out 1 × 1 process of convolution output processing after current signature layer S2, superpositing unit for pair
Last layer characteristic layer and current signature layer are overlapped processing after processing, export present fusion characteristic layer.
Characteristic layer submodule provided in an embodiment of the present invention is made by the way that last layer characteristic layer is carried out up-sampling treatment
One layer of characteristic layer size increasing is twice, and carries out process of convolution to current signature layer so that the number of channels of current signature layer is same as above
The number of channels of one layer of characteristic layer is identical, and by overlap-add procedure, the present fusion characteristic layer of acquisition, present fusion characteristic layer is simultaneously
Possess the strong feature of the characteristic information of last layer characteristic layer, while having the characteristics that the location information of current signature layer is strong.
Example IV
On three basis of embodiment, characteristic layer submodule further includes aliasing effect processing unit, and input terminal is the same as superposition
The output end of unit connects, and aliasing effect processing unit is used to carry out process of convolution to present fusion characteristic layer, and output is finally worked as
Preceding fusion feature layer realizes the aliasing effect of removal up-sampling.
Embodiment five
On embodiment two to the basis of five any one of embodiment, characteristic layer sub-network further includes the N+1 feature
Straton module, candidate region generate sub-network and are equipped with extra play secondary input end, and input terminal is defeated with n-th characteristic layer submodule
Outlet connects, and output end is used to generate the extra play secondary input end connection of sub-network with candidate region, for n-th feature
Straton module output fusion feature layer carry out up-sampling output the N+1 characteristic layer, candidate region generate sub-network and meanwhile from
The N+1 characteristic layer extracts candidate region.
Embodiment six
On embodiment one to the basis of five any one of embodiment, it includes sequentially connected reflect that classification, which returns sub-network,
Submodule, fusion submodule and target discrimination submodule are penetrated, mapping submodule is melted for candidate region to be mapped to different levels
It closes characteristic layer and obtains mapping treated multiple level fusion feature layers, fusion submodule is used for mapping treated multiple layers
Grade fusion feature layer carries out fusion treatment and obtains target discrimination characteristic layer, and target discrimination submodule is used for target discrimination characteristic layer
It carries out target discrimination and exports result.
Embodiment seven
As shown in figure 3, a kind of pyramidal remote sensing image time critical target recognition methods of feature based, includes the following steps：
S110 carries out multi-layer process of convolution to pending image, and using each level convolution processing result as pyramid
The hierarchy characteristic layer in characteristic layer；
I-th of hierarchy characteristic layer in pyramid characteristic layer and i+1 hierarchy characteristic layer overlap-add procedure are obtained i-th by S120
A level fusion feature layer；And i is traversed into preceding N-1 level in pyramid characteristic layer and obtains N-1 level fusion feature layer；It will
N-th hierarchy characteristic layer is as n-th level fusion feature layer in pyramid characteristic layer；
S130 extracts candidate region from N number of level fusion feature layer；
Candidate region is mapped to N number of level fusion feature layer and obtains mapping treated multiple level fusion features by S140
Layer, and treated that multiple level fusion feature layers carry out target discriminations exports results to mapping.
Embodiment eight
Further include following steps between step S110 and step S120 on six basis of embodiment：
Carrying out process of convolution to the i-th hierarchy characteristic layer makes the i-th hierarchy characteristic layer channel with logical to i+1 hierarchy characteristic layer
Road is identical；Carrying out up-sampling treatment to i+1 hierarchy characteristic layer makes the i-th hierarchy characteristic layer size with to i+1 hierarchy characteristic
Layer size is identical；1≤i≤N-1.
Embodiment nine
On seven basis of embodiment six or embodiment, step S140 includes following sub-step：
Step S141：Candidate region is mapped into N number of level fusion feature layer and obtains mapping treated N number of level fusion
Characteristic layer；
Step S142：And treated that N number of level fusion feature layer carries out that fusion treatment obtains target discrimination is special to mapping
Levy layer；
Step S143：Target discrimination characteristic layer carries out target discrimination and exports result.
Present invention firstly provides a kind of feature based pyramid network remote sensing image time critical target recognition methods, design
A kind of feature pyramid target detection network model based on channel weighting promotes target signature from space and two, channel dimension
Three-dimensional ability to express and scale adaptability, by classify Recurrent networks in multi-level pool area Fusion Features so that net
Network scale of model substantially reduces, and realizes network lightweight.
The remote sensing image time critical target recognition methods embodiment of feature based pyramid network provided by the invention, this method
Detailed process is：
Step S110：Extract target signature in image
As shown in figure 4, the extraction network characterized by SE-MobileNet, by MobileNet network structure elements
SE structures are added to realize to target's feature-extraction in image.
Step S111：Squeeze extrusion operations
U primitive character figure of input is shortened to the real number ordered series of numbers of 1*1*C by global average pond into, U indicates channel
Global space character representation, real number ordered series of numbers indicate channel descriptor, i.e., for feature in the expression of channel dimension, specific formula is as follows：
Wherein, W, H are respectively the length and width of characteristic pattern, ucIndicate the characteristic pattern of input.
Step S112：Excitation excitation operations
Formula of the real number ordered series of numbers by Excitation operations after extruding indicates as follows：
S=Fex(z, W)=σ (g (z, W))=δ (W2σ(W1z))
Wherein, δ refers to ReLU activation primitives, and g refers to sigmoid activation primitives,
Two full connection (FC) layers are introduced, i.e. dimensionality reduction layer parameter is W1, and dimensionality reduction ratio is r, passes through a ReLU later, so
It is that the liter that a parameter is W2 ties up layer afterwards.Finally, the real number ordered series of numbers binding characteristic figure U of 1*1*C is carried out by following equation
Scale operates to obtain final output.
Wherein, X=[x1,x2,...,xc] and Fscale(uc,sc) refer to Feature Mapping uc∈RW×HWith scalar scBetween it is right
Answer channel product.
Step S120：Obtain different levels fusion feature layer
Step S121：Path from bottom to top
The pyramidal structure of feature is divided according to level, and it is a pyramid to define the identical characteristic pattern of scale
Rank, and using last characteristic pattern of each level as this layer of pyramidal final output.Use the convolution of each level
Layer the last one characteristic pattern as output, defined on the basis of characteristic pattern scale pyramid structure output be C2, C3, C4,
C5 }, it is respectively from convolutional layer conv2, conv3, conv4 and conv5, and these characteristic layers have relative to input picture
{ 4,8,16,32 } step-length of pixel.
Step S122：Top-down pyramid network path and lateral connection
Top-down pyramid network path builds information flow channel by connection, by high-level characteristic and current spy
Sign fusion.Concrete operations are to up-sample the last layer feature of convolutional network, then pass through lateral connection and preceding layer
Feature is merged, and the semantic information of low level is also strengthened, and the size between two characteristic patterns of lateral connection needs
If identical, the positioning detail information of current signature figure could be preferably utilized.
T1 is carried out 2 times of up-samplings first, obtains characteristic pattern T2, then passes through S1 layers of feature by the feature T1 of last layer
One 1 × 1 convolution operation adjusts number of active lanes, keeps consistent with the number of active lanes of T2 features, finally leads to T2 and adjustment
S2 layer features after road number are merged, and amalgamation mode is direct progress pixel addition.
Step S130：RPN network struction feature pyramid structures extract candidate region
Inputted in Faster RCNN into RPN networks characteristic pattern be fixed 16 × 16 this size, in order to realize
Multiple dimensioned effect needs the characteristic pattern of multiple scales as input.In order to by feature pyramid structure and RPN network integrations,
Fusion feature layer { P2, P3, P4, P5 } all as the input of RPN, is realized multiple dimensioned input by us, and in order to allow ruler
It is more abundant to spend information, maximum pond down-sampling is carried out to P5, generates P6, thus the input of RPN networks for P2, P3, P4, P5,
P6}.RPN networks have the candidate window (Anchors) of 9 kinds of different scales in Faster RCNN, are divided into three kinds of scales and three kinds
Length-width ratio, however for the multiple dimensioned input of RPN networks, multiple dimensioned has little significance, so just remaining three kinds long
Wide ratio.{ P2, P3, P4, P5, P6 } corresponding scale be 32 × 32,64 × 64,128 × 128,256 × 256,512 ×
512 }, three kinds of Aspect Ratios are added, so Anchors different in the RPN networks a total of 15 based on FPN.
Step S140：Target discrimination is carried out to multiple level fusion feature layers and exports result.
S141：According to the candidate frame size that RPN is generated, they are respectively mapped on the characteristic pattern of each scale.Assuming that
The size of candidate frame is w × h, it can should be mapped to P by following calculationkOn characteristic pattern, wherein k gets 5 from 2, specifically
Flow is as shown in Figure 5.
When network model pre-training 224 × 224 be pre-training image size, so benchmark candidate frame is dimensioned to
224, it is denoted as k0, due in Faster RCNN, using C4 as the input in the ponds RoI, so k0=4.Assuming that the w of candidate frame ×
H is respectively 112 × 112, then k=k0- 1=3, candidate frame should be mapped on P3 this characteristic pattern, then do the ponds RoI again
Change.P2-P5 characteristic layers are only used when specific implementation, do not include P6 characteristic layers.
S142：In classification returns sub-network, the result after the ponds multi-level features RoI is merged, then inputs
Classification regression forecasting network, as shown in fig. 6, network parameter will be reduced, promotion network training and test speed merge concrete structure
Shown in the following formula of calculating process.
Op=K × N × C × H × W
Wherein, OpFor the output after fusion, input feature vector figure is X={ x1,x2,x3…,xK, xk=N × C × H × W, K are
The quantity of fusion feature layer, value is 4 in the present embodiment, and H, W are respectively the length and width of characteristic pattern, and C is characterized figure port number, and N is
Characteristic pattern number.
Step S143：Target discrimination characteristic layer carries out target discrimination and exports result.
In embodiment provided by the invention, remote sensing image time critical target identifying system is trained using following steps：
Step S210：Data set makes and analysis
Pass through remote sensing figure under Google Earth satellite remote sensing images software download all over the world airport 0.6m resolution ratio
1000, picture, 2000 × 2000 or more image size.And sample is concentrated to carry out data by LabelImg image labelings tool
Mark, copies PASCAL VOC data set formats.
Data set includes the remote sensing image data collection of 1000 2000 × 2000 or more sizes, wherein training sample in total
340, test sample 330 is opened, and verification sample 330 is opened, and Aircraft Targets are more than 14000 framves.
It is analyzed by the size to all label targets, as shown in Figure 7.Abscissa is K, and ordinate is target
Number.The calculation formula of statistics is：
K in Fig. 7rThe scale size indicated when being 1,2,3 is respectively 16 × 16,32 × 32,64 × 64.
Step S220：According to the scale hyper parameter of data set target scale distribution statistics design system in step S210, and
Design other rational hyper parameters for systematic training, by training dataset input remote sensing image time critical target identifying system in into
Row training.
Step S230：Test data set is inputted in remote sensing image time critical target identifying system and carries out forward calculation, is tested
Remote sensing image time critical target identifying system detection performance and generalization ability.
The validity that feature pyramid network in order to verify proposition detects remote sensing images Aircraft Targets, with existing master
Stream target detection frame Faster RCNN and R-FCN are compared and analyzed, and detect three kinds of scenes, and detection image size is more than
1000 × 1000, Aircraft Targets testing result of the invention and detection result comparison diagram such as Fig. 8, Fig. 9 with other methods
Shown in (a1) to Fig. 9 (a3), Fig. 9 (b1) to Fig. 9 (b3) and Fig. 9 (c1) to Fig. 9 (c3), data set and this that the above method uses
Invention is consistent, and the results are shown in Table 1.
Using Average Accuracy as model-evaluation index, value is bigger, and expression detection performance is better, and detection time is big
Breadth remote sensing images test statistics are got.
Table 1
The present invention designs a kind of feature pyramid target detection network model SEM-FPN based on channel weighting first, packet
Include target's feature-extraction sub-network SE-MobileNet, characteristic layer sub-network, candidate region generation subnet based on channel weighting
Network RPN and classification return four sub-networks of sub-network, return construction feature gold word in sub-network in sub-network RPN and classification respectively
Tower structure, classification return the multi-level pool area Fusion Features of use in sub-network, reduce network parameter；Then remote sensing is made
Image data collection, and concentrate target scale to analyze data；In the training level of network, it is super that suitable training is set
Training sample is inputted in SEM-FPN and is trained end to end, obtains time critical target detection model by parameter；In image measurement
Test sample in data set is inputted time critical target detection model, carries out forward prediction calculating by level.
As it will be easily appreciated by one skilled in the art that the foregoing is merely illustrative of the preferred embodiments of the present invention, not to
The limitation present invention, all within the spirits and principles of the present invention made by all any modification, equivalent and improvement etc., should all include
Within protection scope of the present invention.
Claims (9)
1. a kind of pyramidal remote sensing image time critical target identifying system of feature based, which is characterized in that including：
Target's feature-extraction sub-network is equipped with many levels output end, for carrying out multi-layer process of convolution to pending image,
And it is exported each level convolution processing result as a characteristic layer；
Characteristic layer sub-network is equipped with many levels input terminal and many levels output end, a same target signature of level input terminal
The level output end connection for extracting sub-network, for currently being melted a upper characteristic layer and current signature layer overlap-add procedure
Characteristic layer is closed, top fusion feature layer is top characteristic layer；
Candidate region generates sub-network, is equipped with many levels input terminal and many levels output end, a level input terminal is the same as special
The level output end connection for levying straton network, for extracting candidate region from different levels fusion feature layer；
Classification returns sub-network, is equipped with many levels input terminal and RPN input terminals, a level input terminal is the same as characteristic layer sub-network
One level output end connection, RPN input terminals generate the output end connection of sub-network with candidate region, for reflecting candidate region
It is incident upon different levels fusion feature layer and obtains mapping treated multiple level fusion feature layers, and that treated is multiple to mapping
Level fusion feature layer carries out target discrimination and exports result.
2. remote sensing image time critical target identifying system as described in claim 1, which is characterized in that characteristic layer sub-network includes more
A characteristic layer submodule, be denoted as first characteristic layer submodule, second characteristic layer submodule ..., ith feature straton mould
Block ..., n-th characteristic layer submodule；1≤i≤N-1；
A level input terminal of one input terminal of ith feature straton module as this feature straton network, ith feature
Another input terminal of straton module with i+1 characteristic layer submodule output end connect, n-th characteristic layer submodule it is defeated
Enter a level input terminal of the end as this feature straton network；
Preceding N-1 characteristic layer submodule to last layer characteristic layer and current signature layer overlap-add procedure for obtaining present fusion feature
Layer；N-th characteristic layer submodule is used to export current signature layer as present fusion characteristic layer, wherein last layer characteristic layer
For by being obtained after carrying out convolution to current signature layer.
3. remote sensing image time critical target identifying system as claimed in claim 2, which is characterized in that preceding N-1 characteristic layer submodule
Any one characteristic layer submodule includes in block：
Last layer handles subelement, currently processed subelement and superpositing unit, and last layer handles the output end of subelement with superposition
The first input end of unit connects, and the output end of currently processed subelement is connected with the second input terminal of superpositing unit；
Last layer handle subelement be used for last layer characteristic layer carry out up-sampling treatment output processing after last layer characteristic layer, when
Pre-treatment subelement is used to be used for place current signature layer, superpositing unit after the progress process of convolution output processing of current signature layer
Last layer characteristic layer and current signature layer after processing are overlapped processing after reason, export present fusion characteristic layer.
4. remote sensing image time critical target identifying system as claimed in claim 2 or claim 3, which is characterized in that characteristic layer submodule is also
Including aliasing effect processing unit, input terminal is connected with the output end of superpositing unit, and aliasing effect processing unit is used for working as
Preceding fusion feature layer carries out process of convolution, exports final present fusion characteristic layer.
5. such as claim 2 to 4 any one of them remote sensing image time critical target identifying system, which is characterized in that feature straton
Network further includes the N+1 characteristic layer submodule, and candidate region generates sub-network and is equipped with extra play secondary input end；The the one N+1 special
The output end that straton module input is levied with n-th characteristic layer submodule connects, and the N+1 characteristic layer submodule output end is the same as time
Favored area generates the extra play secondary input end connection of sub-network；
The fusion feature layer that the N+1 characteristic layer submodule is used to export n-th characteristic layer submodule carries out up-sampling output
The N+1 characteristic layer, candidate region generate sub-network and extract candidate region from the N+1 characteristic layer simultaneously.
6. such as remote sensing image time critical target identifying system described in any one of claim 1 to 5, which is characterized in that classification returns
Sub-network includes sequentially connected mapping submodule, fusion submodule and target discrimination submodule, and mapping submodule will be for that will wait
Favored area maps to different levels fusion feature layer and obtains mapping treated multiple level fusion feature layers, and fusion submodule is used
In to mapping, treated that multiple level fusion feature layers carry out fusion treatments obtains target discrimination characteristic layers, target discrimination submodule
Block is used to carry out target discrimination to target discrimination characteristic layer to export result.
7. a kind of recognition methods based on remote sensing image time critical target identifying system described in claim 1, which is characterized in that packet
Include following steps：
S110 carries out multi-layer process of convolution to pending image, and using each level convolution processing result as pyramid feature
The hierarchy characteristic layer in layer；
I-th of hierarchy characteristic layer in pyramid characteristic layer and i+1 hierarchy characteristic layer overlap-add procedure are obtained i-th layer by S120
Grade fusion feature layer；And i is traversed into preceding N-1 level in pyramid characteristic layer and obtains N-1 level fusion feature layer；By golden word
N-th hierarchy characteristic layer is as n-th level fusion feature layer in tower characteristic layer；
S130 extracts candidate region from N number of level fusion feature layer；
Candidate region is mapped to N number of level fusion feature layer and obtains mapping treated multiple level fusion feature layers by S140,
And treated that multiple level fusion feature layers carry out target discriminations exports results to mapping.
8. recognition methods as claimed in claim 7, which is characterized in that between step S110 and step S120 further include as
Lower step：
Carrying out process of convolution to the i-th hierarchy characteristic layer makes the i-th hierarchy characteristic layer channel with to i+1 hierarchy characteristic layer channel phase
Together；Carrying out up-sampling treatment to i+1 hierarchy characteristic layer makes the i-th hierarchy characteristic layer size with to i+1 hierarchy characteristic layer ruler
It is very little identical；1≤i≤N-1.
9. recognition methods as claimed in claim 7 or 8, which is characterized in that step S140 includes following sub-step：
Step S141：Candidate region is mapped into N number of level fusion feature layer and obtains mapping treated N number of level fusion feature
Layer；
Step S142：And treated that N number of level fusion feature layer carries out fusion treatment obtains target discrimination characteristic layer to mapping；
Step S143：Target discrimination characteristic layer carries out target discrimination and exports result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810427107.2A CN108764063B (en) | 2018-05-07 | 2018-05-07 | Remote sensing image time-sensitive target identification system and method based on characteristic pyramid |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810427107.2A CN108764063B (en) | 2018-05-07 | 2018-05-07 | Remote sensing image time-sensitive target identification system and method based on characteristic pyramid |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108764063A true CN108764063A (en) | 2018-11-06 |
| CN108764063B CN108764063B (en) | 2020-05-19 |
Family
ID=64010105
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810427107.2A Active CN108764063B (en) | 2018-05-07 | 2018-05-07 | Remote sensing image time-sensitive target identification system and method based on characteristic pyramid |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108764063B (en) |
Cited By (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109671070A (en) * | 2018-12-16 | 2019-04-23 | 华中科技大学 | A kind of object detection method merged based on characteristic weighing and feature correlation |
| CN109800793A (en) * | 2018-12-28 | 2019-05-24 | 广州海昇教育科技有限责任公司 | A kind of object detection method and system based on deep learning |
| CN109816671A (en) * | 2019-01-31 | 2019-05-28 | 深兰科技（上海）有限公司 | A kind of object detection method, device and storage medium |
| CN109886286A (en) * | 2019-01-03 | 2019-06-14 | 武汉精测电子集团股份有限公司 | Object detection method, target detection model and system based on cascade detectors |
| CN109977963A (en) * | 2019-04-10 | 2019-07-05 | 京东方科技集团股份有限公司 | Image processing method, unit and computer-readable medium |
| CN109977956A (en) * | 2019-04-29 | 2019-07-05 | 腾讯科技（深圳）有限公司 | A kind of image processing method, device, electronic equipment and storage medium |
| CN110070072A (en) * | 2019-05-05 | 2019-07-30 | 厦门美图之家科技有限公司 | A method of generating object detection model |
| CN110321805A (en) * | 2019-06-12 | 2019-10-11 | 华中科技大学 | A kind of dynamic expression recognition methods based on sequential relationship reasoning |
| CN110348311A (en) * | 2019-06-13 | 2019-10-18 | 中国人民解放军战略支援部队信息工程大学 | A kind of intersection identifying system and method based on deep learning |
| CN110378297A (en) * | 2019-07-23 | 2019-10-25 | 河北师范大学 | A kind of Remote Sensing Target detection method based on deep learning |
| CN110490860A (en) * | 2019-08-21 | 2019-11-22 | 北京大恒普信医疗技术有限公司 | Diabetic retinopathy recognition methods, device and electronic equipment |
| CN110533105A (en) * | 2019-08-30 | 2019-12-03 | 北京市商汤科技开发有限公司 | A kind of object detection method and device, electronic equipment and storage medium |
| CN110909642A (en) * | 2019-11-13 | 2020-03-24 | 南京理工大学 | Remote sensing image target detection method based on multi-scale semantic feature fusion |
| CN111160293A (en) * | 2019-12-31 | 2020-05-15 | 珠海大横琴科技发展有限公司 | Small target ship detection method and system based on characteristic pyramid network |
| CN111191531A (en) * | 2019-12-17 | 2020-05-22 | 中南大学 | Rapid pedestrian detection method and system |
| CN111488777A (en) * | 2019-01-28 | 2020-08-04 | 北京地平线机器人技术研发有限公司 | Object identification method, object identification device and electronic equipment |
| CN111882581A (en) * | 2020-07-21 | 2020-11-03 | 青岛科技大学 | Multi-target tracking method for depth feature association |
| CN112330664A (en) * | 2020-11-25 | 2021-02-05 | 腾讯科技（深圳）有限公司 | Pavement disease detection method and device, electronic equipment and storage medium |
| CN113095356A (en) * | 2021-03-03 | 2021-07-09 | 北京邮电大学 | Light weight type neural network and image processing method and device |
| CN113361375A (en) * | 2021-06-02 | 2021-09-07 | 武汉理工大学 | Vehicle target identification method based on improved BiFPN |
| CN113658230A (en) * | 2020-05-12 | 2021-11-16 | 武汉Tcl集团工业研究院有限公司 | Optical flow estimation method, terminal and storage medium |
| WO2021249255A1 (en) * | 2020-06-12 | 2021-12-16 | 青岛理工大学 | Grabbing detection method based on rp-resnet |
| CN114842365A (en) * | 2022-07-04 | 2022-08-02 | 中国科学院地理科学与资源研究所 | Unmanned aerial vehicle aerial photography target detection and identification method and system |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105046197A (en) * | 2015-06-11 | 2015-11-11 | 西安电子科技大学 | Multi-template pedestrian detection method based on cluster |
| CN105894045A (en) * | 2016-05-06 | 2016-08-24 | 电子科技大学 | Vehicle type recognition method with deep network model based on spatial pyramid pooling |
| CN106683091A (en) * | 2017-01-06 | 2017-05-17 | 北京理工大学 | Target classification and attitude detection method based on depth convolution neural network |
| US20170364733A1 (en) * | 2015-08-26 | 2017-12-21 | Digitalglobe, Inc. | System for simplified generation of systems for broad area geospatial object detection |
-
2018
- 2018-05-07 CN CN201810427107.2A patent/CN108764063B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105046197A (en) * | 2015-06-11 | 2015-11-11 | 西安电子科技大学 | Multi-template pedestrian detection method based on cluster |
| US20170364733A1 (en) * | 2015-08-26 | 2017-12-21 | Digitalglobe, Inc. | System for simplified generation of systems for broad area geospatial object detection |
| CN105894045A (en) * | 2016-05-06 | 2016-08-24 | 电子科技大学 | Vehicle type recognition method with deep network model based on spatial pyramid pooling |
| CN106683091A (en) * | 2017-01-06 | 2017-05-17 | 北京理工大学 | Target classification and attitude detection method based on depth convolution neural network |
Cited By (38)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109671070A (en) * | 2018-12-16 | 2019-04-23 | 华中科技大学 | A kind of object detection method merged based on characteristic weighing and feature correlation |
| CN109671070B (en) * | 2018-12-16 | 2021-02-09 | 华中科技大学 | Target detection method based on feature weighting and feature correlation fusion |
| CN109800793B (en) * | 2018-12-28 | 2023-12-22 | 广州海昇教育科技有限责任公司 | Target detection method and system based on deep learning |
| CN109800793A (en) * | 2018-12-28 | 2019-05-24 | 广州海昇教育科技有限责任公司 | A kind of object detection method and system based on deep learning |
| CN109886286A (en) * | 2019-01-03 | 2019-06-14 | 武汉精测电子集团股份有限公司 | Object detection method, target detection model and system based on cascade detectors |
| CN111488777A (en) * | 2019-01-28 | 2020-08-04 | 北京地平线机器人技术研发有限公司 | Object identification method, object identification device and electronic equipment |
| CN109816671A (en) * | 2019-01-31 | 2019-05-28 | 深兰科技（上海）有限公司 | A kind of object detection method, device and storage medium |
| CN109977963B (en) * | 2019-04-10 | 2021-10-15 | 京东方科技集团股份有限公司 | Image processing method, apparatus, device and computer readable medium |
| CN109977963A (en) * | 2019-04-10 | 2019-07-05 | 京东方科技集团股份有限公司 | Image processing method, unit and computer-readable medium |
| US20210271917A1 (en) * | 2019-04-29 | 2021-09-02 | Tencent Technology (Shenzhen) Company Limited | Image processing method and apparatus, electronic device, and storage medium |
| US11741688B2 (en) * | 2019-04-29 | 2023-08-29 | Tencent Technology (Shenzhen) Company Limited | Image processing method and apparatus, electronic device, and storage medium |
| CN109977956A (en) * | 2019-04-29 | 2019-07-05 | 腾讯科技（深圳）有限公司 | A kind of image processing method, device, electronic equipment and storage medium |
| WO2020221013A1 (en) * | 2019-04-29 | 2020-11-05 | 腾讯科技（深圳）有限公司 | Image processing method and apparaus, and electronic device and storage medium |
| CN109977956B (en) * | 2019-04-29 | 2022-11-18 | 腾讯科技（深圳）有限公司 | Image processing method and device, electronic equipment and storage medium |
| CN110070072A (en) * | 2019-05-05 | 2019-07-30 | 厦门美图之家科技有限公司 | A method of generating object detection model |
| CN110321805A (en) * | 2019-06-12 | 2019-10-11 | 华中科技大学 | A kind of dynamic expression recognition methods based on sequential relationship reasoning |
| CN110348311B (en) * | 2019-06-13 | 2021-03-19 | 中国人民解放军战略支援部队信息工程大学 | Deep learning-based road intersection identification system and method |
| CN110348311A (en) * | 2019-06-13 | 2019-10-18 | 中国人民解放军战略支援部队信息工程大学 | A kind of intersection identifying system and method based on deep learning |
| CN110378297B (en) * | 2019-07-23 | 2022-02-11 | 河北师范大学 | Remote sensing image target detection method and device based on deep learning and storage medium |
| CN110378297A (en) * | 2019-07-23 | 2019-10-25 | 河北师范大学 | A kind of Remote Sensing Target detection method based on deep learning |
| CN110490860A (en) * | 2019-08-21 | 2019-11-22 | 北京大恒普信医疗技术有限公司 | Diabetic retinopathy recognition methods, device and electronic equipment |
| CN110533105B (en) * | 2019-08-30 | 2022-04-05 | 北京市商汤科技开发有限公司 | Target detection method and device, electronic equipment and storage medium |
| CN110533105A (en) * | 2019-08-30 | 2019-12-03 | 北京市商汤科技开发有限公司 | A kind of object detection method and device, electronic equipment and storage medium |
| CN110909642A (en) * | 2019-11-13 | 2020-03-24 | 南京理工大学 | Remote sensing image target detection method based on multi-scale semantic feature fusion |
| CN111191531A (en) * | 2019-12-17 | 2020-05-22 | 中南大学 | Rapid pedestrian detection method and system |
| CN111160293A (en) * | 2019-12-31 | 2020-05-15 | 珠海大横琴科技发展有限公司 | Small target ship detection method and system based on characteristic pyramid network |
| CN113658230A (en) * | 2020-05-12 | 2021-11-16 | 武汉Tcl集团工业研究院有限公司 | Optical flow estimation method, terminal and storage medium |
| CN113658230B (en) * | 2020-05-12 | 2024-05-28 | 武汉Tcl集团工业研究院有限公司 | Optical flow estimation method, terminal and storage medium |
| WO2021249255A1 (en) * | 2020-06-12 | 2021-12-16 | 青岛理工大学 | Grabbing detection method based on rp-resnet |
| CN111882581A (en) * | 2020-07-21 | 2020-11-03 | 青岛科技大学 | Multi-target tracking method for depth feature association |
| CN111882581B (en) * | 2020-07-21 | 2022-10-28 | 青岛科技大学 | Multi-target tracking method for depth feature association |
| CN112330664A (en) * | 2020-11-25 | 2021-02-05 | 腾讯科技（深圳）有限公司 | Pavement disease detection method and device, electronic equipment and storage medium |
| CN112330664B (en) * | 2020-11-25 | 2022-02-08 | 腾讯科技（深圳）有限公司 | Pavement disease detection method and device, electronic equipment and storage medium |
| CN113095356B (en) * | 2021-03-03 | 2023-10-31 | 北京邮电大学 | Light-weight neural network system and image processing method and device |
| CN113095356A (en) * | 2021-03-03 | 2021-07-09 | 北京邮电大学 | Light weight type neural network and image processing method and device |
| CN113361375B (en) * | 2021-06-02 | 2022-06-07 | 武汉理工大学 | Vehicle target identification method based on improved BiFPN |
| CN113361375A (en) * | 2021-06-02 | 2021-09-07 | 武汉理工大学 | Vehicle target identification method based on improved BiFPN |
| CN114842365A (en) * | 2022-07-04 | 2022-08-02 | 中国科学院地理科学与资源研究所 | Unmanned aerial vehicle aerial photography target detection and identification method and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108764063B (en) | 2020-05-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108764063A (en) | A kind of pyramidal remote sensing image time critical target identifying system of feature based and method | |
| CN107944442B (en) | Based on the object test equipment and method for improving convolutional neural networks | |
| Peng et al. | Detecting heads using feature refine net and cascaded multi-scale architecture | |
| CN112861978B (en) | Multi-branch feature fusion remote sensing scene image classification method based on attention mechanism | |
| CN107564025B (en) | Electric power equipment infrared image semantic segmentation method based on deep neural network | |
| CN103955702B (en) | SAR image terrain classification method based on depth RBF network | |
| CN108564097A (en) | A kind of multiscale target detection method based on depth convolutional neural networks | |
| CN108241871A (en) | Laser point cloud and visual fusion data classification method based on multiple features | |
| Xiao et al. | Enhancing multiscale representations with transformer for remote sensing image semantic segmentation | |
| CN109543632A (en) | A kind of deep layer network pedestrian detection method based on the guidance of shallow-layer Fusion Features | |
| Wang et al. | Semantic segmentation of remote sensing ship image via a convolutional neural networks model | |
| Zhang et al. | Self-attention guidance and multiscale feature fusion-based UAV image object detection | |
| Liu et al. | Coastline extraction method based on convolutional neural networks—A case study of Jiaozhou Bay in Qingdao, China | |
| CN109670555A (en) | Instance-level pedestrian detection and pedestrian's weight identifying system based on deep learning | |
| Tang et al. | Integrated feature pyramid network with feature aggregation for traffic sign detection | |
| Chen et al. | ASF-Net: Adaptive screening feature network for building footprint extraction from remote-sensing images | |
| CN111639697B (en) | Hyperspectral image classification method based on non-repeated sampling and prototype network | |
| Zhai et al. | An improved faster R-CNN pedestrian detection algorithm based on feature fusion and context analysis | |
| Tu et al. | Hyperspectral image classification using a superpixel–pixel–subpixel multilevel network | |
| CN104463210B (en) | Classification of Polarimetric SAR Image method based on object-oriented and spectral clustering | |
| Li et al. | DSPCANet: Dual-channel scale-aware segmentation network with position and channel attentions for high-resolution aerial images | |
| Hu et al. | Supervised multi-scale attention-guided ship detection in optical remote sensing images | |
| Li et al. | Adaptive mask sampling and manifold to Euclidean subspace learning with distance covariance representation for hyperspectral image classification | |
| CN109284752A (en) | A kind of rapid detection method of vehicle | |
| CN116503602A (en) | Unstructured environment three-dimensional point cloud semantic segmentation method based on multi-level edge enhancement |
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: 20240315 Address after: Room A348, 4th Floor, Building 1, Phase III, International Enterprise Center, No. 1 Guanggu Avenue, Donghu New Technology Development Zone, Wuhan City, Hubei Province, 430073 (Wuhan Free Trade Zone) Patentee after: AVIC Tianhai (Wuhan) Technology Co.,Ltd. Country or region after: China Address before: 430074 Hubei Province, Wuhan city Hongshan District Luoyu Road No. 1037 Patentee before: HUAZHONG University OF SCIENCE AND TECHNOLOGY Country or region before: China |