CN109284704A - Complex background SAR vehicle target detection method based on CNN - Google Patents
Complex background SAR vehicle target detection method based on CNN Download PDFInfo
- Publication number
- CN109284704A CN109284704A CN201811047934.5A CN201811047934A CN109284704A CN 109284704 A CN109284704 A CN 109284704A CN 201811047934 A CN201811047934 A CN 201811047934A CN 109284704 A CN109284704 A CN 109284704A
- Authority
- CN
- China
- Prior art keywords
- frame
- network model
- target detection
- layer
- cnn
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/08—Detecting or categorising vehicles
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Image Analysis (AREA)
Abstract
The present invention discloses a kind of complex background SAR vehicle target detection method based on CNN, comprising steps of S1, acquires pattern data and obtain sample data set through processing;S2 forms fusion frame, carries out retraining to the fusion frame on the basis of pre-training weight after being merged ResNet and Faster-RCNN frame;S3 carries out Target detection and identification to the pattern data using the fusion frame after retraining;The present invention realizes that target detection process realizes full-automatic target detection end to end by merging ResNet and Faster-RCNN frame, using Faster-RCNN frame, convenient for engineering application;The phenomenon that solving the problems, such as that network model present in depth convolutional network model is degenerated using residual error network model simultaneously, gradient existing for depth convolutional network model avoided to disappear.
Description
Technical field
The present invention relates to vehicle target detection technique fields, and in particular to a kind of complex background SAR vehicle mesh based on CNN
Mark detection method.
Background technique
The picture characteristics of synthetic aperture radar (Synthetic Aperture Radar, SAR) image can be with different
Imaging parameters, imaging posture, substance environment etc. have greatly changed so that the Target detection and identification of SAR image becomes non-
It is often difficult.It is traditional based on constant false alarm rate (Constant False Alarm Rate, CFAR) algorithm its derivative algorithm in target
Under background contrast with higher and scene simple scenario, detection threshold can preferably isolate target from background, when
When in face of the totally different clutter of many kinds of, scattering properties, usual detection performance can be declined.
With the continuous development of artificial intelligence, the method for deep learning is also introduced into SAR object detection field.Convolution mind
It is artificial nerve network model (Artificial through network model (Convolutional Neural Network, CNN)
Neural NetWork, ANN) one kind, the strategy that uses weight shared due to it reduces parameter, the translation, contracting to image
Put, tilt or other forms deformation have height invariance, therefore be widely used in two dimensional image target detection and
In identification.CNN can take out the feature of different levels in the study of different phase, avoid in conventional machines learning algorithm
The process of artificial design features and classifier.But its drawback is it is also obvious that the size dimension of its input picture needs to fix, nothing
Method realizes the target detection under large scene end to end.
Summary of the invention
To solve above-mentioned technological deficiency, the technical solution adopted by the present invention is, provides a kind of complex background based on CNN
SAR vehicle target detection method, comprising steps of
S1 acquires pattern data and obtains sample data set through processing;
S2 forms fusion frame, on the basis of pre-training weight after being merged ResNet and Faster-RCNN frame
On to the fusion frame carry out retraining;
S3 carries out Target detection and identification to the pattern data using the fusion frame after retraining.
Preferably, Same Scene is imaged by different radars, by data prediction sample orientation and away from
Descriscent is all the pattern data of 0.3m distance amplitude.
Preferably, manually extracting to vehicle sample in the pattern data, original sample is obtained;To the original sample
This, which expand, forms vehicle sample, and the vehicle sample is synthesized sample number with background sample in a manner of inserting at random
According to collection.
Preferably, using large data collection ILSVRC-2012 data set in ZF network model and VGG-16 network model
Random initializtion parameter carry out the pre-training, using ResNet-50 data set to random in ResNet-50 network model
Initiation parameter carries out the pre-training.
Preferably, utilizing the sample data on the network model feature extraction layer parameter basis after the pre-training
Collect and the retraining is carried out to feature extraction layer, candidate extract layer and the Classification and Identification layer in the fusion frame.
Preferably, in the step S2, the Faster-RCNN frame include feature extraction layer, Area generation network,
The pond ROI layer and classification extract layer;The feature extraction layer is by extracting the characteristic pattern of the pattern data as the classification
The input of identification layer;The Area generation network by the last layer characteristic pattern that the feature extraction layer inputs sliding window mention
Take candidate frame;The pond ROI layer collects the characteristic pattern and the candidate frame of input, the comprehensive characteristic pattern and the time
It selects and extracts candidate frame characteristic pattern after frame, the Classification and Identification layer carries out Classification and Identification to the candidate frame characteristic pattern, and carries out the
Two frames return;The Faster-RCNN frame uses one of ZF network model or VGG network model.
Preferably, 9 sizes and length and width are arranged to each of characteristic pattern pixel in the Area generation network
The candidate frame is tentatively obtained than different anchor points and in conjunction with the recurrence of the first frame.
Preferably, ResNet-50 network model and Faster RCNN frame are carried out fusion shape in the step S2
At fusion frame;The fusion frame handles the pattern data;The fusion frame is according to the ResNet-50 net
Network model obtains 50 layers of residual block, and the residual block includes feature residual block and classification residual block, the feature residual block setting
It is 40 layers, the classification residual block is set as 10 layers.
Preferably, the feature extraction layer is set as the feature residual block, the spy of the feature extraction layer output
Sign figure size remains the 1/16 of the pattern data.
Preferably, the Classification and Identification layer after the layer of the pond ROI substitutes the ZF with the classification residual block
Two layers of full articulamentum in network model.
Compared with the prior art the beneficial effects of the present invention are: 1, the present invention is by ResNet and Faster-RCNN frame
It is merged, realizes that target detection process realizes full-automatic target detection end to end using Faster-RCNN frame, just
It is applied in engineering;Asking for the degeneration of network model present in depth convolutional network model is solved using residual error network model simultaneously
The phenomenon that inscribing, gradient existing for depth convolutional network model avoided to disappear;2, the ResNet-50 based on Faster-RCNN frame
Network model can not only obtain good detection effect, under the support of NVIDIA video card Tesla K40m, using CUDA+
The GPU of CUDNN accelerates, and reaches 0.46s for the average time-consuming of the SAR image target detection of 2000*2000 or so size, also reaches
It is horizontal very high real-time has been arrived.
Detailed description of the invention
Fig. 1 is the exemplary diagram of original sample;
Fig. 2 is the exemplary diagram of samples pictures and labeling position;
Fig. 3 is the frame diagram of Faster-RCNN;
Fig. 4 is transfer learning and directly trained comparison diagram;
Fig. 5 is the training loss curve graph of the fusion frame;
Fig. 6 is the comparison diagram of object detection results and true annotation results.
Digital representation in figure:
1- original sample;2- background sample;3- transfer learning loses curve;4- directly training loss curve.
Specific embodiment
Below in conjunction with attached drawing, the forgoing and additional technical features and advantages are described in more detail.
Embodiment one
The present invention is based on the complex background SAR vehicle target detection method of CNN comprising steps of
S1 acquires pattern data and obtains sample data set through processing;
S2 is formed after being merged ResNet (network model depth residual error network model) and Faster-RCNN frame
Frame structure is merged, retraining is carried out to the fusion frame on the basis of pre-training weight;
S3 carries out Target detection and identification to the pattern data using the fusion frame after retraining.
Step S1 is pre- by data specifically, Same Scene is imaged in the airborne X-band radar by different flight numbers
Processing samples orientation and distance to being all the pattern data of 0.3m distance amplitude, is fixed using 128 × 128 (unit is pixel)
Size manually extracts vehicle sample in the pattern data, in total acquisition obtain 500 comprising a variety of vehicles (truck,
Bus and crane) original sample 1 original sample slice of data collection, as shown in FIG. 1, FIG. 1 is the exemplary diagrams of original sample;
A, b, c are the different original samples 1 in Fig. 1.
To the original sample 1 to rotate, overturning and the mode of (multiplicative noise) of making an uproar is added to be expanded, to increase the original
Sample type in beginning sample slice data set, so that formation includes the vehicle sample data set of several vehicle samples;It chooses
Several width scenes are as background sample 2 (including road, the interference such as building), the vehicle that the vehicle sample data is concentrated
Sample is synthesized in a manner of inserting at random with the background sample, and vehicle sample database, i.e. sample are finally fabricated to
Data set, the vehicle sample database generally comprise the sample graph of 7500 width about 2000 × 2000 (unit is pixel) size
Piece.
Preferably, background sample 2 described in each width inserts the 5-15 different vehicle samples at random in synthesis process
This.
The format of the vehicle sample database finally obtained generally uses VOC formatted data, that is, includes figure
(JREGImages) file and annotation (Annotations) file;The graphic file clip pack is containing the sample after synthesis
Picture, xml document of the comment file clip pack containing corresponding mark bounding box (rectangular selection frame) position.Fig. 2 is
The exemplary diagram of samples pictures and bounding box labeling position;A and c in Fig. 2 are different samples pictures, and b is a's
The exemplary diagram of bounding box labeling position, d are the exemplary diagram of the bounding box labeling position of c.
The training of CNN network model needs the support of a large amount of marker samples, and detects and identify field in SAR image at present
In, due to the limitation of data source, using it is more be MSTAR data set and MiniSAR data set.It is each in MSTAR data set
The size of width image is 128 × 128 (unit is pixel), the target comprising 37 kinds of models of major class, to a kind of model in data set
Target all acquire the image of different orientations and pitch angle.MSTAR data set is since picture size is smaller and background clutter
Interference very little, be generally used for Study on Target Recognition, be not suitable for the target detection of large scene image.MiniSAR data set
Picture size is 2510 × 1638 (unit is pixel), includes the target of many attitude in image, and include a variety of interference signals
(building, trees etc.), but data volume is seldom, can not train up to network model, can only examine as verifying collection
The precision of network model prediction.
And real vehicles are sliced sample by step S1 through the invention and SAR background image is merged and carrys out exptended sample
Step method avoids improving this hair for the situation of the vehicle target sample deficiency of large scene complexity SAR image in the prior art
The accuracy of the bright complex background SAR vehicle target detection method based on CNN.
Embodiment two
As shown in figure 3, Fig. 3 is the frame diagram of Faster-RCNN.It is needed in step s 2 through Faster-RCNN frame pair
The pattern data is handled.
Specifically, Faster-RCNN (Faster Region CNN) frame mainly includes feature extraction layer, Area generation
Network (Region Proposal Network, RPN), the pond ROI layer and classification extract layer.
The feature extraction layer mainly includes multiple convolutional layers, active coating and pond layer, and the feature extraction layer is to described
Pattern data extracts characteristic pattern as the input of the Classification and Identification layer.The feature extraction layer is according to network model
Depth can be divided into ZF network model and VGG network model, wherein the VGG-16 network model includes 13 conv (convolution
Layer), 13 relu (active coating) and 4 pooling (pond layer), the ZF network model include 5 conv, 4 relu and 2
The characteristic pattern size of a pooling, the ZF network model and VGG network model output is all the 1/ of the pattern data
16。
The Area generation network passes through the sliding window extraction time in the last layer characteristic pattern that the feature extraction layer inputs
Frame is selected, 9 sizes anchor point (anchor) different with length-width ratio is arranged to each of characteristic pattern pixel, and combine
Frame returns the candidate frame for tentatively obtaining the pattern data.
The pond ROI layer collects the characteristic pattern and the candidate frame of input, the comprehensive characteristic pattern and the time
Candidate frame characteristic pattern is extracted after selecting frame, the subsequent Classification and Identification layer is sent into and determines target category.
The Classification and Identification layer carries out Classification and Identification to the candidate frame characteristic pattern, and carries out more accurate frame and return,
The final target detection realized to the pattern data.
The Faster-RCNN frame four steps of target detection (characteristic area generation, feature extraction, classification and
Candidate region returns) unify within the frame of a depth network model, the complex background the present invention is based on CNN can be accelerated
The detection speed of SAR vehicle target detection method improves detection accuracy.
Embodiment three
Preferably, in the present embodiment, step S2 is specially by ResNet-50 network model and Faster RCNN frame
It carries out fusion and forms fusion frame;The fusion process of ResNet-50 network model and Faster RCNN frame specifically includes that
50 layers of residual block, the feature extraction layer setting of the fusion frame are obtained according to ResNet-50 network model
The characteristic pattern size for 40 layers of the residual block, output remains the 1/16 of the pattern data;
Sliding window is extracted candidate in the last layer characteristic pattern that the Area generation network is inputted by the feature extraction layer
Frame is arranged 9 sizes anchor point (anchor) different with length-width ratio to each of characteristic pattern pixel, and combines side
Frame returns the candidate frame for tentatively obtaining the pattern data.
The pond ROI layer collects the characteristic pattern and the candidate frame of input, the comprehensive characteristic pattern and the time
Candidate frame characteristic pattern is extracted after selecting frame, the subsequent Classification and Identification layer is sent into and determines target category.
The Classification and Identification layer after the layer of the pond ROI substitutes the ZF network mould with remaining 10 layers of residual block
Two layers of full articulamentum in type ultimately forms fusion frame and carries out target detection to pattern data.
By merging ResNet-50 network model and Faster RCNN frame, target detection end to end is realized
Process, to guarantee that the present invention is based on the full-automatic target detections of the complex background SAR vehicle target detection method of CNN, just
In engineering application of the invention;Network model present in depth convolutional network model is solved using residual error network model simultaneously
The problem of degeneration, avoids gradient extinction tests existing for depth convolutional network model, improves detection effect.
Example IV
Preferably, being trained using transfer learning, specifically include that
Pre-training: in step S2, large data collection ILSVRC-2012 data set (ImageNet Large Scale is utilized
The extensive visual identity contest of Visual Recognition Challenge, IMANEET) to ZF network model and VGG-16 net
Random initializtion parameter in network model carries out pre-training, using ResNet-50 data set in ResNet-50 network model
Random initializtion parameter carries out pre-training.
Retraining: on the network model feature extraction layer parameter basis after the pre-training, the sample data is utilized
Collection carries out retraining to the feature extraction layer, the candidate extract layer and the Classification and Identification layer.
The final fusion frame using after retraining completes Target detection and identification.
The network model parameter that the trained and retraining obtains described in through the invention carries out detection identification, detects
Recognition effect will be far superior to random initializtion parameter, and the training expense of network model can be greatly decreased.
As shown in figure 4, Fig. 4 is transfer learning and directly trained loss curve comparison figure, Fig. 4 includes transfer learning damage
Curve 3 and directly training loss curve 4 are lost, curve 3 is lost by the transfer learning and curve 4 is lost in the direct training
It compares, it can be seen that the transfer learning obviously accelerates the convergence rate of network model, and the penalty values after convergence are less than
Directly trained penalty values.
Embodiment five
Preferably, using Average Accuracy (mean average to the evaluation of the testing result of network model
Precision, mAP) it is used as detection effect evaluation criterion.The mAP calculation formula are as follows:
Wherein, P is precision ratio, and R is recall ratio.
MAP is that solve Traditional measurements normalized recall rate according to the integral of recall ratio and the drawn curve of precision ratio, look into standard
The One-Point-Value limitation of rate and F-Score (F score comprehensively considers the reconciliation value of P and R), therefore mAP is evaluated as detection effect and is marked
Will definitely more effectively comprehensive assessment algorithm validity and accuracy.
System is concentrated in 500 test samples to ZF network model, VGG-16 network model, fusion three kinds of network models of frame
It is as shown in table 1 to count mAP:
1 ZF network model of table, VGG-16 network model, fusion frame indicator-specific statistics table
By comparison T1, T2 and T3 experiment, it can be found that the testing result of fusion frame is optimal, but its individual figure is average
The time-consuming also longest of detection.The detection time-consuming of image of traditional two-parameter CFAR method in 2000*2000 or so size reaches 10s
Magnitude, and use the inspection of the ResNet-50 of GPU (Graphics Processing Unit graphics processor) progress parallel computation
Time-consuming only 0.46s is surveyed, there is high real-time.
As shown in figure 5, Fig. 5 is the training loss curve graph for merging frame under 4 kinds of different scenes.It can be with from Fig. 5
It obtains, just basic convergence is not declining after iteration 2000 times for trained loss, therefore defeated with regard to stopping iteration at iteration 3000 times
Input weight when network model weight is as test out.By object detection results of the invention and really in actual experiment
Annotation results compare statistics and (think to be greater than 50% detection in target area and true callout box overlapping area when statistics
Frame is regarded as effectively detecting), as shown in fig. 6, Fig. 6 is the comparison diagram of object detection results and true annotation results;A in Fig. 6,
B, c, d are respectively the object detection results of four different scenes and the comparison diagram of true annotation results, it can be found that four from Fig. 6
The precision ratio of a scene reaches 100%, and recall ratio reaches 95%, F-Score and reaches 97%.
The foregoing is merely presently preferred embodiments of the present invention, is merely illustrative for the purpose of the present invention, and not restrictive
's.Those skilled in the art understand that in the spirit and scope defined by the claims in the present invention many changes can be carried out to it,
It modifies or even equivalent, but falls in protection scope of the present invention.
Claims (10)
1. a kind of complex background SAR vehicle target detection method based on CNN, which is characterized in that comprising steps of
S1 acquires pattern data and obtains sample data set through processing;
S2 forms fusion frame after being merged ResNet and Faster-RCNN frame, right on the basis of pre-training weight
The fusion frame carries out retraining;
S3 carries out Target detection and identification to the pattern data using the fusion frame after retraining.
2. the complex background SAR vehicle target detection method based on CNN as described in claim 1, which is characterized in that by not
Same Scene is imaged with radar, samples orientation and distance to being all 0.3m distance amplitude by data prediction
The pattern data.
3. the complex background SAR vehicle target detection method based on CNN as claimed in claim 2, which is characterized in that artificial right
Vehicle sample extracts in the pattern data, obtains original sample;The original sample expand and forms vehicle sample
This, and the vehicle sample is synthesized into sample data set with background sample in a manner of inserting at random.
4. the complex background SAR vehicle target detection method based on CNN as described in claim 1, which is characterized in that using greatly
Type data set ILSVRC-2012 data set carries out institute to the random initializtion parameter in ZF network model and VGG-16 network model
Pre-training is stated, the pre- instruction is carried out to the random initializtion parameter in ResNet-50 network model using ResNet-50 data set
Practice.
5. the complex background SAR vehicle target detection method based on CNN as described in claim 1, which is characterized in that described
On network model feature extraction layer parameter basis after pre-training, using the sample data set to the spy in the fusion frame
It levies extract layer, candidate extract layer and Classification and Identification layer and carries out the retraining.
6. the complex background SAR vehicle target detection method based on CNN as described in claim 1, which is characterized in that described
In step S2, the Faster-RCNN frame includes that feature extraction layer, Area generation network, the pond ROI layer and classification are extracted
Layer;The feature extraction layer is by extracting input of the characteristic pattern of the pattern data as the Classification and Identification layer;The area
Domain generates network and passes through the sliding window extraction candidate frame in the last layer characteristic pattern that the feature extraction layer inputs;The pond ROI
Change the characteristic pattern and the candidate frame that layer collects input, it is special to extract candidate frame after the comprehensive characteristic pattern and the candidate frame
Sign figure, the Classification and Identification layer carries out Classification and Identification to the candidate frame characteristic pattern, and carries out the second frame recurrence;It is described
Faster-RCNN frame uses one of ZF network model or VGG network model.
7. the complex background SAR vehicle target detection method based on CNN as claimed in claim 6, which is characterized in that the area
Domain generates network and 9 sizes anchor point different with length-width ratio is arranged to each of characteristic pattern pixel and combines first
Frame recurrence tentatively obtains the candidate frame.
8. the complex background SAR vehicle target detection method based on CNN as claimed in claim 7, which is characterized in that described
In step S2, ResNet-50 network model and Faster RCNN frame are subjected to fusion and form fusion frame;The fusion frame
Frame handles the pattern data;The fusion frame obtains 50 layers of residual block according to the ResNet-50 network model,
The residual block includes feature residual block and classification residual block, and the feature residual block is set as 40 layers, the classification residual block
It is set as 10 layers.
9. the complex background SAR vehicle target detection method based on CNN as claimed in claim 8, which is characterized in that the spy
Sign extract layer is set as the feature residual block, and the characteristic pattern size of the feature extraction layer output remains the pattern
The 1/16 of data.
10. the complex background SAR vehicle target detection method based on CNN as claimed in claim 8, which is characterized in that in institute
It states the Classification and Identification layer after the layer of the pond ROI and substitutes two layers in the ZF network model with the classification residual block and connect entirely
Connect layer.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811047934.5A CN109284704A (en) | 2018-09-07 | 2018-09-07 | Complex background SAR vehicle target detection method based on CNN |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811047934.5A CN109284704A (en) | 2018-09-07 | 2018-09-07 | Complex background SAR vehicle target detection method based on CNN |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109284704A true CN109284704A (en) | 2019-01-29 |
Family
ID=65183914
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811047934.5A Pending CN109284704A (en) | 2018-09-07 | 2018-09-07 | Complex background SAR vehicle target detection method based on CNN |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109284704A (en) |
Cited By (18)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109901129A (en) * | 2019-03-06 | 2019-06-18 | 中国人民解放军海军航空大学 | Object detection method and system in a kind of sea clutter |
CN110059654A (en) * | 2019-04-25 | 2019-07-26 | 台州智必安科技有限责任公司 | A kind of vegetable Automatic-settlement and healthy diet management method based on fine granularity identification |
CN110110722A (en) * | 2019-04-30 | 2019-08-09 | 广州华工邦元信息技术有限公司 | A kind of region detection modification method based on deep learning model recognition result |
CN110136134A (en) * | 2019-04-03 | 2019-08-16 | 深兰科技(上海)有限公司 | A kind of deep learning method, apparatus, equipment and medium for road surface segmentation |
CN110379178A (en) * | 2019-07-25 | 2019-10-25 | 电子科技大学 | Pilotless automobile intelligent parking method based on millimetre-wave radar imaging |
CN110503643A (en) * | 2019-08-23 | 2019-11-26 | 闽江学院 | A kind of object detection method and device based on the retrieval of multiple dimensioned rapid scene |
CN110826457A (en) * | 2019-10-31 | 2020-02-21 | 上海融军科技有限公司 | Vehicle detection method and device under complex scene |
CN110929632A (en) * | 2019-11-19 | 2020-03-27 | 复旦大学 | Complex scene-oriented vehicle target detection method and device |
CN111209975A (en) * | 2020-01-13 | 2020-05-29 | 北京工业大学 | Ship target identification method based on multitask learning |
CN111292349A (en) * | 2020-01-17 | 2020-06-16 | 北京大学深圳研究生院 | Data enhancement method for target detection based on fusion of recommendation candidate boxes |
CN111582339A (en) * | 2020-04-28 | 2020-08-25 | 江西理工大学 | Vehicle detection and identification method based on deep learning |
CN111784031A (en) * | 2020-06-15 | 2020-10-16 | 上海东普信息科技有限公司 | Logistics vehicle classification prediction method, device, equipment and storage medium |
CN112133100A (en) * | 2020-09-16 | 2020-12-25 | 北京影谱科技股份有限公司 | Vehicle detection method based on R-CNN |
CN112132032A (en) * | 2020-09-23 | 2020-12-25 | 平安国际智慧城市科技股份有限公司 | Traffic sign detection method and device, electronic equipment and storage medium |
CN112580408A (en) * | 2019-09-30 | 2021-03-30 | 杭州海康威视数字技术股份有限公司 | Deep learning model training method and device and electronic equipment |
CN112734641A (en) * | 2020-12-31 | 2021-04-30 | 百果园技术(新加坡)有限公司 | Training method and device of target detection model, computer equipment and medium |
CN114463729A (en) * | 2021-12-13 | 2022-05-10 | 浙江大华技术股份有限公司 | Milk cow identification method and device |
CN114973023A (en) * | 2022-08-01 | 2022-08-30 | 中国科学院空天信息创新研究院 | High-resolution SAR image vehicle target key part extraction method based on fast RCNN |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107945173A (en) * | 2017-12-11 | 2018-04-20 | 深圳市宜远智能科技有限公司 | A kind of skin disease detection method and system based on deep learning |
CN108280460A (en) * | 2017-12-04 | 2018-07-13 | 西安电子科技大学 | Based on the SAR vehicle target recognition methods for improving convolutional neural networks |
CN108288271A (en) * | 2018-02-06 | 2018-07-17 | 上海交通大学 | Image detecting system and method based on three-dimensional residual error network |
-
2018
- 2018-09-07 CN CN201811047934.5A patent/CN109284704A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108280460A (en) * | 2017-12-04 | 2018-07-13 | 西安电子科技大学 | Based on the SAR vehicle target recognition methods for improving convolutional neural networks |
CN107945173A (en) * | 2017-12-11 | 2018-04-20 | 深圳市宜远智能科技有限公司 | A kind of skin disease detection method and system based on deep learning |
CN108288271A (en) * | 2018-02-06 | 2018-07-17 | 上海交通大学 | Image detecting system and method based on three-dimensional residual error network |
Non-Patent Citations (3)
Title |
---|
LI, HX 等: "Learning Deep Appearance Feature for Multi-target Tracking", 《2017 INTERNATIONAL CONFERENCE ON VIRTUAL REALITY AND VISUALIZATION (ICVRV 2017)》 * |
SHAOQING REN等: "Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks", 《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》 * |
彭刚 等: "改进的基于区域卷积神经网络的微操作系统目标检测方法", 《模式识别与人工智能》 * |
Cited By (25)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109901129A (en) * | 2019-03-06 | 2019-06-18 | 中国人民解放军海军航空大学 | Object detection method and system in a kind of sea clutter |
CN110136134A (en) * | 2019-04-03 | 2019-08-16 | 深兰科技(上海)有限公司 | A kind of deep learning method, apparatus, equipment and medium for road surface segmentation |
CN110059654A (en) * | 2019-04-25 | 2019-07-26 | 台州智必安科技有限责任公司 | A kind of vegetable Automatic-settlement and healthy diet management method based on fine granularity identification |
CN110110722A (en) * | 2019-04-30 | 2019-08-09 | 广州华工邦元信息技术有限公司 | A kind of region detection modification method based on deep learning model recognition result |
CN110379178A (en) * | 2019-07-25 | 2019-10-25 | 电子科技大学 | Pilotless automobile intelligent parking method based on millimetre-wave radar imaging |
CN110503643A (en) * | 2019-08-23 | 2019-11-26 | 闽江学院 | A kind of object detection method and device based on the retrieval of multiple dimensioned rapid scene |
CN110503643B (en) * | 2019-08-23 | 2021-10-01 | 闽江学院 | Target detection method and device based on multi-scale rapid scene retrieval |
CN112580408B (en) * | 2019-09-30 | 2024-03-12 | 杭州海康威视数字技术股份有限公司 | Deep learning model training method and device and electronic equipment |
CN112580408A (en) * | 2019-09-30 | 2021-03-30 | 杭州海康威视数字技术股份有限公司 | Deep learning model training method and device and electronic equipment |
CN110826457A (en) * | 2019-10-31 | 2020-02-21 | 上海融军科技有限公司 | Vehicle detection method and device under complex scene |
CN110826457B (en) * | 2019-10-31 | 2022-08-19 | 上海融军科技有限公司 | Vehicle detection method and device under complex scene |
CN110929632A (en) * | 2019-11-19 | 2020-03-27 | 复旦大学 | Complex scene-oriented vehicle target detection method and device |
CN111209975A (en) * | 2020-01-13 | 2020-05-29 | 北京工业大学 | Ship target identification method based on multitask learning |
CN111292349A (en) * | 2020-01-17 | 2020-06-16 | 北京大学深圳研究生院 | Data enhancement method for target detection based on fusion of recommendation candidate boxes |
CN111292349B (en) * | 2020-01-17 | 2023-04-18 | 北京大学深圳研究生院 | Data enhancement method for target detection based on fusion of recommendation candidate boxes |
CN111582339A (en) * | 2020-04-28 | 2020-08-25 | 江西理工大学 | Vehicle detection and identification method based on deep learning |
CN111784031A (en) * | 2020-06-15 | 2020-10-16 | 上海东普信息科技有限公司 | Logistics vehicle classification prediction method, device, equipment and storage medium |
CN112133100A (en) * | 2020-09-16 | 2020-12-25 | 北京影谱科技股份有限公司 | Vehicle detection method based on R-CNN |
CN112133100B (en) * | 2020-09-16 | 2022-04-22 | 北京影谱科技股份有限公司 | Vehicle detection method based on R-CNN |
CN112132032A (en) * | 2020-09-23 | 2020-12-25 | 平安国际智慧城市科技股份有限公司 | Traffic sign detection method and device, electronic equipment and storage medium |
CN112734641A (en) * | 2020-12-31 | 2021-04-30 | 百果园技术(新加坡)有限公司 | Training method and device of target detection model, computer equipment and medium |
CN112734641B (en) * | 2020-12-31 | 2024-05-31 | 百果园技术(新加坡)有限公司 | Training method and device for target detection model, computer equipment and medium |
CN114463729A (en) * | 2021-12-13 | 2022-05-10 | 浙江大华技术股份有限公司 | Milk cow identification method and device |
CN114973023A (en) * | 2022-08-01 | 2022-08-30 | 中国科学院空天信息创新研究院 | High-resolution SAR image vehicle target key part extraction method based on fast RCNN |
CN114973023B (en) * | 2022-08-01 | 2022-10-04 | 中国科学院空天信息创新研究院 | High-resolution SAR image vehicle target key part extraction method based on fast RCNN |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109284704A (en) | Complex background SAR vehicle target detection method based on CNN | |
CN108510467B (en) | SAR image target identification method based on depth deformable convolution neural network | |
CN106874889B (en) | Multiple features fusion SAR target discrimination method based on convolutional neural networks | |
CN106156744B (en) | SAR target detection method based on CFAR detection and deep learning | |
CN105809198B (en) | SAR image target recognition method based on depth confidence network | |
CN105427314B (en) | SAR image object detection method based on Bayes's conspicuousness | |
CN108009509A (en) | Vehicle target detection method | |
CN108898065B (en) | Deep network ship target detection method with candidate area rapid screening and scale self-adaption | |
CN108334848A (en) | A kind of small face identification method based on generation confrontation network | |
CN103729854B (en) | A kind of method for detecting infrared puniness target based on tensor model | |
CN110163275B (en) | SAR image target classification method based on deep convolutional neural network | |
CN107247930A (en) | SAR image object detection method based on CNN and Selective Attention Mechanism | |
CN113012150A (en) | Feature-fused high-density rice field unmanned aerial vehicle image rice ear counting method | |
CN112257605B (en) | Three-dimensional target detection method, system and device based on self-labeling training sample | |
CN110097575B (en) | Target tracking method based on local features and scale pool | |
CN104573731A (en) | Rapid target detection method based on convolutional neural network | |
CN109389080A (en) | Hyperspectral image classification method based on semi-supervised WGAN-GP | |
CN103824079B (en) | Multi-level mode sub block division-based image classification method | |
CN110647802A (en) | Remote sensing image ship target detection method based on deep learning | |
CN113160062B (en) | Infrared image target detection method, device, equipment and storage medium | |
CN109886147A (en) | A kind of more attribute detection methods of vehicle based on the study of single network multiple-task | |
CN109635789B (en) | High-resolution SAR image classification method based on intensity ratio and spatial structure feature extraction | |
CN105488541A (en) | Natural feature point identification method based on machine learning in augmented reality system | |
CN106023134A (en) | Automatic grain boundary extraction method for steel grain | |
CN104657717A (en) | Pedestrian detection method based on layered kernel sparse representation |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190129 |