CN109871873A - Millimeter-wave image object detection and recognition method based on Fast R-CNN - Google Patents
Millimeter-wave image object detection and recognition method based on Fast R-CNN Download PDFInfo
- Publication number
- CN109871873A CN109871873A CN201910045667.6A CN201910045667A CN109871873A CN 109871873 A CN109871873 A CN 109871873A CN 201910045667 A CN201910045667 A CN 201910045667A CN 109871873 A CN109871873 A CN 109871873A
- Authority
- CN
- China
- Prior art keywords
- fast
- millimeter
- cnn
- wave image
- training
- 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.)
- Withdrawn
Links
Landscapes
- Image Analysis (AREA)
Abstract
The millimeter-wave image object detection and recognition method based on Fast R-CNN that the invention discloses a kind of, includes the following steps;S1: conditional depth convolution generates confrontation network (C-DCGAN) model;S2: being added depth convolution for condition and generate the generator fought in network, is aided with training plus condition using the ability that convolutional network extracts feature;S3: the arbiter part in trained C-DCGAN is extracted, and forms the new network structure for being used for image recognition after adding Softinax;Feature is extracted using this method and is used for image recognition, and condition depth convolution confrontation network can not only generate expected sample under the limitation of condition, but also more high using the ability that convolutional layer extracts feature.
Description
Technical field
The present invention relates to detection identification technology field more particularly to a kind of millimeter-wave image targets based on Fast R-CNN
Detection and recognition methods.
Background technique
Millimeter wave refers to that wavelength is the electromagnetic wave of 1-10mm, is located at the wave-length coverage that microwave and far infrared wave overlap mutually,
Also there is oneself unique characteristic while with two kinds of spectral characteristics of microwave and far infrared wave.Compared with infrared, millimeter wave
Atmospheric attenuation is small, and the ability for distinguishing metal target and ambient enviroment is strong, and compared with microwave, the directive property of millimeter wave is good, anti-interference
Ability is strong, detection performance is good, and due to the unique property that millimeter wave has, having millimeter-wave radiation device can be in cloud, mist, flue dust etc.
The ability to work all-time anf all-weather is realized under harsh environmental conditions.Since millimeter wave has preferable penetrability and higher
Spatial resolution can be imaged the prohibited items being hidden under clothing using millimeter-wave radiation device, to reach detection
The purpose of identification.
CNN (convolutional neural networks) is a kind of efficient identification method that developed recently gets up, and is a kind of feedforward neural network,
Its artificial neuron can respond the surrounding cells in a part of coverage area, can directly input original image and be identified, be kept away
The pretreatment complicated early period to image is exempted from.R-CNN is the new model of CNN, for the convolutional neural networks model based on region,
It generates candidate region according to Selective Search or Edge boxes, then with convolutional neural networks to the time of generation
Favored area carries out feature extraction, and Fast R-CNN is further improved on the basis of R-CNN, proposes the pond ROI
Layer, further improves the accuracy and candidate region computational efficiency of identification.
Summary of the invention
The millimeter-wave image Target detection and identification method based on Fast R-CNN that the purpose of the present invention is to provide a kind of,
The detection identification of object is realized by the combination of mm-wave imaging technology and Fast R-CNN technology, the method has
The features such as detection performance is good, strong antijamming capability, directive property is strong, recognition accuracy is high, response quickly.
In order to achieve the above objectives, the technical solution of the present invention is as follows:
A kind of millimeter-wave image Target detection and identification method based on Fast R-CNN, includes the following steps:
S1, the millimeter-wave image sample for obtaining object carry out tag processes and are fabricated to sample set;
S2, training set is input to Fast R-CNN network, extracts object using the convolution kernel in Fast R-CNN
The feature of millimeter-wave image carries out multitask training as feature vector;
S3, the millimeter-wave image that object to be identified is obtained by the acquisition of millimeter wave data acquisition equipment;
The trained Fast R-CNN network of millimeter-wave image input of S4, the object to be identified for obtaining acquisition, utilize
Fast R-CNN network carries out object identification.
Further, the sample in the step S1 has carried out enhancing processing.
Further, the sample enhancing processing mode in the step S1 is first to carry out left and right overturning to original image block, then right
Segment after original image block and left and right overturning is spun upside down respectively, and the segment after overturning is all classified as sample.
Further, in the step S1, the sample set is divided into training set, verifying collection and test set;Training set is used for
Training Fast R-CNN network, verifying collection for observing whether the loss function of Fast R-CNN network is received in the training process
It holds back, to judge whether to terminate training, test set is used to test the classification accuracy of Fast R-CNN network.
Further, the Fast R-CNN network include 13 convolutional layers, 4 pond layers, 1 pond ROI layer, 2
Full articulamentum and 2 sane level layers.
Further, the complete connection output of multitask training Fast R-CNN network wrap cls_score layers with
Bbox_pred layers, described cls_score layers is used to classify, and described bbox_pred layers for adjusting candidate frame position.
Advantageous effects of the invention are as follows: the present invention is organic by mm-wave imaging technology and Fast R-CNN technology
It is implemented in combination with the detection identification of object, good, strong antijamming capability, directive property are strong with detection performance for the method, it is quasi- to identify
The features such as exactness height, response quickly.
Specific embodiment
Technical solution of the present invention is clearly and completely described below in conjunction with specific embodiment, it is clear that described
Embodiment be only a part of the embodiment of the present invention, instead of all the embodiments.Based on the embodiment of the present invention, this field
Those of ordinary skill's every other embodiment obtained, belongs to protection scope of the present invention.
A kind of millimeter-wave image Target detection and identification method based on Fast R-CNN, comprising:
S1, the millimeter-wave image sample for obtaining object carry out tag processes and are fabricated to sample set;
S2, training set is input to Fast R-CNN network, extracts object using the convolution kernel in Fast R-CNN
The feature of millimeter-wave image carries out multitask training as feature vector;
S3, the millimeter wave figure that object to be identified is obtained by millimeter wave data acquisition equipment (such as millimeter wave transceiver) acquisition
Picture;
The trained Fast R-CNN network of millimeter-wave image input of S4, the object to be identified for obtaining acquisition, utilize
Fast R-CNN network carries out object identification.
Sample in the S1 has carried out enhancing processing, first carries out left and right overturning to original image block, then to original image block and a left side
Segment after right overturning is spun upside down respectively, and the segment after overturning is all classified as sample.By a certain percentage by sample set
It is divided into training set, verifying collection and test set;Training set is for training Fast R-CNN network, and verifying collection is in training process
Whether the loss function of middle observation Fast R-CNN network restrains, and to judge whether to terminate training, test set is for testing Fast
The classification accuracy of R-CNN network.
The Fast R-CNN network includes 13 convolutional layers, 4 pond layers, 1 pond ROI layer, 2 full articulamentums
With 2 sane level layers.Original layer parameter needs to initialize by training method, and the full connection for classification is with mean value 0, standard deviation
It is initialized for 0.01 Gaussian Profile;For recurrence full articulamentum with mean value be 0, standard deviation be 0.001 Gaussian Profile into
Row initialization, biasing are initialized to 0.
In tuning training, N full pictures are firstly added, the R candidate frame chosen from N picture is then added.
The shared calculating of R/N candidate frame convolution of same image and memory, reduce computing overhead.The composition of R candidate frame is such as
Under: the candidate frame for overlapping [0.5,1] with some true value is defined as prospect, accounts for the 25% of total amount;The maximum value Chong Die with true value
Candidate frame in [0.1,0.5] is defined as background, accounts for the 75% of total amount.
The method of the Fast R-CNN Network Recognition object are as follows: the millimeter image of object to be identified is inputted into Fast R-
CNN network obtains characteristic pattern by several convolutional layers and pond layer;Using characteristic pattern mapping relations, found in characteristic pattern every
The corresponding feature frame of a candidate frame, and by each feature frame, this word arrives fixed size in the layer of the pond ROI;By feature frame by complete
Articulamentum obtains the feature vector of fixed size, and described eigenvector respectively obtains classification score via respective full articulamentum
Rear hatch returns two output vectors;All results are subjected to non-maxima suppression processing and generate final Target detection and identification
As a result.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Within mind and principle, all any modification, equivalent substitution, improvement and etc. be should all be included in the protection scope of the present invention.
Claims (6)
1. a kind of millimeter-wave image Target detection and identification method based on Fast R-CNN, which is characterized in that including walking as follows
It is rapid:
S1, the millimeter-wave image sample for obtaining object carry out tag processes and are fabricated to sample set;
S2, training set is input to Fast R-CNN network, extracts object millimeter using the convolution kernel in Fast R-CNN
The feature of wave image carries out multitask training as feature vector;
S3, the millimeter-wave image that object to be identified is obtained by the acquisition of millimeter wave data acquisition equipment;
The trained Fast R-CNN network of millimeter-wave image input of S4, the object to be identified for obtaining acquisition, utilize Fast R-
CNN network carries out object identification.
2. the millimeter-wave image Target detection and identification method based on Fast R-CNN, feature exist as described in claim 1
In the sample in the step S1 has carried out enhancing processing.
3. the millimeter-wave image Target detection and identification method based on Fast R-CNN, feature exist as claimed in claim 2
In the sample enhancing processing mode in the step S1 is first to carry out left and right overturning to original image block, then to original image block and left and right
Segment after overturning is spun upside down respectively, and the segment after overturning is all classified as sample.
4. the millimeter-wave image Target detection and identification method as described in any one of claims 1 to 3 based on Fast R-CNN,
It is characterized in that, the sample set is divided into training set, verifying collection and test set in the step S1;Training set is for training
Whether Fast R-CNN network, loss function of the verifying collection for observation Fast R-CNN network in the training process restrain, with
Judge whether to terminate training, test set is used to test the classification accuracy of Fast R-CNN network.
5. the millimeter-wave image Target detection and identification method based on Fast R-CNN, feature exist as claimed in claim 4
It include 13 convolutional layers in, the Fast R-CNN network, 4 pond layers, 1 pond ROI layer, 2 full articulamentums and 2 put down
Grade layer.
6. the millimeter-wave image Target detection and identification method based on Fast R-CNN, feature exist as claimed in claim 5
In the full connection output of the multitask training Fast R-CNN network includes cls_score layers and bbox_pred layers, described
Cls_score layers are used to classify, and described bbox_pred layers for adjusting candidate frame position.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910045667.6A CN109871873A (en) | 2019-01-17 | 2019-01-17 | Millimeter-wave image object detection and recognition method based on Fast R-CNN |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910045667.6A CN109871873A (en) | 2019-01-17 | 2019-01-17 | Millimeter-wave image object detection and recognition method based on Fast R-CNN |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109871873A true CN109871873A (en) | 2019-06-11 |
Family
ID=66917804
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910045667.6A Withdrawn CN109871873A (en) | 2019-01-17 | 2019-01-17 | Millimeter-wave image object detection and recognition method based on Fast R-CNN |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109871873A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111243223A (en) * | 2020-02-26 | 2020-06-05 | 福州大学 | Automobile anti-scratch monitoring alarm method and system |
-
2019
- 2019-01-17 CN CN201910045667.6A patent/CN109871873A/en not_active Withdrawn
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111243223A (en) * | 2020-02-26 | 2020-06-05 | 福州大学 | Automobile anti-scratch monitoring alarm method and system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Du et al. | Saliency-guided single shot multibox detector for target detection in SAR images | |
Kang et al. | A modified faster R-CNN based on CFAR algorithm for SAR ship detection | |
Sun et al. | Oriented ship detection based on strong scattering points network in large-scale SAR images | |
CN102496016B (en) | Infrared target detection method based on space-time cooperation framework | |
Kim | Double layered-background removal filter for detecting small infrared targets in heterogenous backgrounds | |
Wang et al. | Concealed object detection for millimeter-wave images with normalized accumulation map | |
Yuan et al. | A deep learning method for detection of dangerous equipment | |
Mou et al. | Sea clutter suppression for radar PPI images based on SCS-GAN | |
Wang et al. | Dense docked ship detection via spatial group-wise enhance attention in SAR images | |
Wang et al. | Study on the method of transmission line foreign body detection based on deep learning | |
Peng et al. | Improved YOLOX’s anchor-free SAR image ship target detection | |
Du et al. | An SAR target detector based on gradient harmonized mechanism and attention mechanism | |
Yang et al. | A novel deformable body partition model for MMW suspicious object detection and dynamic tracking | |
Song et al. | Multi-view HRRP generation with aspect-directed attention GAN | |
Hou et al. | An integrated method of ship detection and recognition in SAR images based on deep learning | |
CN109871873A (en) | Millimeter-wave image object detection and recognition method based on Fast R-CNN | |
Tao et al. | Radio frequency interference signature detection in radar remote sensing image using semantic cognition enhancement network | |
Zhang et al. | MSIF: Multisize inference fusion-based false alarm elimination for ship detection in large-scale SAR images | |
Lv et al. | ESP-LRSMD: A two-step detector for ship detection using SLC SAR imagery | |
Wang et al. | SCFNet: Semantic condition constraint guided feature aware network for aircraft detection in SAR Images | |
Zhang et al. | Oriented Ship Detection Based on Soft Thresholding and Context Information in SAR Images of Complex Scenes | |
Kim et al. | Three plot correlation-based small infrared target detection in dense sun-glint environment for infrared search and track | |
Su et al. | Ship detection based on RetinaNet-plus for high-resolution SAR imagery | |
Qosja et al. | Benchmarking Convolutional Neural Network Backbones for Target Classification in SAR | |
Zhang et al. | Infrared target detection using intensity saliency and self-attention |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
WW01 | Invention patent application withdrawn after publication |
Application publication date: 20190611 |
|
WW01 | Invention patent application withdrawn after publication |