CN108470139A - A kind of small sample radar image human action sorting technique based on data enhancing - Google Patents
A kind of small sample radar image human action sorting technique based on data enhancing Download PDFInfo
- Publication number
- CN108470139A CN108470139A CN201810073423.4A CN201810073423A CN108470139A CN 108470139 A CN108470139 A CN 108470139A CN 201810073423 A CN201810073423 A CN 201810073423A CN 108470139 A CN108470139 A CN 108470139A
- Authority
- CN
- China
- Prior art keywords
- data
- image
- radar
- noise
- radar image
- 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
- 238000000034 method Methods 0.000 title claims abstract description 78
- 230000009471 action Effects 0.000 title claims abstract description 34
- 230000002708 enhancing effect Effects 0.000 title claims abstract description 18
- 238000012549 training Methods 0.000 claims abstract description 25
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 16
- 238000013135 deep learning Methods 0.000 claims abstract description 8
- 230000003542 behavioural effect Effects 0.000 claims abstract description 4
- 230000003287 optical effect Effects 0.000 claims abstract description 4
- 230000006835 compression Effects 0.000 claims description 11
- 238000007906 compression Methods 0.000 claims description 11
- 230000009466 transformation Effects 0.000 claims description 9
- 238000001914 filtration Methods 0.000 claims description 6
- 230000002146 bilateral effect Effects 0.000 claims description 4
- 235000002566 Capsicum Nutrition 0.000 claims description 3
- 239000006002 Pepper Substances 0.000 claims description 3
- 235000016761 Piper aduncum Nutrition 0.000 claims description 3
- 235000017804 Piper guineense Nutrition 0.000 claims description 3
- 235000008184 Piper nigrum Nutrition 0.000 claims description 3
- 150000003839 salts Chemical class 0.000 claims description 3
- 238000000638 solvent extraction Methods 0.000 claims description 3
- 238000003672 processing method Methods 0.000 claims description 2
- 244000203593 Piper nigrum Species 0.000 claims 1
- 230000000694 effects Effects 0.000 description 10
- 238000012360 testing method Methods 0.000 description 10
- 238000001514 detection method Methods 0.000 description 7
- 230000006870 function Effects 0.000 description 7
- 230000008569 process Effects 0.000 description 7
- 230000004044 response Effects 0.000 description 7
- 230000008859 change Effects 0.000 description 6
- 230000006399 behavior Effects 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- VJYFKVYYMZPMAB-UHFFFAOYSA-N ethoprophos Chemical compound CCCSP(=O)(OCC)SCCC VJYFKVYYMZPMAB-UHFFFAOYSA-N 0.000 description 4
- 238000013507 mapping Methods 0.000 description 4
- 239000011159 matrix material Substances 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 238000013461 design Methods 0.000 description 3
- 238000000605 extraction Methods 0.000 description 3
- 238000005286 illumination Methods 0.000 description 3
- 230000003993 interaction Effects 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 230000011218 segmentation Effects 0.000 description 3
- 241000722363 Piper Species 0.000 description 2
- 230000004913 activation Effects 0.000 description 2
- 238000013480 data collection Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000003709 image segmentation Methods 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- HPTJABJPZMULFH-UHFFFAOYSA-N 12-[(Cyclohexylcarbamoyl)amino]dodecanoic acid Chemical compound OC(=O)CCCCCCCCCCCNC(=O)NC1CCCCC1 HPTJABJPZMULFH-UHFFFAOYSA-N 0.000 description 1
- 241001269238 Data Species 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000000903 blocking effect Effects 0.000 description 1
- 210000000988 bone and bone Anatomy 0.000 description 1
- QBWCMBCROVPCKQ-UHFFFAOYSA-N chlorous acid Chemical compound OCl=O QBWCMBCROVPCKQ-UHFFFAOYSA-N 0.000 description 1
- 230000001149 cognitive effect Effects 0.000 description 1
- 238000010835 comparative analysis Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000005520 cutting process Methods 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 239000003607 modifier Substances 0.000 description 1
- 230000008520 organization Effects 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 239000000047 product Substances 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 239000013589 supplement Substances 0.000 description 1
- 230000000153 supplemental effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/20—Movements or behaviour, e.g. gesture recognition
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/70—Denoising; Smoothing
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Psychiatry (AREA)
- Social Psychology (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Processing (AREA)
Abstract
The present invention relates to a kind of small sample radar image human action sorting techniques based on data enhancing, including:Using emulation radar image as the source of training data, the human body behavioral data acquired using optical motion catcher, radar return is calculated by establishing human body ellipsoid motion model, radar spectrogram is being obtained by Short Time Fourier Transform, is generating data set;Data enhancing is realized to the radar image in data set using a variety of data enhancement methods;Convolutional neural networks model is established, and carries out the training of deep learning using CAFFE.
Description
Technical field
The invention belongs to classification of radar targets, data enhancing and deep learning field, it is related to increasing the data of radar image
It is strong go forward side by side pedestrian's body classification of motion the problem of.
Background technology
The correlative study of Activity recognition [1] analysis can trace back to one of [2] experiment of Johansson in 1975, author
12 manikins are proposed, the point model method of this description behavior is to the behavior description algorithm later based on organization of human body
Play important directive function.In intelligent video monitoring, patient monitoring system, human-computer interaction, virtual reality, smart home,
Intelligent security guard, sportsman's supplemental training and military field are suffered from and are widely applied.Human bodys' response is in action at present
Cognitive phase, and the process that action recognition can regard feature extraction as and classifier design is combined.Characteristic extraction procedure by
To blocking, dynamic background, dollying head, the influence of the factors such as visual angle and illumination variation and there is prodigious challenge.At present
The main Research Challenges of Human bodys' response include acting class in class to change greatly, space complexity, time difference opposite sex [3] etc..
For mostly acting, even same action has the different forms of expression.Space complexity refer to different illumination,
It will present different action scenes under the conditions of visual angle and background etc., and identical human body behavior exists in different action scenes
Difference is will produce in posture and characteristic, even if larger degree of freedom is had if human action in constant action scene, and
And each identical action has prodigious otherness in terms of direction, angle, shape and size.In addition, human body blocks certainly, portion
Point block, human body individual difference, more person recognition objects the problems such as be all the embodiment of action recognition complexity spatially.When
Between otherness refer to human action occur time point it is unpredictable, and act duration intervals be also not quite similar.In addition,
Action is also likely to be present action white space within action time.The time difference opposite sex requires to distinguish action in identification process
Beginning and ending time, while the effectively effective time of judgement action effect and interval is carried out to acting within the scope of time domain and sequential
More careful analysis causes action that can all be had differences under different rates, sequence and combined situation.
Doppler radar identifying system is it is possible to prevente effectively from weather, illumination, all kinds of influence factors such as block.Doppler radar
It is a kind of Active Radar, according to doppler principle, electromagnetic wave is emitted to moving target and target is differentiated by its echo-signal
Motion state.Radar detection suffers from important application in many fields, such as unmanned, all various aspects such as post-disaster search and rescue.Base
In Doppler radar Human bodys' response technology be in recent years since the new technology that grows up, the radar image after ovennodulation
It contains partes corporis humani and divides the fine motion Doppler frequency that modulation generates, and then human motion can be differentiated, this makes base
It is possibly realized in the human action identification of Doppler radar.
The data enhancing technology of Human bodys' response based on Doppler radar has higher researching value, data enhancing
Method is widely used in field of image recognition, and relative skill is more mature, however but yet there are no deep answer for radar image
With and research.
Invention content
The object of the present invention is to provide a kind of small sample radar image human action sorting techniques.The present invention is in radar data
Measure it is insufficient under the premise of, with reference to natural image data enhancement method to small sample radar data realize data enhancing, recycle
Convolutional neural networks realization in deep learning classifies to human action in the enhanced radar image of data, and classification results are good
Data set before enhancing.Technical solution is as follows:
A kind of small sample radar image human action sorting technique based on data enhancing, includes the following steps:
1) data set is built:Using emulation radar image as the source of training data, using optical motion catcher
The human body behavioral data of acquisition chooses the action of 7 classes, is to run, jump, walk, climb, stand, box and crawl respectively, by establishing human body
Radar return is calculated in ellipsoid motion model, is obtaining radar spectrogram by Short Time Fourier Transform, is generating data set;
2) data enhancing is realized to the radar image in data set using a variety of data enhancement methods:Realize data enhancing
Method includes bicubic transformation in compression of images enhancement method, bilinear transformation, Method of Partitioning, approximate point processing method;Filtering
Four kinds of mean filter method, bilateral fuzzy special efficacy filter method, Gaussian Blur filter method and median filtering method sides in enhancement method
Method;The Gaussian noise in Noise enhancement mode, the white Gaussian noise that mean value related with gradation of image is zero, pepper is added to make an uproar
Sound, salt noise, poisson noise, s&p noises, speckle noise method;Expose the method for changing pixel value gamma in enhancement method;
3) establishes convolutional neural networks model, and the training of deep learning is carried out using CAFFE, is importing training network
Before, the size of each spectrogram is adjusted to 100 × 100, and the size of convolution kernel is 9 × 9, and stride is 1 pixel.
Data enhancement method of the present invention according to natural image proposes a kind of suitable for improving radar image recognition effect
Data enhancement method improves the accuracy rate of Human bodys' response.How general data set is with the radar based on MOCAP data set generations
It is research object to strangle image and pass through the enhanced image of data, includes enhancing and making, the convolutional neural networks mould of data set
Foundation, training and the test of type.The advantages of this patent is according to radar system is carried for the insufficient objective condition of radar data collection
A kind of suitable data enhancement methods are gone out and have carried out EDS extended data set, and then the identifiability of radar data collection is made to be improved.It should
Invention can further increase the accuracy rate of Human bodys' response on the basis of legacy data, be required more for recognition accuracy
High application provides help.
Description of the drawings
Fig. 1 is high-resolution partes corporis humani position radar image;
Fig. 2 is human body ellipsoidal model figure;
Fig. 3 function relation figures between gamma and accuracy.
Specific implementation mode
The present invention is summarized first:
1. building data set.Model training, which is carried out, by convolutional neural networks needs mass data.Due to no any public affairs
Radar measuring image data set altogether, the present invention uses source of the emulation radar image as training data, using Ka Neijimeilong
The human body behavioral data of university's optical motion catcher acquisition generates data set.The model is without directly optimizing grid and bone
Bone parameter, but radar return is calculated by establishing human body ellipsoid motion model, it is obtained by Short Time Fourier Transform
To radar spectrogram.
2. radar image data enhances.The present invention uses a variety of data enhancement methods, in compression of images enhancement method
The methods of bicubic transformation, bilinear transformation, Method of Partitioning, approximate point processing;Filter mean filter method, bilateral in enhancement method
Four kinds of fuzzy special efficacy filter method, Gaussian Blur filter method and median filtering method methods;Gauss in addition Noise enhancement mode makes an uproar
White Gaussian noise that sound, mean value related with gradation of image are zero, pepper noise, salt noise, poisson noise, s&p noises, spot
The methods of spot noise;Expose the method for changing pixel value gamma in enhancement method.
3. establishing convolutional neural networks model.The present invention uses convolutional neural networks (Convolutional Neural
Networks, CNN), 3 convolutional layers and 1 full articulamentum are contained in network structure, and each convolutional layer is followed by one
Down-sampling layer.The present invention uses CAFFE (Convolutional Architecture for Fast Feature
Embedding) training of deep learning is carried out.Before importing training network, the size of each spectrogram is adjusted to
100×100.The size of convolution kernel is 9 × 9, and stride is 1 pixel.
4. training convolutional neural networks model.Prepare spectrogram appropriate by data enhancement method to be trained.In number
The action of 7 classes is chosen according to the present invention on the structure of collection, is to run, jump, walk, climb, stand, box and crawl respectively.Each action is with 100
Based on Zhang Shengcheng images, the image that 100 various data enhancement methods generate is added respectively, in deep learning frame CAFFE
Lower training pattern, and on test set verify model validity.
5. testing above-mentioned model on test set, the effect according to the classification accuracy comparative analysis data enhancement methods.
To keep technical scheme of the present invention clearer, the specific embodiment of the invention is further described through.This hair
It is bright to implement according to the following steps:
1. radar time-frequency image data set is built
(1) radar image based on MOCAP data sets generates
This data set is established by Carnegie Mellon University graph experiment room.This data set contains six kinds of movement feelings
Total 2605 tests movement including scape.This six kinds of movement scenarios are between interpersonal interaction, man and nature respectively
Interaction, sports, autogenic movement, scene change and test activity in movement.MOCAP data set uses are by Vicon
The motion capture system of Industries exploitations collects mankind's activity data.This system represents human body with 41 mark points
The movement of partes corporis humani point can be reduced to movement a little by various pieces when collecting data.The motion capture system by
12 infrared ray MX40 video cameras compositions, each video camera can be with 120Hz frame speed recording images.High-resolution human body is each
Position radar image is as shown in Figure 1.The data set includes 2605 groups of experimental datas, select in process of the present invention wherein seven kinds often
The action seen is used for generating radar image, this seven kinds actions are respectively:It runs, walk, jump, creep, creep, stand
And boxing.MOCAP data sets construct a human body spheroid action model, and human body ellipsoidal model is as shown in Fig. 2, entire people
Body Model is made of multiple ellipsoids, each spheroid radar reflection wave-amplitude can by be approximately ellipse RCS obtain,
It is handled using Short Time Fourier Transform and obtains radar spectrogram from the reflected radar return of human body various pieces, the present invention
The characteristics of according to radar spectrogram, data set is obtained by " sliding window method ", can get size for each being acted in classification task
For the data set of 200 pictures, the data set of each action is divided into two parts by the present invention, respectively 100 training sets and
100 test sets.
(2) radar image data enhances
Data enhance using image procossing as theoretical foundation, contain luminance transformation, space filtering, compression of images, image point
It cuts and a variety of methods such as image restoration.
Luminance transformation is built upon with space filtering on the basis to processes pixel, and spatial domain handles expression formula
G (x, y)=T [f (x, y)] (1)
It indicates, wherein f (x, y) is inputted as image, and g (x, y) is the image after output, and T is the behaviour handled image f
It accords with.In calculating process, the point on each image of definition is moved along image progressive, is calculated every output valve and is only needed to use
To the neighborhood of a point.Brightness change needs that exposure modules, the last parameter gamma of function is called to specify pixel value
Size, newly-generated image is darker than original image when gamma values are more than 1, otherwise brightness improves.In addition in handling image, pixel
Codomain just can be achieved on by bearing, but practical want to save or check pictures when are to the processing of negative value non-
Often difficult, therefore the present invention is needed image scale in the scale that maximum magnitude is [0,255].By brightness change extraction
Histogram information plays the role of basic in compression of images, image segmentation.
The compact model of image can simply be divided into encoder and decoder two parts.Encoder can be divided into mapping variation
Device, quantizer, symbol encoder three parts, decoder can be divided into symbol decoder, anti-mapping modifiers two parts.When defeated
When entering image f (x, y) feeding encoders, encoder can establish one group of coded sequence to describe image, by right according to image
Variation than compressing front and back image information bit determines compression ratio, then is quantified to compression image with compression ratio.In order to make
It with compression image, needs image being re-fed into decoder, to generate the image of a reconstruct.In general reconstruct image
Precise Representation as being likely to be input picture, if it is, the system can regard an error free, letter as
Breath preserves complete system.If it is not, illustrate that there are errors between reconstructed image and input picture, therefore the present invention needs
One error function is set to define the size cases of error.I.e.
Encoder section is responsible for reducing the redundancy on input picture coding, pixel and psycho-visual.In the mapping of encoder
The variator stage changes image into a kind of invisible format and is reduced according to compression ratio for reducing redundancy between pixel, quantizer
The accuracy of output is to eliminate the redundancy on psycho-visual, finally by symbol encoder to changing from quantizer and mapping
The code word exported in device is recombinated, and the elimination to coding redundancy is completed.
Image segmentation is also a kind of basic mode of data enhancing, and cutting operation can say that image is subdivided into heterogeneity,
It segments the needs that degree needs foundation practical problem.Segmentation fine degree is an important embodiment of computer recognition capability.
It is generally basede on brightness for the segmentation of monochrome image to be split, figure is differentiated according to the continuity of luminance information and similitude
As edge.In most basic point detection, it usually needs define the response R for being a little in filter of image by the point
Gray level is obtained with the sum of products of corresponding design factor, and formula is as follows:
W indicates that design factor, z indicate and the relevant pixel intensities of w.When the value of R is more than preset threshold value of the invention, then
Illustrate that the point is an opposite isolated point.The detection of line is relatively more more complex, and each filter model can be melted into one
A 3*3 matrixes, each optimum point of matrix by 2 weightings, can form horizontal, vertical of the optimum point in this matrix and
45 degree of inclined lines.Therefore the detection of line and threshold value set closely related, ordinary circumstance, and the present invention is right in the detection to line
It is interested in directive line, thus can with the threshold value of the line of the independent analysis direction, so again by after threshold process just
It can obtain one group and respond most strong point in the direction, eventually detect line.The detection of point and the detection of line are for image edge
Divide important role, but it is to detect the continuity of brightness that a kind of most important means are divided at edge so far.This company
Continue property judgement to need to be judged by single order and second dervative, gradient can be defined as
It is zero in the middle Grad of constant brightness, Grad is directly proportional to brightness change value.The most basic spy of gradient vector
Point is exactly the direction that can be directed toward f (x, y) maximum rate of change.Noise factor has second dervative extremely strong influence, therefore needs
It introduces Laplace operator to supplement it, obtains the edge segmentation that gradient can be used for image in this way.
Compression of images is also a kind of important way of data enhancing, such as BICUBIC, NEAREST, LANCZOS.
BICUBIC (bicubic interpolation) is a kind of interpolation method of complexity, it can create the image side more smoother than bilinear interpolation
Edge.Bicubic interpolation method is usually used in a part of image processing software, printed driver and digital camera, right
The some regions of original image or original image are amplified.At present in commercial image editing software, often using speed
It is most fast, but be also least accurate " nearest neighbor " (Nearest) interpolation.LANCZOS algorithms are a kind of to lead to symmetrical matrix
The algorithm that orthogonal similarity transformation becomes symmetric triple-diagonal matrix is crossed, the power of each pixel of image can be calculated by the algorithm
Weight, further according to weight selected pixels value.
2. the human action disaggregated model structure based on convolutional neural networks
(1) basic convolutional neural networks model construction
3 convolutional layers and 1 full articulamentum are contained in the CNN network structures that the present invention uses, behind each convolutional layer
With a down-sampling layer.The present invention carries out the training of deep learning using CAFFE.Before importing training network, Mei Gepin
The size of spectrogram is adjusted to 100 × 100.The size of convolution kernels is 9 × 9, and stride is 1 pixel.Image is imported from convolutional layer
20 characteristic patterns are generated, and are handled successively with ReLU activation primitives and 2 × 2 maximum pond layers (max pooling).So
The characteristic pattern generated from last convolutional layer is sent to full articulamentum afterwards, then is handled with Softmax activation primitives.
The convolutional neural networks model training 3. radar human action is classified
1 parameter setting of table
Highest measuring accuracy can be obtained by the hyper parameter experience optimization of CNN.In addition, NVIDIA Titan X GPU
Training process is also accelerated with the libraries CUDA (CUDNN).The method of stochastic gradient descent can be used for the power in adjusting training network
Weight.Parameter setting is as shown in table 1, and basic learning rate is using the coefficient before gradient when SGD algorithms, and setting is too small to be caused
Optimization algorithm is too slow, and setting is excessive can not optimize, and the present invention is set to 0.001, and momentum is empirical value, generally
Between 0.9 to 0.95, weight attenuation rate and threshold value are default values, need not excessively be adjusted.A training can be obtained in this way
Good CNN graders.
4. the classifying quality of model is tested
Table 2 has the data enhancement methods for promoting effect
3 ineffective data enhancement methods of table
The present invention using 100 generate figure as training dataset 100 test sets recognition accuracy as benchmark, divide
The picture for increasing the generation of 100 different data enhancement methods on the basis of 100 artworks is not compared as training set and identical
Accuracy rate under test set.
The present invention enumerates all data enhancement methods that can increase classification accuracy in table 2.Table 3 is listed to radar
The unfruitful data enhancement methods of image classification, therefore not all data enhancement method can improve radar image
Recognition accuracy.In the method enhanced as data using no method for compressing image, bicubic interpolation (BICUBIC),
Nearest neighbor interpolation (NEAREST) and Lanczos methods all have promotion effect to radar image recognition accuracy.In addition it can send out
The data enhancement methods for now increasing noise have accuracy rate the effect generally promoted, this may be because Noise Method data enhance
Method increases the robustness of data set.In addition Bilateral Filter also have obviously the promotion of accuracy rate during image is fuzzy
Effect.During brightness is reconciled, there is very big promotion in when image dimmed (gamma 1.5) to accuracy rate, improves nearly 6 percentage points, and
When image is excessively bright, accuracy rate declines extremely serious, has dropped 4 percentage points.Fig. 3 shows gamma variations and accuracy
Between variation relation.The classification accuracy of radar image is in certain brightness range and image light and shade as seen in Figure 3
It spends related.
By image it is observed that test accuracy rate increases as gamma values increase, this explanation is by reducing radar
Picture contrast can effectively enhance recognition effect.Gamma values are skimage.exposure.adjust_gamma
(image, gamma=1).From its bottom document it may be seen that the processing formula to pixel brightness value is:
I=Ig (5)
The brightness value I of pixel is fixed between (0~1), therefore gamma values are bigger, and treated, and pixel brightness value is got over
It is low.Contrast refer in an image brightness it is maximum place and the place of brightness minimum between difference, difference range more it is big then
Contrast is bigger, and the smaller then contrast of difference is smaller.It is exponential type function that Exposure, which handles function, when gamma values are more than 1
When, pixel brightness value generally reduces in treated image, but due to the non-linear nature of exponential function, pixel brightness value is high
Place lacking of reducing, the low place drop of pixel brightness value is much lower, thus while image overall brightness is dimmed, but image
Contrast but gets a promotion, and the higher place of brightness is radar echo signal in radar image.Therefore it is equivalent to radar simulation
The background signal of image weakens, and enhances radar echo signal, and then enhances the accuracy of radar image identification.
Claims (1)
1. a kind of small sample radar image human action sorting technique based on data enhancing, includes the following steps:
1) data set is built:Using emulation radar image as the source of training data, acquired using optical motion catcher
Human body behavioral data, choose 7 classes action, be runs, jump, walk, climb, stand, box and crawl respectively, by establish human body ellipsoid transport
Radar return is calculated in movable model, is obtaining radar spectrogram by Short Time Fourier Transform, is generating data set;
2) data enhancing is realized to the radar image in data set using a variety of data enhancement methods:Including compression of images enhancing side
Bicubic transformation, bilinear transformation, Method of Partitioning, approximate point processing method in formula;Filter enhancement method in mean filter method,
Four kinds of bilateral fuzzy special efficacy filter method, Gaussian Blur filter method and median filtering method methods;Add the height in Noise enhancement mode
White Gaussian noise that this noise, mean value related with gradation of image are zero, pepper noise, salt noise, poisson noise, s&p noises,
Speckle noise method;Expose the method for changing pixel value gamma in enhancement method;
3) establishes convolutional neural networks model, and the training of deep learning is carried out using CAFFE, before importing training network,
The size of each spectrogram is adjusted to 100 × 100, and the size of convolution kernel is 9 × 9, and stride is 1 pixel.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810073423.4A CN108470139A (en) | 2018-01-25 | 2018-01-25 | A kind of small sample radar image human action sorting technique based on data enhancing |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810073423.4A CN108470139A (en) | 2018-01-25 | 2018-01-25 | A kind of small sample radar image human action sorting technique based on data enhancing |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108470139A true CN108470139A (en) | 2018-08-31 |
Family
ID=63266150
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810073423.4A Pending CN108470139A (en) | 2018-01-25 | 2018-01-25 | A kind of small sample radar image human action sorting technique based on data enhancing |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108470139A (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110400252A (en) * | 2019-06-28 | 2019-11-01 | 董立 | Stock ground contour digitazation method and system |
CN110796206A (en) * | 2019-11-06 | 2020-02-14 | 国网山东省电力公司电力科学研究院 | Data enhancement method and device for partial discharge map |
CN111507361A (en) * | 2019-01-30 | 2020-08-07 | 富士通株式会社 | Microwave radar-based action recognition device, method and system |
CN111898652A (en) * | 2020-07-10 | 2020-11-06 | 西北工业大学 | Spatial target posture classification and identification method based on convolutional neural network |
CN111968048A (en) * | 2020-07-30 | 2020-11-20 | 国网智能科技股份有限公司 | Method and system for enhancing image data of few samples in power inspection |
CN112116140A (en) * | 2020-09-10 | 2020-12-22 | 同济大学 | Building energy consumption prediction method based on twin model |
WO2021129569A1 (en) * | 2019-12-25 | 2021-07-01 | 神思电子技术股份有限公司 | Human action recognition method |
CN113420448A (en) * | 2021-06-25 | 2021-09-21 | 中国兵器装备集团自动化研究所有限公司 | Digital twinning system and method for ammunition fusion casting charging forming process |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105759265A (en) * | 2016-01-28 | 2016-07-13 | 国家海洋局第二海洋研究所 | Synthetic aperture radar (SAR) image moving target parameter extraction method |
CN106680794A (en) * | 2015-11-10 | 2017-05-17 | 核工业北京地质研究院 | Rapid radar data de-noising method based on spatial modeling technology |
CN107169435A (en) * | 2017-05-10 | 2017-09-15 | 天津大学 | A kind of convolutional neural networks human action sorting technique based on radar simulation image |
-
2018
- 2018-01-25 CN CN201810073423.4A patent/CN108470139A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106680794A (en) * | 2015-11-10 | 2017-05-17 | 核工业北京地质研究院 | Rapid radar data de-noising method based on spatial modeling technology |
CN105759265A (en) * | 2016-01-28 | 2016-07-13 | 国家海洋局第二海洋研究所 | Synthetic aperture radar (SAR) image moving target parameter extraction method |
CN107169435A (en) * | 2017-05-10 | 2017-09-15 | 天津大学 | A kind of convolutional neural networks human action sorting technique based on radar simulation image |
Non-Patent Citations (2)
Title |
---|
王思雨等: "基于卷积神经网络的高分辨率SAR图像飞机目标检测方法", 《雷达学报》 * |
颜志国等: "《多摄像机协同关注目标检测跟踪技术》", 30 June 2017 * |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111507361A (en) * | 2019-01-30 | 2020-08-07 | 富士通株式会社 | Microwave radar-based action recognition device, method and system |
CN111507361B (en) * | 2019-01-30 | 2023-11-21 | 富士通株式会社 | Action recognition device, method and system based on microwave radar |
CN110400252A (en) * | 2019-06-28 | 2019-11-01 | 董立 | Stock ground contour digitazation method and system |
CN110400252B (en) * | 2019-06-28 | 2022-09-06 | 中科航宇(北京)自动化工程技术有限公司 | Material yard contour line digitalization method and system |
CN110796206A (en) * | 2019-11-06 | 2020-02-14 | 国网山东省电力公司电力科学研究院 | Data enhancement method and device for partial discharge map |
WO2021129569A1 (en) * | 2019-12-25 | 2021-07-01 | 神思电子技术股份有限公司 | Human action recognition method |
CN111898652A (en) * | 2020-07-10 | 2020-11-06 | 西北工业大学 | Spatial target posture classification and identification method based on convolutional neural network |
CN111968048A (en) * | 2020-07-30 | 2020-11-20 | 国网智能科技股份有限公司 | Method and system for enhancing image data of few samples in power inspection |
CN111968048B (en) * | 2020-07-30 | 2024-03-26 | 国网智能科技股份有限公司 | Method and system for enhancing image data of less power inspection samples |
CN112116140A (en) * | 2020-09-10 | 2020-12-22 | 同济大学 | Building energy consumption prediction method based on twin model |
CN113420448A (en) * | 2021-06-25 | 2021-09-21 | 中国兵器装备集团自动化研究所有限公司 | Digital twinning system and method for ammunition fusion casting charging forming process |
CN113420448B (en) * | 2021-06-25 | 2023-05-23 | 中国兵器装备集团自动化研究所有限公司 | Digital twin system and method for ammunition fusion casting charging forming process |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108470139A (en) | A kind of small sample radar image human action sorting technique based on data enhancing | |
Huang et al. | Faster R-CNN for marine organisms detection and recognition using data augmentation | |
Lore et al. | LLNet: A deep autoencoder approach to natural low-light image enhancement | |
CN108764085B (en) | Crowd counting method based on generation of confrontation network | |
CN107169435A (en) | A kind of convolutional neural networks human action sorting technique based on radar simulation image | |
EP3211596A1 (en) | Generating a virtual world to assess real-world video analysis performance | |
US20230214976A1 (en) | Image fusion method and apparatus and training method and apparatus for image fusion model | |
KR102103770B1 (en) | Apparatus and method for pedestrian detection | |
US20240062530A1 (en) | Deep perceptual image enhancement | |
US20230281913A1 (en) | Radiance Fields for Three-Dimensional Reconstruction and Novel View Synthesis in Large-Scale Environments | |
Liu et al. | Normalized face image generation with perceptron generative adversarial networks | |
CN106951870A (en) | The notable event intelligent detecting prewarning method of monitor video that active vision notes | |
CN114445715A (en) | Crop disease identification method based on convolutional neural network | |
CN110390673A (en) | Cigarette automatic testing method based on deep learning under a kind of monitoring scene | |
CN108256567A (en) | A kind of target identification method and system based on deep learning | |
Liu et al. | Image edge recognition of virtual reality scene based on multi-operator dynamic weight detection | |
US11138812B1 (en) | Image processing for updating a model of an environment | |
Nentwig et al. | Concerning the applicability of computer graphics for the evaluation of image processing algorithms | |
CN114067172A (en) | Simulation image generation method, simulation image generation device and electronic equipment | |
CN114708615A (en) | Human body detection method based on image enhancement in low-illumination environment, electronic equipment and storage medium | |
CN116391209A (en) | Realistic audio-driven 3D avatar generation | |
CN112927127A (en) | Video privacy data fuzzification method running on edge device | |
CN111862278A (en) | Animation obtaining method and device, electronic equipment and storage medium | |
CN116740808A (en) | Animal behavior recognition method based on deep learning target detection and image classification | |
Tang et al. | Learning from natural noise to denoise micro-doppler spectrogram |
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 |
Application publication date: 20180831 |
|
RJ01 | Rejection of invention patent application after publication |