CN109002848A - A kind of detection method of small target based on Feature Mapping neural network - Google Patents
A kind of detection method of small target based on Feature Mapping neural network Download PDFInfo
- Publication number
- CN109002848A CN109002848A CN201810729648.0A CN201810729648A CN109002848A CN 109002848 A CN109002848 A CN 109002848A CN 201810729648 A CN201810729648 A CN 201810729648A CN 109002848 A CN109002848 A CN 109002848A
- Authority
- CN
- China
- Prior art keywords
- neural network
- indicate
- target
- detection
- spindle
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—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
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace 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
- G06V10/30—Noise filtering
-
- 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
- G06V2201/07—Target detection
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of detection method of small target based on Feature Mapping neural network, are related to Dim targets detection field；It includes step 1: building, training spindle-type deep neural network；Step 2: by the amplitude figure that the spindle-type deep neural network that the Weak target image input of acquisition has been trained obtains targets improvement and background inhibits；Step 3: amplitude figure completes Dim targets detection using constant false alarm rate method.The present invention uses spindle network structure, the powerful expression ability of network is improved, solves the problems, such as that existing Weak target is caused detection accuracy low by noise and interference effect, has reached and has improved the powerful expression ability of network, under high-noise environment, the effect of Dim targets detection precision is improved.
Description
Technical field
The present invention relates to Dim targets detection field, especially a kind of Weak target inspection based on Feature Mapping neural network
Survey method.
Background technique
Passive millimeter wave PMMW and infrared imaging have the good characteristic radiationless, penetration capacity is strong, in military field
In using of increasing concern, therefore under millimeter wave and infrared imaging to Dim targets detection carry out research have it is very heavy
The meaning wanted.Detection of Small and dim targets is developed rapidly in recent years, and Weak target refers to that diameter is the mesh of 3-5 pixel
Mark, but great difficulty is still faced for millimeter wave and the detection of infrared imaging condition small target with high precision under: firstly, target
Image-forming range generally farther out, detected target area is smaller, and noise is relatively low, and texture-free feature is extractable；Second, mesh
Mark imaging is usually by the interference of complex background, and a large amount of clutter, noise, there are also some marginal informations such as: cloud edge, extra large space-based
The presence at line, building edge etc., causes target and is submerged among background.
For Dim targets detection, academia proposes a series of detection methods in recent years；Background suppression method is small and weak
The most common method in target detection, this method carry out target detection by the background of estimation image to be detected on this basis.
It is broadly divided into three classes detection method: the first kind is the method based on filtering, and background is estimated by image filtering, finally makes mesh
Mark is enhanced；It inhibits the effect of background preferable in the better simply situation of background, background is more complex, signal-to-noise ratio is lower
Situation, false-alarm probability increase, detection accuracy decline；Second class is the method based on recurrence, and homing method can be divided into linearly again
It returns and nonlinear regression, classical linear regression method depends on specific background clutter model and seek the ginseng of hypothesized model
Number estimation；And non-linear regression method only relies upon data itself to estimate regression function；Kernel regression algorithm NRRKR is typical
Nonlinear regression algo；In practical applications, due to the priori knowledge of shortage background clutter, non-linear regression method is more suitable for multiple
The detection of Weak target under miscellaneous background condition, but its there are clearly disadvantageous: each regional area requires repeatedly to be returned
Return iteration, total algorithm efficiency is extremely low；Third class method is inhibited to background according to local contrast difference, to target into
Row enhancing, completes the detection to target.Such methods detection effect in the better simply situation of background is preferable, and in complex background
It is lower to be easy to increase false-alarm targets quantity and easily affected by noise.
Other than background suppression method, there are also a kind of detection method based on machine learning, such method pattern classifications
Thought go solve target detection problems, it is trained modeling to target and background respectively, is then determined according to decision rule
Whether the subimage block of test image contains target, such as NLPCA, SPCA, FLD.Later, going out with sparse representation theory
It is existing, to solve the problems, such as that Dim targets detection brings new method.The small IR targets detection indicated based on image sparse is calculated
Method SR, this method generate target dictionary using binary Gauss model, then by background sub-block and target sub-block in target dictionary
The difference of middle sparse coefficient judges the position of target.Gauss dictionary is only applicable to height as the excessively complete dictionary of typical structuring
The Weak target of this distribution, and for the target of unstructuredness, rarefaction representation coefficient is not enough to distinguish target and background clutter.
Later, Wang et al. proposed the sparse dictionary of multi-scale self-adaptive form to detect Weak target, by using different size of
Atom describes the heterogeneity of image, and the more subtle local feature of capture image improves detection accuracy；Then, old et al.
The method based on degree of rarefication is proposed, it is dilute to improve that this method constructs different doubledictionary manually in a manner of off-line learning
Dredge the difference indicated；Later, it is thus proposed that new method: be based on the sparse reconstruct weak moving target detection algorithm of space-time joint
STCSR, the content that this method passes through study sequence image first construct adaptive kenel it is excessively complete empty when dictionary, then using more
Dictionary and dictionary when background sky when first Gauss model extracts target empty from excessively complete dictionary, by multiple image respectively in target
Dictionary and dictionary carries out sparse reconstruct when background sky when empty, distinguishes target and background using reconstruct difference.Exist for this method
Deficiency in terms of dictionary learning, it is thus proposed that improved ISR method, this method propose a kind of significant background and target doubledictionary
Building method has better target and background modeling ability.Above method all improves detection accuracy to a certain extent, but
The above method has the disadvantage that the interference being on the one hand easy by noise, the difference of another aspect target and background sparse features
It is unobvious, it is easy mixed in together, increases detection difficulty.
Therefore detection is improved while needing a kind of detection method of small target that can overcome the noise factor of Weak target
Precision.
Summary of the invention
It is an object of the invention to: the present invention provides a kind of Dim targets detection sides based on Feature Mapping neural network
Method solves the problems, such as that existing Weak target is caused detection accuracy low by noise and interference effect.
The technical solution adopted by the invention is as follows:
A kind of detection method of small target based on Feature Mapping neural network, includes the following steps:
Step 1: building, training spindle-type deep neural network；
Step 2: the spindle-type deep neural network that the Weak target image input of acquisition has been trained is obtained into targets improvement
The amplitude figure inhibited with background；
Step 3: amplitude figure completes Dim targets detection using constant false alarm rate method.
Preferably, the step 1 includes the following steps:
Step 1.1: building include input layer, decoding layer, coding layer and softmax output layer spindle-type depth nerve net
The structure of network；
Step 1.2: determining that the hyper parameter of spindle-type deep neural network obtains spindle-type depth using cross validation method
Neural network；
Step 1.3: building training dataset；
Step 1.4: training dataset input spindle-type deep neural network is trained acquisition using unsupervised mode
It initializes network weight and completes training.
Preferably, the decoding layer, coding layer training calculating are as follows:
hk=σ (WkX+bk)
Wherein, WkIndicate weight matrix, bkIndicate that bias vector, σ indicate activation primitive, X={ x1,x2,...,xmIndicate
The input of current layer, hkIndicate the output of current layer.
Preferably, the step 2 includes the following steps:
Step 2.1: after the spindle-type deep neural network that the Weak target image input of acquisition has been trained, by small and weak mesh
Mark sample label is set as 1, and background sample label is set as 0；
Step 2.2: differentiate Weak target in acquisition image to Weak target image using the method for sliding window
Probability value；
Step 2.3: width is responded as window coordinates point using result, that is, multiple probability values that output layer logistic is returned
The amplitude figure that value obtains the corresponding targets improvement of multiple Weak targets and background inhibits.
Preferably, the step 3 includes the following steps:
Step 3.1: each value in amplitude figure extracts sub-block using sliding windowIt is detected, inputs spindle-type
Deep neural network obtains amplitude；
Step 3.2: statistics being carried out to each amplitude after detection using constant false alarm rate and obtains false-alarm probability；
Wherein, T indicates the threshold value of Likelihood ration test,Indicate the mean value of sliding window sub-block, p indicates point in window sub-block
Number, PfaIndicate the false-alarm probability of constant false alarm rate detection setting, τCFARIndicate detection threshold value, F1,p-1(τCFAR) indicate center F
The cumulative distribution function of stochastic variable；
Step 3.3: candidate target sum being calculated according to false-alarm probability, false-alarm probability is sorted from high to low, from multiple width
Detection positioning target in value.
Preferably, the building training dataset in the step 1.3 includes the following steps:
Step 1.3.1: coordinate points are randomly generated in not including Weak target image as simulation objectives, extract N*N window
Mouth region domain is as background sample；
Step 1.3.2: a simulation objectives are added as target sample using dimensional Gaussian strength model in background sample
This, dimensional Gaussian model is as follows:
Wherein, (x0,y0) indicate target image center, s (i, j) indicate target image position (i, j) pixel
Value, sEIndicate generate target intensity, value be (0,1] between random number, σxAnd σyRespectively indicate horizontal and vertical distribution ginseng
Number, value is between [0,2]；
Step 1.3.3: the different parameters for adjusting target sample generate the Weak target completion training data of different signal-to-noise ratio
The building of collection.
In conclusion by adopting the above-described technical solution, the beneficial effects of the present invention are:
1. the present invention uses spindle network structure, the Weak target Block Characteristic of low-dimensional is mapped to higher-dimension sky first
Between, high resoluting characteristic is then extracted by encoding nerve network, completes background and target-recognition, the strong of output is differentiated according to network
Degree obtains the image that background inhibits targets improvement, finally completes to examine Weak target using the detection method based on constant false alarm rate
It surveys, solves the problems, such as that existing Weak target is caused detection accuracy low by noise and interference effect, it is powerful to have reached raising network
Expression ability, under high-noise environment, improve Dim targets detection precision effect；
2. the millimeter wave and infrared image of the invention under different scenes detects, decoding behaviour is realized in the front end of network
Make, so that network has more powerful expression ability, first by Weak target area pixel Feature Mapping to high-dimensional feature space,
High dimensional feature is passed through into coding mapping to the low-dimensional feature space for being easy to differentiate again, reaches lower false alarm rate, higher detection
The effect of precision, stronger robustness；
3. the unsupervised mode of learning pre-training that inventive network uses, building deeper time structure, by unsupervised
Network stabilization can be improved in the initialization network weight for learning to obtain, during directly training deep neural network
Local Minimum problem, while unsupervised learning obtains a series of internal feature of associated data sets, removes the redundancy of input data
Ingredient can differentiate feature conducive to height is obtained, further increase the discrimination precision of network；
4. being detected in constant false alarm rate method of the invention using sliding block, two target ranges are assumed according to the actual situation
Have and permanent false rate detection is carried out using sliding block when a certain distance, on the one hand can accelerate to detect speed, on the other hand be conducive to mention
High measurement accuracy.
Detailed description of the invention
In order to illustrate the technical solution of the embodiments of the present invention more clearly, below will be to needed in the embodiment attached
Figure is briefly described, it should be understood that the following drawings illustrates only certain embodiments of the present invention, therefore is not construed as pair
The restriction of range for those of ordinary skill in the art without creative efforts, can also be according to this
A little attached drawings obtain other relevant attached drawings.
Fig. 1 is spindle-type deep neural network structural schematic diagram of the invention；
Fig. 2 is that schematic diagram is embodied in spindle-type deep neural network of the invention；
Fig. 3 is training data X-Y scheme of the invention；
Fig. 4 is training data three-dimensional map of the invention；
Fig. 5 is training data two-dimensional map figure of the invention；
Fig. 6 is training data schematic diagram of the invention；
Fig. 7 is emulation testing data set schematic diagram data of the invention；
Fig. 8 is testing result schematic diagram after analogous diagram background of the invention inhibits；
Detection probability P when Fig. 9 is SNR of the invention < 10dWith false alarm rate PfaBetween relational graph；
Detection probability P when Figure 10 is 10 < SNR of the invention < 20dWith false alarm rate PfaBetween relational graph；
Detection probability P when Figure 11 is 20 < SNR of the invention < 30dWith false alarm rate PfaBetween relational graph；
Detection probability P when Figure 12 is 30 < SNR of the invention < 40dWith false alarm rate PfaBetween relational graph；
Figure 13 is P of the inventionfaUnder=1e-4, PdFigure is detected with the variation effect of SNR；
Figure 14 is P of the inventionfa=10e-4, PdFigure is detected with the variation effect of SNR；
Figure 15 is P of the inventionfa=20e-4, PdFigure is detected with the variation effect of SNR；
Figure 16 is P of the inventionfa=30e-4, PdFigure is detected with the variation effect of SNR；
Figure 17 is that the effect of DL algorithm of the invention illustrates enlarged drawing.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that described herein, specific examples are only used to explain the present invention, not
For limiting the present invention, i.e., described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is logical
The component for the embodiment of the present invention being often described and illustrated herein in the accompanying drawings can be arranged and be designed with a variety of different configurations.
Therefore, the detailed description of the embodiment of the present invention provided in the accompanying drawings is not intended to limit below claimed
The scope of the present invention, but be merely representative of selected embodiment of the invention.Based on the embodiment of the present invention, those skilled in the art
Member's every other embodiment obtained without making creative work, shall fall within the protection scope of the present invention.
It should be noted that the relational terms of term " first " and " second " or the like be used merely to an entity or
Operation is distinguished with another entity or operation, and without necessarily requiring or implying between these entities or operation, there are any
This actual relationship or sequence.Moreover, the terms "include", "comprise" or its any other variant be intended to it is non-exclusive
Property include so that include a series of elements process, method, article or equipment not only include those elements, but also
Further include other elements that are not explicitly listed, or further include for this process, method, article or equipment it is intrinsic
Element.In the absence of more restrictions, the element limited by sentence "including a ...", it is not excluded that including described
There is also other identical elements in the process, method, article or equipment of element.
Technical problem: solve the problems, such as that existing Weak target is caused detection accuracy low by noise and interference effect
Technological means:
A kind of detection method of small target based on Feature Mapping neural network, includes the following steps:
Step 1: building, training spindle-type deep neural network；
Step 2: the spindle-type deep neural network that the Weak target image input of acquisition has been trained is obtained into targets improvement
The amplitude figure inhibited with background；
Step 3: amplitude figure completes Dim targets detection using constant false alarm rate method.
Step 1 includes the following steps:
Step 1.1: building include input layer, decoding layer, coding layer and softmax output layer spindle-type depth nerve net
The structure of network；
Step 1.2: determining that the hyper parameter of spindle-type deep neural network obtains spindle-type depth using cross validation method
Neural network；
Step 1.3: building training dataset；
Step 1.4: training dataset input spindle-type deep neural network is trained acquisition using unsupervised mode
It initializes network weight and completes training.
Decoding layer, coding layer training calculate as follows:
hk=σ (WkX+bk)
Wherein, WkIndicate weight matrix, bkIndicate that bias vector, σ indicate activation primitive, X={ x1,x2,...,xmIndicate
The input of current layer, hkIndicate the output of current layer.
Step 2 includes the following steps:
Step 2.1: after the spindle-type deep neural network that the Weak target image input of acquisition has been trained, by small and weak mesh
Mark sample label is set as 1, and background sample label is set as 0；
Step 2.2: differentiate Weak target in acquisition image to Weak target image using the method for sliding window
Probability value；
Step 2.3: width is responded as window coordinates point using result, that is, multiple probability values that output layer logistic is returned
The amplitude figure that value obtains the corresponding targets improvement of multiple Weak targets and background inhibits.
Step 3 includes the following steps:
Step 3.1: each value in amplitude figure extracts sub-block using sliding windowIt is detected, inputs spindle-type
Deep neural network obtains amplitude；
Step 3.2: statistics being carried out to each amplitude after detection using constant false alarm rate and obtains false-alarm probability；
Wherein, T indicates the threshold value of Likelihood ration test,Indicate the mean value of sliding window sub-block, p indicates point in window sub-block
Number, PfaIndicate the false-alarm probability of constant false alarm rate detection setting, τCFARIndicate detection threshold value, F1,p-1(τCFAR) indicate center F
The cumulative distribution function of stochastic variable；
Step 3.3: candidate target sum being calculated according to false-alarm probability, false-alarm probability is sorted from high to low, from multiple width
Detection positioning target in value.
Building training dataset in step 1.3 includes the following steps:
Step 1.3.1: coordinate points are randomly generated in not including Weak target image as simulation objectives, extract N*N window
Mouth region domain is as background sample；
Step 1.3.2: a simulation objectives are added as target sample using dimensional Gaussian strength model in background sample
This, dimensional Gaussian model is as follows:
Wherein, (x0,y0) indicate target image center, s (i, j) indicate target image position (i, j) pixel
Value, sEIndicate generate target intensity, value be (0,1] between random number, σxAnd σyRespectively indicate horizontal and vertical distribution ginseng
Number, value is between [0,2]；
Step 1.3.3: the different parameters for adjusting target sample generate the Weak target completion training data of different signal-to-noise ratio
The building of collection.
Technical effect:
The present invention uses spindle network structure, and the Weak target Block Characteristic of low-dimensional is mapped to higher dimensional space first,
Then high resoluting characteristic is extracted by encoding nerve network, completes background and target-recognition, the intensity of output is differentiated according to network
The image that background inhibits targets improvement is obtained, is finally completed using the detection method based on constant false alarm rate to Dim targets detection,
It solves the problems, such as that existing Weak target is caused detection accuracy low by noise and interference effect, has reached and improved the powerful table of network
Show ability, under high-noise environment, improves the effect of Dim targets detection precision.
Feature and performance of the invention are described in further detail with reference to embodiments.
Embodiment 1
For the training data of network as shown in figure 3, small circle indicates an intermediate item, x point indicates in addition a kind of sample, training sample
Originally it is linearly inseparable, there is biggish differentiation difficulty；After being trained by training sample to depth network, output second
After layer and third layer feature are as shown in figure 4, the two dimensional character of linearly inseparable is mapped to three-dimensional by neural network, in three-dimensional spy
Levying space has linear discriminability, after three-dimensional feature is re-encoded into two dimension, as shown in figure 5, originally can not linearly inseparable
Low-dimensional feature be mapped to and can linearly differentiate feature space, and blue x point sample distribution is in compact region.Such as figure
Shown in 2, it includes 6 layers that whole network, which has altogether, can be divided into 2 parts, input layer is decoding layer to the 2nd layer, completes the liter of feature
Dimension, the mapping of low-dimensional to higher-dimension；2 layers to 5 layers are typical sparse self-encoding encoders, realize that the coding of abstract advanced features extracts,
The last layer is softmax output layer；Network model proposed in this paper and the typical deep neural network model main distinction are
For first layer to the decoding layer of the second layer, conventional depth neural network is mainly to handle high dimensional data, extracts the spy of abstract dimensionality reduction
Sign；Smaller based on Dim targets detection window, the lower feature of dimension realizes decoding operate in the front end of network first, so that
Network has more powerful expression ability；Under high-noise environment, Dim targets detection precision is higher；Building, training spindle-type
Deep neural network: step 1.1: building include input layer, decoding layer, coding layer and softmax output layer spindle-type depth
The structure of neural network；Step 1.2: the hyper parameter of spindle-type deep neural network is determined using cross validation method；Step
1.3: building training dataset；Step 1.4: by training dataset input spindle-type deep neural network using unsupervised mode into
Row training obtains initialization network weight and completes training.Complete training each layer of network model size be [81,512,256,
121,81,1], wherein input layer { I1, I2 ..., IN } be image in Dim targets detection window pixel linear array；Feature
Conversion, we use sparse self-encoding encoder, and calculation formula is as follows:
Wherein, x(i)It indicates to encode the activity of neural network hidden neuron j certainly when given input is x,
Indicate the average active degree of hidden neuron j；ρ indicates sparsity parameter, and β indicates hyper parameter.
ρ is set as 0.5, β and is set as 3, and using gradient descent method learning network parameter, learning rate is set as 0.01；It is instructing
When white silk, Weak target sample label is set as 1, and background sample label is set as 0, using the method for sliding window to image
All region input networks are differentiated, respond width as window coordinates point using the result that output layer logistic is returned
Value, the value illustrates that probability of the detection window comprising Weak target is higher, obtains mesh by spindle-type deep neural network
The amplitude figure that mark enhancing and background inhibit is Io；Amplitude figure completes Dim targets detection using constant false alarm rate method；Trained number
It is obtained according to collection using emulation mode, by manually adding Weak target in the image that 220 width do not include Weak target, emulates structure
Make training set；Coordinate points are randomly generated in every piece image, extract the region of 9*9 as background sample, in background sample
Use dimensional Gaussian strength model plus a simulation objectives as target sample, dimensional Gaussian model is as follows:
Wherein, (x0,y0) indicate target image center, s (i, j) indicate target image position (i, j) pixel
Value, sEIndicate generate target intensity, value be (0,1] between random number, σxAnd σyRespectively indicate horizontal and vertical distribution ginseng
Number, value is between [0,2].
The Weak target that different signal-to-noise ratio are generated by adjusting different parameters, the Weak target signal-to-noise ratio generated herein between
Between [0-120], each sample is substantially uniformly distributed in this signal-to-noise ratio interval range, altogether comprising 26400 positive samples and
26400 negative samples, part generate training sample image as shown in fig. 6, emulation testing data set partial data is as shown in Figure 7.
Test set includes emulation testing collection and truthful data test set, and image data includes millimeter-wave image and infrared figure
Picture, infrared image use the identical Weak target emulation mode of training data from multiple data sets, emulation testing data set,
The Weak target of different signal-to-noise ratio is added at random in background image, addition 1920 is small and weak in 32 width images altogether by the application
Target, equally their signal-to-noise ratio are approached and are evenly distributed between [0-120] Db；Test set data are inputted into housebroken spinning
Capitate neural network completes test, and the Weak target Block Characteristic of low-dimensional is mapped to higher-dimension first by spindle neural network
Then high resoluting characteristic is extracted by encoding nerve network in space, complete background and target-recognition, differentiates output according to network
Intensity obtains the image that background inhibits targets improvement, finally completes to examine Weak target using the detection method based on constant false alarm rate
It surveys, effectively improves the detection accuracy of Weak target.
Embodiment 2
Step 1 includes the following steps:
Step 1.1: building include input layer, decoding layer, coding layer and softmax output layer spindle-type depth nerve net
The structure of network；
Step 1.2: determining that the hyper parameter of spindle-type deep neural network obtains spindle-type depth using cross validation method
Neural network；
Step 1.3: building training dataset；
Step 1.4: training dataset input spindle-type deep neural network is trained acquisition using unsupervised mode
It initializes network weight and completes training.
Decoding layer, coding layer training calculate as follows:
hk=σ (WkX+bk)
Wherein, WkIndicate weight matrix, bkIndicate that bias vector, σ indicate activation primitive, X={ x1,x2,...,xmIndicate
The input of current layer, hkIndicate the output of current layer.
Step 2 includes the following steps:
Step 2.1: after the spindle-type deep neural network that the Weak target image input of acquisition has been trained, by small and weak mesh
Mark sample label is set as 1, and background sample label is set as 0, image it is specific as follows:
Wherein, s indicates that sensor acquires image, stIndicate echo signal, sbIndicate that background signal, n indicate noise.
Step 2.2: differentiate Weak target in acquisition image to Weak target image using the method for sliding window
Probability value；
Step 2.3: width is responded as window coordinates point using result, that is, multiple probability values that output layer logistic is returned
The amplitude figure that value obtains the corresponding targets improvement of multiple Weak targets and background inhibits.
Step 3 includes the following steps:
Step 3.1: each value in amplitude figure extracts sub-block using sliding windowIt is detected, input network obtains
Take amplitude；
Step 3.2: statistics being carried out to each amplitude after detection using constant false alarm rate and obtains false-alarm probability；
Wherein, T indicates the threshold value of Likelihood ration test,Indicate the mean value of sliding window sub-block, p indicates point in window sub-block
Number, PfaIndicate the false-alarm probability of constant false alarm rate detection setting, τCFARIndicate detection threshold value, F1,p-1(τCFAR) indicate center F
The cumulative distribution function of stochastic variable；
Step 3.3: candidate target sum being calculated according to false-alarm probability, false-alarm probability is sorted from high to low, from multiple width
Detection positioning target in value.
Building training dataset in step 1.3 includes the following steps:
Step 1.3.1: coordinate points are randomly generated in not including Weak target image as simulation objectives, extract N*N window
Mouth region domain is as background sample；
Step 1.3.2: a simulation objectives are added as target sample using dimensional Gaussian strength model in background sample
This, dimensional Gaussian model is as follows:
Wherein, (x0,y0) indicate target image center, s (i, j) indicate target image position (i, j) pixel
Value, sEIndicate generate target intensity, value be (0,1] between random number, σxAnd σyRespectively indicate horizontal and vertical distribution ginseng
Number, value is between [0,2]；
Step 1.3.3: the different parameters for adjusting target sample generate the Weak target completion training data of different signal-to-noise ratio
The building of collection.
By the detection effect of comparative analysis other several mainstream algorithm detection effects and the application, the essence of the application is embodied
Degree.Using two class curves as evaluation index, first kind curve is ROC curve, and what it reflected in target detection is that detection is general
Rate PdWith false alarm rate PfaBetween variation relation, the area under ROC curve is bigger, and detection performance is better, PdWith PfaCalculating it is public
Formula is as follows:
Wherein, NtIndicate the destination number being correctly detecting, NaIndicate the total quantity of target, NfExpression detects target
False quantity, N indicate the quantity of all pixels point in image.
Second class curve is detection probability PdVariation relation between Signal to Noise Ratio (SNR), with the increase of SNR value, PdIt will be by
Gradual change is big, finally levels off to 1, the SNR calculation formula of use are as follows:
Wherein, gtIndicate the average value of target regional area pixel, gbAnd σbIt is average to respectively indicate background regional area pixel
Value and standard deviation.
Several mainstream algorithm detection effects include: ACSDM, CSCD, SR, ISTCR, ISTCSR-CSCD, what the application indicated
It is DL algorithm；It is as shown in Figure 8: to respectively represent ACSDM, CSCD, SR, ISTCR, ISTCSR-CSCD and DL algorithm from left to right；
Wherein, green solid lines frame indicate target actual position, red block indicate detector output, if detector output position and
Target actual position is overlapped, then red block can cover green frame, the target point only real red block, the mesh of only green frame
Mark indicates that detector missing inspection occurs under specified false alarm rate, and DL algorithm is relative to other several algorithms as can be seen from Fig.,
The real goal number detected is most, and the destination number of missing inspection is minimum；
The quantitative analysis results of 6 kinds of algorithms as shown in figs9-12, Fig. 9-12 describe under different Signal to Noise Ratio (SNR) (SNR <
10,10 < SNR < 20,20 < SNR < 30,30 < SNR < 40) detection probability PdWith false alarm rate PfaBetween relationship, indicate for star
Solid line be DL algorithm proposed in this paper result, it can be seen that the application algorithm generally will on 4 kinds of difference sections SNR
Better than other 5 kinds of algorithms；In SNR < 10, in Pfa=1 × 10-4When, ISTCR algorithm is better than context of methods, has best
As a result, our method Dl of remaining situation has the precision preferably detected；In 10 < SNR < 20, this paper algorithm is each
Verification and measurement ratio is all much higher than remaining method under false alarm rate, is higher by about 20% compared to the method CSCD, DL being number two；In 20 < SNR
When < 30, it is higher than similar best method about 12%；In 30 < SNR < 40, context of methods is also significantly better than congenic method, about high
8% or so；
Figure 13-16 is indicated in identical PfaUnder, PdWith the situation of change of SNR, indicate that the solid line for star represents this Shen
The DL algorithm that please be propose, it can be seen that in 4 kinds of different constant false alarm rates, DL algorithm all obtains best as a result, outstanding
It is in Pfa>10×10-4When, context of methods is averagely higher than similar best method 20%；In Pfa=1 × 10-4When, DL, ISTCR
Relatively with tri- kinds of algorithm detection method performances of ISTCR-CSCD, but it is substantially better than its excess-three kind algorithm.Figure 17 is DL algorithm
Detection effect enlarged diagram because this detection using color distinguish actual position frame and detection output box, according to Patent Law
Carry out black and white after effect may not be it is clear that it may be necessary to when can provide cromogram.
By constructing spindle neural network structure, learns the feature of Weak target under complex background, pass through depth nerve
Network carries out image block to differentiate output destination probability, and target area possesses higher probability, and background area probability is lower, with this
Probability constructs target strength map, then carries out target detection and positioning with constant false alarm rate, that is, CFAR method, with mainstream algorithm phase
Than, the detection accuracy of DL algorithm averagely improves on about 15%, especially authentic testing image, the far super congenic method of DL performance,
Based on the method for study under complex background, there is better detectability, in the case of illustrating that DL method has high measurement accuracy,
Also possess the generalization ability of remote super congenic method.
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
Made any modifications, equivalent replacements, and improvements etc., should all be included in the protection scope of the present invention within mind and principle.
Claims (6)
1. a kind of detection method of small target based on Feature Mapping neural network, characterized by the following steps:
Step 1: building, training spindle-type deep neural network；
Step 2: the spindle-type deep neural network that the Weak target image input of acquisition has been trained is obtained into targets improvement and back
The amplitude figure that scape inhibits；
Step 3: amplitude figure completes Dim targets detection using constant false alarm rate method.
2. a kind of detection method of small target based on Feature Mapping neural network according to claim 1, feature exist
In: the step 1 includes the following steps:
Step 1.1: building includes the spindle-type deep neural network of input layer, decoding layer, coding layer and softmax output layer
Structure；
Step 1.2: determining that the hyper parameter of spindle-type deep neural network obtains spindle-type depth nerve using cross validation method
Network；
Step 1.3: building training dataset；
Step 1.4: training dataset input spindle-type deep neural network is trained acquisition initially using unsupervised mode
Change network weight and completes training.
3. a kind of detection method of small target based on Feature Mapping neural network according to claim 2, feature exist
In: the decoding layer, coding layer training calculate as follows:
hk=σ (WkX+bk)
Wherein, WkIndicate weight matrix, bkIndicate that bias vector, σ indicate activation primitive, X={ x1,x2,...,xmIndicate current
The input of layer, hkIndicate the output of current layer.
4. a kind of detection method of small target based on Feature Mapping neural network according to claim 1, feature exist
In: the step 2 includes the following steps:
Step 2.1: after the spindle-type deep neural network that the Weak target image input of acquisition has been trained, by Weak target sample
This label is set as 1, and background sample label is set as 0；
Step 2.2: Weak target image being carried out to differentiate the probability for obtaining Weak target in image using the method for sliding window
Value；
Step 2.3: being obtained using result, that is, multiple probability values that output layer logistic is returned as window coordinates point response amplitude
The amplitude figure for taking the corresponding targets improvement of multiple Weak targets and background to inhibit.
5. a kind of detection method of small target based on Feature Mapping neural network according to claim 1, feature exist
In: the step 3 includes the following steps:
Step 3.1: each value in amplitude figure extracts sub-block using sliding windowIt is detected, inputs spindle-type depth
Neural network obtains amplitude；
Step 3.2: statistics being carried out to each amplitude after detection using constant false alarm rate and obtains false-alarm probability；
In, T indicates the threshold value of Likelihood ration test,Indicate the mean value of sliding window sub-block, p indicates put in window sub-block
Number, PfaIndicate the false-alarm probability of constant false alarm rate detection setting, τCFARIndicate detection threshold value, F1,p-1(τCFAR) indicate that center F is random
The cumulative distribution function of variable；
Step 3.3: candidate target sum being calculated according to false-alarm probability, false-alarm probability is sorted from high to low, from multiple amplitudes
Detection positioning target.
6. a kind of detection method of small target based on Feature Mapping neural network according to claim 2, feature exist
In: the building training dataset in the step 1.3 includes the following steps:
Step 1.3.1: coordinate points are randomly generated in not including Weak target image as simulation objectives, extract N*N window region
Domain is as background sample；
Step 1.3.2: in background sample using dimensional Gaussian strength model plus simulation objectives as target sample, two
It is as follows to tie up Gauss model:
Wherein, (x0,y0) indicate target image center, s (i, j) indicate target image position (i, j) pixel value,
sEIndicate generate target intensity, value be (0,1] between random number, σxAnd σyHorizontal and vertical distribution parameter is respectively indicated,
Its value is between [0,2]；
Step 1.3.3: the different parameters for adjusting target sample generate the Weak targets of different signal-to-noise ratio and complete training dataset
Building.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810729648.0A CN109002848B (en) | 2018-07-05 | 2018-07-05 | Weak and small target detection method based on feature mapping neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810729648.0A CN109002848B (en) | 2018-07-05 | 2018-07-05 | Weak and small target detection method based on feature mapping neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109002848A true CN109002848A (en) | 2018-12-14 |
| CN109002848B CN109002848B (en) | 2021-11-05 |
Family
ID=64598687
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810729648.0A Active CN109002848B (en) | 2018-07-05 | 2018-07-05 | Weak and small target detection method based on feature mapping neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109002848B (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109767434A (en) * | 2019-01-07 | 2019-05-17 | 西安电子科技大学 | Time domain detection method of small target neural network based |
| CN109902715A (en) * | 2019-01-18 | 2019-06-18 | 南京理工大学 | A kind of method for detecting infrared puniness target based on context converging network |
| CN110210518A (en) * | 2019-05-08 | 2019-09-06 | 北京互金新融科技有限公司 | The method and apparatus for extracting dimensionality reduction feature |
| CN112288778A (en) * | 2020-10-29 | 2021-01-29 | 电子科技大学 | Infrared small target detection method based on multi-frame regression depth network |
| CN113327253A (en) * | 2021-05-24 | 2021-08-31 | 北京市遥感信息研究所 | Weak and small target detection method based on satellite-borne infrared remote sensing image |
| CN109002848B (en) * | 2018-07-05 | 2021-11-05 | 西华大学 | Weak and small target detection method based on feature mapping neural network |
| CN115222775A (en) * | 2022-09-15 | 2022-10-21 | 中国科学院长春光学精密机械与物理研究所 | Weak and small target detection tracking device and detection tracking method thereof |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2005103929A1 (en) * | 2004-04-20 | 2005-11-03 | Pluck Corporation | Method, system, and computer program product for sharing information within a global computer network |
| CN103812577A (en) * | 2012-11-06 | 2014-05-21 | 西南交通大学 | Method for automatically identifying and learning abnormal radio signal type |
| CN105844627A (en) * | 2016-03-21 | 2016-08-10 | 华中科技大学 | Sea surface object image background inhibition method based on convolution nerve network |
| CN106056097A (en) * | 2016-08-17 | 2016-10-26 | 西华大学 | Millimeter wave weak small target detection method |
| CN107067413A (en) * | 2016-12-27 | 2017-08-18 | 南京理工大学 | A kind of moving target detecting method of time-space domain statistical match local feature |
| CN108122003A (en) * | 2017-12-19 | 2018-06-05 | 西北工业大学 | A kind of Weak target recognition methods based on deep neural network |
| CN112507840A (en) * | 2020-12-02 | 2021-03-16 | 中国船舶重工集团公司第七一六研究所 | Man-machine hybrid enhanced small target detection and tracking method and system |
| CN113065558A (en) * | 2021-04-21 | 2021-07-02 | 浙江工业大学 | Lightweight small target detection method combined with attention mechanism |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109002848B (en) * | 2018-07-05 | 2021-11-05 | 西华大学 | Weak and small target detection method based on feature mapping neural network |
-
2018
- 2018-07-05 CN CN201810729648.0A patent/CN109002848B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2005103929A1 (en) * | 2004-04-20 | 2005-11-03 | Pluck Corporation | Method, system, and computer program product for sharing information within a global computer network |
| CN103812577A (en) * | 2012-11-06 | 2014-05-21 | 西南交通大学 | Method for automatically identifying and learning abnormal radio signal type |
| CN105844627A (en) * | 2016-03-21 | 2016-08-10 | 华中科技大学 | Sea surface object image background inhibition method based on convolution nerve network |
| CN106056097A (en) * | 2016-08-17 | 2016-10-26 | 西华大学 | Millimeter wave weak small target detection method |
| CN107067413A (en) * | 2016-12-27 | 2017-08-18 | 南京理工大学 | A kind of moving target detecting method of time-space domain statistical match local feature |
| CN108122003A (en) * | 2017-12-19 | 2018-06-05 | 西北工业大学 | A kind of Weak target recognition methods based on deep neural network |
| CN112507840A (en) * | 2020-12-02 | 2021-03-16 | 中国船舶重工集团公司第七一六研究所 | Man-machine hybrid enhanced small target detection and tracking method and system |
| CN113065558A (en) * | 2021-04-21 | 2021-07-02 | 浙江工业大学 | Lightweight small target detection method combined with attention mechanism |
Non-Patent Citations (10)
| Title |
|---|
| GAO, ZHISHENG等: "Dim and small target detection based on feature mapping neural networks", 《JOURNAL OF VISUAL COMMUNICATION & IMAGE REPRESENTATION》 * |
| LONG GENG等: "Dim small target detection in single frame complex background", 《2016 2ND IEEE INTERNATIONAL CONFERENCE ON COMPUTER AND COMMUNICATIONS (ICCC)》 * |
| MIAO ZHANG等: "Space debris detection methods utilizing hyperspectral sequence analysis based on Hilbert-Huang transformation", 《2012 IEEE INTERNATIONAL INSTRUMENTATION AND MEASUREMENT TECHNOLOGY CONFERENCE PROCEEDINGS》 * |
| 于周吉: "基于卷积神经网络的红外弱小目标检测算法", 《光学与光电技术》 * |
| 刘俊明等: "融合全卷积神经网络和视觉显著性的红外小目标检测", 《光子学报》 * |
| 娄康等: "基于卷积神经网络与高斯混合建模的红外弱小目标检测方法", 《第三十八届中国控制会议论文集（7）》 * |
| 流浪的MOOK: "红外弱小目标检测算法综述", 《HTTPS://BLOG.CSDN.NET/M0_51625082/ARTICLE/DETAILS/118963429》 * |
| 蔡智富: "基于自适应背景估计的复杂红外背景抑制技术", 《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》 * |
| 高志升等: "具有高斯噪声不变性的特征描述算子", 《计算机工程与应用》 * |
| 高志升等: "采用目标背景建模的毫米波弱小目标检测", 《光学精密工程》 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109002848B (en) * | 2018-07-05 | 2021-11-05 | 西华大学 | Weak and small target detection method based on feature mapping neural network |
| CN109767434A (en) * | 2019-01-07 | 2019-05-17 | 西安电子科技大学 | Time domain detection method of small target neural network based |
| CN109767434B (en) * | 2019-01-07 | 2023-04-07 | 西安电子科技大学 | Time domain weak and small target detection method based on neural network |
| CN109902715A (en) * | 2019-01-18 | 2019-06-18 | 南京理工大学 | A kind of method for detecting infrared puniness target based on context converging network |
| CN110210518A (en) * | 2019-05-08 | 2019-09-06 | 北京互金新融科技有限公司 | The method and apparatus for extracting dimensionality reduction feature |
| CN110210518B (en) * | 2019-05-08 | 2021-05-28 | 北京互金新融科技有限公司 | Method and device for extracting dimension reduction features |
| CN112288778A (en) * | 2020-10-29 | 2021-01-29 | 电子科技大学 | Infrared small target detection method based on multi-frame regression depth network |
| CN113327253A (en) * | 2021-05-24 | 2021-08-31 | 北京市遥感信息研究所 | Weak and small target detection method based on satellite-borne infrared remote sensing image |
| CN113327253B (en) * | 2021-05-24 | 2024-05-24 | 北京市遥感信息研究所 | Weak and small target detection method based on satellite-borne infrared remote sensing image |
| CN115222775A (en) * | 2022-09-15 | 2022-10-21 | 中国科学院长春光学精密机械与物理研究所 | Weak and small target detection tracking device and detection tracking method thereof |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109002848B (en) | 2021-11-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109002848A (en) | A kind of detection method of small target based on Feature Mapping neural network | |
| CN110472627B (en) | End-to-end SAR image recognition method, device and storage medium | |
| CN113050042B (en) | Radar signal modulation type identification method based on improved UNet3+ network | |
| CN110766051A (en) | Lung nodule morphological classification method based on neural network | |
| CN109101897A (en) | Object detection method, system and the relevant device of underwater robot | |
| CN109871902B (en) | SAR small sample identification method based on super-resolution countermeasure generation cascade network | |
| CN106780485A (en) | SAR image change detection based on super-pixel segmentation and feature learning | |
| CN107784663A (en) | Correlation filtering tracking and device based on depth information | |
| CN109902715B (en) | Infrared dim target detection method based on context aggregation network | |
| CN110213244A (en) | A kind of network inbreak detection method based on space-time characteristic fusion | |
| CN104408481B (en) | Classification of Polarimetric SAR Image method based on depth wavelet neural network | |
| CN109543688A (en) | A kind of novel meter reading detection and knowledge method for distinguishing based on multilayer convolutional neural networks | |
| CN106056097B (en) | Millimeter wave detection method of small target | |
| CN105844279A (en) | Depth learning and SIFT feature-based SAR image change detection method | |
| CN108830331A (en) | A kind of Ground Penetrating Radar object detection method based on full convolutional network | |
| CN109740485A (en) | Reservoir or dyke recognition methods based on spectrum analysis and depth convolutional neural networks | |
| Qi et al. | FTC-Net: Fusion of transformer and CNN features for infrared small target detection | |
| CN108280412A (en) | High Resolution SAR image object detection method based on structure changes CNN | |
| CN109543585A (en) | Underwater optics object detection and recognition method based on convolutional neural networks | |
| Liu et al. | Image edge recognition of virtual reality scene based on multi-operator dynamic weight detection | |
| CN110348434A (en) | Camera source discrimination method, system, storage medium and calculating equipment | |
| CN108399430A (en) | A kind of SAR image Ship Target Detection method based on super-pixel and random forest | |
| CN111798404B (en) | Iris image quality evaluation method and system based on deep neural network | |
| CN112784777A (en) | Unsupervised hyperspectral image change detection method based on antagonistic learning | |
| Marques et al. | Target detection in SAR images based on a level set approach |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20230413 Address after: 1002, Building A, Zhiyun Industrial Park, No. 13 Huaxing Road, Henglang Community, Dalang Street, Longhua District, Shenzhen City, Guangdong Province, 518110 Patentee after: Shenzhen Wanzhida Technology Co.,Ltd. Address before: 610039, No. 999, Jin Zhou road, Jinniu District, Sichuan, Chengdu Patentee before: XIHUA University Effective date of registration: 20230413 Address after: Room 3303-3306, Building 2, Binjiang Business Center, No. 1888 Ganjiang South Avenue, Honggutan District, Nanchang City, Jiangxi Province, 330000 Patentee after: Jiangxi Chengan Technology Co.,Ltd. Address before: 1002, Building A, Zhiyun Industrial Park, No. 13 Huaxing Road, Henglang Community, Dalang Street, Longhua District, Shenzhen City, Guangdong Province, 518110 Patentee before: Shenzhen Wanzhida Technology Co.,Ltd. |
|
| TR01 | Transfer of patent right |