CN107330447A - The outline identifying system that a kind of reaction type ICM neutral nets and FPF are combined - Google Patents
The outline identifying system that a kind of reaction type ICM neutral nets and FPF are combined Download PDFInfo
- Publication number
- CN107330447A CN107330447A CN201710414874.5A CN201710414874A CN107330447A CN 107330447 A CN107330447 A CN 107330447A CN 201710414874 A CN201710414874 A CN 201710414874A CN 107330447 A CN107330447 A CN 107330447A
- Authority
- CN
- China
- Prior art keywords
- image
- fpf
- icm
- target
- outline
- 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
- 238000006757 chemical reactions by type Methods 0.000 title claims abstract description 31
- 230000007935 neutral effect Effects 0.000 title claims abstract description 18
- 230000001965 increasing effect Effects 0.000 claims abstract description 9
- 230000008713 feedback mechanism Effects 0.000 claims abstract description 8
- 230000009467 reduction Effects 0.000 claims abstract description 5
- 238000005728 strengthening Methods 0.000 claims abstract description 4
- 230000001629 suppression Effects 0.000 claims abstract description 4
- 239000000284 extract Substances 0.000 claims abstract description 3
- 238000012549 training Methods 0.000 claims description 32
- 239000011159 matrix material Substances 0.000 claims description 27
- 210000002569 neuron Anatomy 0.000 claims description 21
- 230000006870 function Effects 0.000 claims description 16
- 230000001054 cortical effect Effects 0.000 claims description 14
- 230000000007 visual effect Effects 0.000 claims description 14
- 235000009421 Myristica fragrans Nutrition 0.000 claims description 12
- 239000001115 mace Substances 0.000 claims description 12
- 238000000034 method Methods 0.000 claims description 12
- 230000000694 effects Effects 0.000 claims description 9
- 230000002708 enhancing effect Effects 0.000 claims description 8
- 238000012545 processing Methods 0.000 claims description 5
- 230000015572 biosynthetic process Effects 0.000 claims description 3
- 238000010586 diagram Methods 0.000 claims description 3
- 230000005284 excitation Effects 0.000 claims description 3
- 230000036279 refractory period Effects 0.000 claims description 2
- 238000004519 manufacturing process Methods 0.000 claims 1
- 238000010998 test method Methods 0.000 claims 1
- 238000004364 calculation method Methods 0.000 abstract description 5
- 230000008859 change Effects 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 5
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 3
- 230000008878 coupling Effects 0.000 description 2
- 238000010168 coupling process Methods 0.000 description 2
- 238000005859 coupling reaction Methods 0.000 description 2
- 235000013399 edible fruits Nutrition 0.000 description 2
- 230000005281 excited state Effects 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 230000000644 propagated effect Effects 0.000 description 2
- 238000000926 separation method Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 241000282836 Camelus dromedarius Species 0.000 description 1
- 241000124008 Mammalia Species 0.000 description 1
- 210000001367 artery Anatomy 0.000 description 1
- 230000002238 attenuated effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 210000003710 cerebral cortex Anatomy 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000005764 inhibitory process Effects 0.000 description 1
- 238000012804 iterative process Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 210000005036 nerve Anatomy 0.000 description 1
- 210000000653 nervous system Anatomy 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 230000006641 stabilisation Effects 0.000 description 1
- 238000011105 stabilization Methods 0.000 description 1
- 238000003786 synthesis reaction Methods 0.000 description 1
- 230000009897 systematic effect Effects 0.000 description 1
- 238000000844 transformation Methods 0.000 description 1
- 230000017105 transposition Effects 0.000 description 1
- 210000003462 vein Anatomy 0.000 description 1
- 210000000857 visual cortex Anatomy 0.000 description 1
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
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
A kind of outline identifying system that reaction type ICM neutral nets and FPF are combined, the pulse-couple characteristic having using ICMNN extracts the integrity profile of target image;And the feedback mechanism used is then continuous to original image strengthens, so as to reach the non-similar target of suppression while strengthening the purpose of similar target;With continuous iteration, FPF is in the image treated through ICM, constantly search for the candidate target similar with object edge, larger relevant peaks will be produced in relevant position when finding reliable similar target, and because the addition of feedback mechanism effectively suppresses other nontarget areas, so that the correlation reduction in other regions, so as to realize the reliable recognition of target.Present system can preferably recognize also there is preferable stability to the image with certain rotation and dimensional variation for image generic in figure.Amount of calculation has reduced compared with other outline recognizers, and workload is reduced accordingly, and speed is increased.
Description
Technical field
The outline identifying system that a kind of reaction type ICM neutral nets of the present invention and FPF are combined, is related to image recognition technology
Field.
Background technology
Outline identification is a kind of image recognition technology based on objective contour, is widely used in robot crawl, medical science
The field such as image procossing and CBIR.For many years, occur in that many outline recognizers, it is classical just like
Under several algorithms:1) method based on statistical learning, the method based on statistical learning is broken generally into the study stage (for training point
Class device) and sorting phase.The study stage is in the property field under different classes of pattern, to assess the distribution per class object edge;
Sorting phase is that the object edge distribution obtained using evaluation stage is classified to sketch figure picture.Such method is dependent on a large amount of
Statistical nature, and easily occurred learn or cross training phenomenon.2) method of neutral net, based on traditional neural network
Method, extracts a series of characteristic parameter, in sorting phase extraction in the study stage from different classes of target image
Profile feedback is classified to neutral net to image, and this method needs that substantial amounts of test image is carried out to learn that phase could be carried out
The image processing step of pass.3) method based on Dynamic Programming, dynamic programming is the image and former mesh unknown object edge
Then logo image calculates the beeline between two variables as a series of variables.If such method is not in advance using exhaustion
Method calculates connection weight, can not just use the progressively method of recursion from back to front commonly used in Dynamic Programming, therefore amount of calculation
It is very big.
The content of the invention
The deficiency of several classical outline recognizers for more than, the present invention proposes a kind of reaction type ICM neutral nets
The outline identifying system being combined with FPF, the system can be recognized preferably for image generic in figure, to certain
The image of rotation and dimensional variation also has preferable stability.Amount of calculation has subtracted compared with other outline recognizers
Small, workload is reduced accordingly, and speed is increased.
The technical scheme that the present invention takes is:
The outline identifying system that a kind of reaction type ICM neutral nets and FPF are combined, the pulse coupling having using ICMNN
Characteristic is closed, the integrity profile of target image is extracted;And the feedback mechanism used is then continuous to original image strengthens, so that
Reach the non-similar target of suppression while strengthening the purpose of similar target;With continuous iteration, FPF is in the image treated through ICM
In, the candidate target similar with object edge is constantly searched for, will be produced when finding reliable similar target in relevant position
Larger relevant peaks, and because the addition of feedback mechanism effectively suppresses other nontarget areas, so that the phase in other regions
Pass value reduction, so as to realize the reliable recognition of target.
The outline identifying system that a kind of reaction type ICM neutral nets and FPF are combined,
Step one:Set up intersecting visual cortical model ICM:
The minimum system that intersecting visual cortical model ICM is constituted is described with three below formula:
Fij[n+1]=fFij[n]+Sij+W{Y[n]}ij (l)
θij[n+1]=g θij[n]+hYij[n+1] (3)
Wherein, input matrix is S, and the state matrix of neuron is F, and output matrix is Y, and dynamic threshold matrix is θ, scalar
F and g is less than 1.0, makes g<F, it can be ensured that threshold value promotes neuron to excite terminating below the state value of neuron, scalar h are one
Very big value, when neuron is excited so that threshold value is sharply increased, so as to suppress the pulse granting of neuron, makes it into not
Ying Qi, i.e., no longer provide pulse over a period to come;
Step 2:With intersecting visual cortical model ICM to sketch figure picture processing, an arteries and veins for including objective contour can be produced
Image is rushed, FPF effect is scanned in the pulse image-region, if the candidate target similar to object edge, then
Correlation peak can be produced in its correlation surface, relevant peaks synthetic discriminant functions wave filter SDF and the minimum average B configuration correlation energy such as FPF is
Measure wave filter MACE extension;
SDF basic thought is:Certain class image and its fault image are constituted into a training set, comprehensive discriminating letter is found out
Number, the purpose of synthetic discriminant functions is to seek such a filter function:When correct target is inputted, there is a correlation in correlation surface
Peak value;Select one group of training image vi, it is h to make wave filter, and SDF is that matrix V is to pass through in weighted input linear combination, formula (6)
The Fourier transform of input vector is combined to create,
It is V Fourier transform.Each vectorAll with limits value CiIt is associated, therefore SDF wave filters h is to pass through
Following formula is constrained,
As formula (8) is usedInverse matrix go to find h,
Average minimum correlation energy wave filter MACE is developed on the basis of synthetic discriminant functions, utilizes MACE
Can minimize the average correlation energy after training image related operation;
MACE wave filters are to calculate to obtain by following formula:
Wherein:
It is i-th of element of k-th training image, two-dimentional fractional power exponential filter can be described as follows:
Wherein
X is created by training image, similar to formula (6),Element produce inside the image,It is Y+ kConjugation,
In formula (14), it is referred to as Minimum Average Correlation Energy Filter if power P=2, P=0 formulas are referred to as mirror
Other function filter, is referred to as fractional power exponential filter FPF when P is between 0-2.
Step 3:Result images are handled again using reaction type ICM, the figure that only target image is present is finally obtained
Picture.Based on this, realize that outline is recognized using one closed-loop system of reaction type ICM combinations FPF formation.
The outline identifying system that a kind of reaction type ICM neutral nets of the present invention and FPF are combined, has the beneficial effect that:
(1) system can be recognized preferably for image generic in figure, to becoming with certain rotation and yardstick
The image of change also has preferable stability.
(2) amount of calculation has reduced compared with other outline recognizers, and workload is reduced accordingly, and speed is increased
Plus.
Brief description of the drawings
Fig. 1 is curvature stream picture one;
Fig. 2 is curvature stream picture two;
Fig. 3 is curvature stream picture three;
Fig. 4 is curvature stream picture four.
Fig. 5 is ICM model framework charts.
Fig. 6 is to rotate the FPF training image figures under 10 ° of angles;
Fig. 7 is to rotate the FPF training image figures under 20 ° of angles;
Fig. 8 is to rotate the FPF training image figures under 30 ° of angles;
Fig. 9 is the result figure that fractional power exponential filter power P takes 0.1.
Figure 10 is the result figure that fractional power exponential filter power P takes 0.5.
Figure 11 is the result figure that fractional power exponential filter power P takes 0.9.
Figure 12 is FPF combinations ICM system block diagram.
Embodiment
Because non-rigid targets change in shape is various, therefore require that recognizer can not only effectively represent objective contour, together
When need to also it is extensive differentiate between carry out balance.The present invention is by the intersecting sight cortex model with biological vision nervous system
Neutral net (ICMNN) is combined with fractional power exponential filter (FPF), devises one kind using reaction type ICMNN separation mesh
Profile is marked, and realizes using FPF the outline identifying system of fast correlation.
The outline identifying system that a kind of reaction type ICM neutral nets and FPF are combined, the pulse coupling having using ICMNN
Characteristic is closed, the integrity profile of target image is extracted;And the feedback mechanism used is then continuous to original image strengthens, so that
Reach the non-similar target of suppression while strengthening the purpose of similar target;With continuous iteration, FPF is in the image treated through ICM
In, the candidate target similar with object edge is constantly searched for, will be produced when finding reliable similar target in relevant position
Larger relevant peaks, and because the addition of feedback mechanism effectively suppresses other nontarget areas, so that the phase in other regions
Pass value reduction, so as to realize the reliable recognition of target.
The outline identifying system that a kind of reaction type ICM neutral nets of the present invention and FPF are combined, including:
1:Intersecting visual cortical model (ICM) and reaction type visual cortex model (FICM):
1), intersecting visual cortical model:
Intersecting visual cortical model (ICM) is mainly based upon Eckhorn models and Rybak model evolutions, and it is many
The product of Cerebral cortex model cross-synthesis is planted, is a kind of Pulse-coupled Neural Network Model of simplification.ICM is than traditional artificial god
Through network model closer to real biological vision neural network, and ICM amounts of calculation are less;In addition, ICM is also using based on song
The centripetal wave technology that rate stream theory is realized overcomes auto-wave interference problem present in pulse-couple network.Using this technology,
The auto-wave in Pulse Coupled Neural Network can be made no longer outwards to propagate but be propagated to target's center, be finally punctured into a bit,
As shown in the waveform transmission of the target image of each in Fig. 1-4.This can be used for profile of the reduction caused by target scale changes greatly
Small change recognizes caused influence to outline.
The minimum system that ICM is constituted is described with three below formula, and Fig. 5 is its model framework chart.
Fij[n+1]=fFij[n]+Sij+W{y[n]}ij (1)
θij[n+1]=g θij[n]+hYij[n+1] (3)
Wherein, outside stimulus input is matrix S, and the current state matrix of neuron is F [n+1], and previous state is F [n],
W { * } represents the connection between neuron.Neuron is output as matrix Y, if Y, which is 1 expression neuron, is in excited state, such as
Fruit represents to be in OFF state for 0.Dynamic threshold matrix is θ, and θ is set according to the demand of user.Scalar f and g are less than
1.0, make g<F, it can be ensured that threshold value promotes neuron to excite terminating below the state value of neuron, scalar h are one very big
Value, the value 20 in testing herein.When neuron is excited so that threshold θ is sharply increased, so as to suppress the pulse hair of neuron
Put, make it into refractory period, i.e., no longer provide pulse over a period to come.
2), reaction type intersecting visual cortical model:
After ICM iteration, the region of simultaneous shot to certain time can extinguish, and this extinguishing, which is called, desynchronizes, and uses
This mechanism desynchronized can be partitioned into the texture information of image.The small variations that striated band comes constantly are propagated over time,
The change of neuron state is ultimately caused, such texture information is just extracted.In order to obtain more preferable texture information we
A kind of reaction type intersecting visual cortical model is used, reaction type intersecting visual cortical model presses down ICMNN output with one kind
The mode of property processed feeds back to input, and reaction type intersecting visual cortical model is similar with the olfactory system of mouse in mammal.ICM
In an iterative process can be by image interior zone and edge separation, but for reaction type ICM, it, which is exported, can feed back to input generation
Shunting, so as to cause uneven influence to whole input.Here it is ICM and reaction type ICM difference.
The neuron excited in ICM iteration will be output as input feedback to other neurons, like this complete in iteration
The regular hour is needed to can be only achieved stable state afterwards.Want to reach stabilization, it is necessary to meet some requirements, first, input
Value requirement have to be larger than threshold value, if input value is too low, and input gray level value will be attenuated (attenuation factor) quickly and finally disappear
Lose.The granting stage is rushed next to that exporting and safe must enter institute's dicrotism, in this case all nerves in excited state
Member will be excited always.When meeting secondary condition, feedback and connection weight matrix can become constant.In fact, output is only each other
Vertical, and exactly these outputs cause other fluctuations all in network.When exporting stable, there will be no change input and
The other factors of other network parameters.
The feed back input A of system is calculated according to the time weight average mode in formula (4), this dynamic threshold with ICM
Calculate similar, difference is that constant V value is different.
Wherein VAV than being used in dynamic threshold is much smaller, and V is made hereinA=1, after inhibition feedback is calculated
Input and be:
Reaction type intersects the computing that cortical visual model carries out formula (9) and (10) after each iteration has been calculated.
2:Fractional power exponential filter (FPF):With ICM to sketch figure picture processing, one can be produced and include objective contour
Pulse diagram picture, FPF effect is scanned in the pulse image-region, if the candidate target similar to object edge,
Correlation peak then can be produced in its correlation surface.FPF such as is at relevant peaks synthetic discriminant functions wave filter (SDF) and the minimum average B configuration phase
Close the extension of energy filter (MACE).
SDF basic thought is:Certain class image and its fault image are constituted into a training set, comprehensive discriminating letter is found out
Number, the purpose of synthetic discriminant functions is to seek such a filter function:When correct target is inputted, there is a correlation in correlation surface
Peak value.
Select one group of training image vi, it is h to make wave filter.SDF is that matrix V is logical in weighted input linear combination, formula (6)
The Fourier transform for crossing input vector is combined to create.
It is V Fourier transform.Each vectorAll with limits value CiIt is associated, therefore SDF wave filters h is to pass through
Following formula is constrained
As formula (8) is usedInverse matrix go to find h, whereinThe Fourier transformation shape of as minimum correlation energy wave filter
Formula.To carry out transposition to Fourier's matrix.[*]-1For to inverse of a matrix computing.C is limits value, be withMatrix abscissa is big
A small column matrix, its value size is 1.
Average minimum correlation energy wave filter (MACE) develops on the basis of synthetic discriminant functions[14], profit
Can minimize the average correlation energy after training image related operation with MACE.
MACE wave filters are to calculate to obtain by following formula:
D is built by formula (10), and c is the limits value that formula (7) is mentioned.Mean that average minimum correlation energy wave filter.
Wherein:
N is training image number, δijIt is the diagonal matrix that size is d*d, element value is that 1, d sizes are diagonal for training image
Element number.It is i-th of element of k-th training image.So DijIt is just the matrix that size is d*d, its element value is exactly
The pixel average of each training image same position.
Two-dimentional fractional power exponential filter can be described as follows:
Built by formula (13), andObtained by formula (12).C is the limits value in formula (7), and D is obtained by formula (10).Its
In:
Column vector is changed into for the 2D Fourier transformations of d*1 matrix, i.e. training image.D-1/2Represent to diagonal matrix D
Open radical sign.
Element produce inside the image,It isConjugation.
With formula (10) equally, power P=2 is then referred to as Minimum Average Correlation Energy Filter, P=0 formulas to formula (14)
Referred to as Discrimination Functions wave filter, is referred to as fractional power exponential filter FPF when P is between 0-2.
The camel sketch figure picture under three width different rotary angles is shown in Fig. 6-8, can be built using this three width image
One FPF, the corresponding space area images of FPF of different P values are as shown in figs. 9-11.As can be seen that the bigger FPF of P values is corresponding
As a result closer to the edge of training image.
3.. the outline identification that reaction type ICM and FPF are combined:
In outline identification, if only target be identified with ICM and FPF, the enhancing of obtained simply target image,
Non-object image is still present, in order that effect more preferably, is handled result images, finally obtained again using reaction type ICM
The image that only target image is present.Based on this, realize that outline is known using one closed-loop system of reaction type ICM combinations FPF formation
Not.
The workflow of system is shown in Figure 12:The sketch figure picture progress ICM iteration generations edge of input is swashed first
The image of hair, then edge enhancing is carried out to the image.Training image is inputted into FPF, the enhanced image in edge is obtained.Original
The enhanced result in image border carries out related operation with FPF, obtains correlation peak, and the region of correlation peak maximum point is carried out
Modification, then using amended image as original image input system, finally obtain enhancing of the target image on original image.
Experimental procedure:
Step 1:Sketch figure picture to input carries out the image that ICM iteration produces edge excitation, then carries out side to the image
Edge strengthens.
Step 2:Training image is inputted into FPF, the enhanced image in edge is obtained.
Step 3:The enhanced result in original image edge and FPF are subjected to related operation, correlation peak is obtained.
Step 4:The region of correlation peak maximum point is modified.
Step 5:Using amended image as original image input system, the enhancing on target image is finally obtained.
Systematic function is tested in experiment, has selected Kimia outline data to be tested:
(1) original images to be recognized and training image, are inputted;
(2) it is 0.3, to take FPF powers P, and image of the original images to be recognized after ICM is related to the FPF after training to be obtained
The peak value arrived is 0.99;
(3), different peak value thresholds are set, the enhanced result of different feedbacks can be obtained.It can obtain, when threshold value thresholding
When bigger, system identification effect is better;
(4) P=0.4, is taken, the correlation peak of generation is 0.1;
(5), again successively increase threshold value thresholding can obtain and (3) in identical result.That is the bigger identification of threshold value thresholding is imitated
Fruit is better;
(6) P value, is increased successively, repeating the step in (3), can to obtain the bigger image border of P values more obvious, with original graph
The correlation peak that picture is produced is bigger, and recognition effect is better.
Claims (5)
1. the outline identifying system that a kind of reaction type ICM neutral nets and FPF are combined, it is characterised in that:Have using ICMNN
Pulse-couple characteristic, extract target image integrity profile;And the feedback mechanism used then progress continuous to original image
Enhancing, so as to reach the non-similar target of suppression while strengthening the purpose of similar target;With continuous iteration, FPF is through ICM processing
In the image crossed, the candidate target similar with object edge is constantly searched for, will be corresponding when finding reliable similar target
Position produces larger relevant peaks, and because the addition of feedback mechanism effectively suppresses other nontarget areas, so that other
The correlation reduction in region, so as to realize the reliable recognition of target.
2. the outline identifying system that a kind of reaction type ICM neutral nets and FPF are combined, it is characterised in that:
Step one:Set up intersecting visual cortical model ICM:
The minimum system that intersecting visual cortical model ICM is constituted is described with three below formula:
Fij[n+1]=fFij[n]+Sij+W{Y[n]}ij (1)
θij[n+1]=g θij[n]+hYij[n+1] (3)
Wherein, input matrix is S, and the state matrix of neuron is F, and output matrix is Y, and dynamic threshold matrix is θ, scalar f and g
Less than 1.0, g is made<F, it can be ensured that threshold value promotes neuron to excite terminating below the state value of neuron, scalar h are one very big
Value, when neuron is excited so that threshold value is sharply increased, so as to suppress the pulse granting of neuron, make it into refractory period,
Pulse is no longer provided over a period to come;
Step 2:With intersecting visual cortical model ICM to sketch figure picture processing, a pulse diagram for including objective contour can be produced
Picture, FPF effect is scanned in the pulse image-region, if the candidate target similar to object edge, then can be
Produce correlation peak in its correlation surface, relevant peaks synthetic discriminant functions wave filter SDF and the minimum Average Correlation Energy filter such as FPF is
Ripple device MACE extension;
SDF basic thought is:Certain class image and its fault image are constituted into a training set, synthetic discriminant functions are found out, it is comprehensive
The purpose for closing Discrimination Functions is to seek such a filter function:When correct target is inputted, there is a correlation peak in correlation surface;
Select one group of training image vi, it is h to make wave filter, SDF be in weighted input linear combination, formula (6) matrix V be by input to
The Fourier transform of amount is combined to create,
It is V Fourier transform;Each vectorAll with limits value CiIt is associated, therefore SDF wave filters h is by following formula
Come what is constrained,
As formula (8) is usedInverse matrix go to find h,
Average minimum correlation energy wave filter MACE is developed on the basis of synthetic discriminant functions, can using MACE
Minimize the average correlation energy after training image related operation;
MACE wave filters are to calculate to obtain by following formula:
Wherein:
It is i-th of element of k-th training image, two-dimentional fractional power exponential filter can be described as follows:
Wherein:
X is created by training image, similar to formula (6),Element produce inside the image,It isConjugation,
In formula (14), it is referred to as Minimum Average Correlation Energy Filter if power P=2, P=0 formulas are referred to as to differentiate letter
Wavenumber filter, is referred to as fractional power exponential filter FPF when P is between 0-2;
Step 3:Result images are handled again using reaction type ICM, the image that only target image is present is finally obtained;
Based on this, realize that outline is recognized using one closed-loop system of reaction type ICM combinations FPF formation.
3. the outline recognition methods that a kind of reaction type ICM neutral nets and FPF are combined, it is characterised in that:First to input
Sketch figure picture carries out the image that ICM iteration produces edge excitation, then carries out edge enhancing to the image;Input training image is arrived
In FPF, the enhanced image in edge is obtained;At original image edge, enhanced result carries out related operation with FPF, obtains correlation
Peak value, modifies to the region of correlation peak maximum point, then using amended image as original image input system, finally
Obtain enhancing of the target image on original image.
4. the outline recognition methods that a kind of reaction type ICM neutral nets and FPF are combined according to claim 3, its feature
It is to comprise the following steps:
Step 1:Sketch figure picture to input carries out the image that ICM iteration produces edge excitation, then carries out edge increasing to the image
By force;
Step 2:Training image is inputted into FPF, the enhanced image in edge is obtained;
Step 3:The enhanced result in original image edge and FPF are subjected to related operation, correlation peak is obtained;
Step 4:The region of correlation peak maximum point is modified;
Step 5:Using amended image as original image input system, the enhancing on target image is finally obtained.
5. the outline identifying system that a kind of reaction type ICM neutral nets according to claims 1 to 2 and FPF are combined, is selected
Kimia outline data are tested, it is characterised in that method of testing is as follows:
(1) original images to be recognized and training image are inputted;
(2) it is 0.3, image of the original images to be recognized after ICM obtained peak related to the FPF after training to take FPF powers P
It is worth for 0.99;
(3) different peak value thresholds are set, the enhanced result of different feedbacks can be obtained;It can obtain, when threshold value thresholding is bigger
When, system identification effect is better;
(4) P=0.4 is taken, the correlation peak of generation is 0.1;
(5) increase threshold value thresholding can be obtained with identical result, the i.e. bigger recognition effect of threshold value thresholding in (3) more successively again
It is good;
(6) P value is increased successively, repeating the step in (3), can to obtain the bigger image border of P values more obvious, with original image production
Raw correlation peak is bigger, and recognition effect is better.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710414874.5A CN107330447B (en) | 2017-06-05 | 2017-06-05 | Silhouette recognition system combining feedback type ICM neural network and FPF |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710414874.5A CN107330447B (en) | 2017-06-05 | 2017-06-05 | Silhouette recognition system combining feedback type ICM neural network and FPF |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107330447A true CN107330447A (en) | 2017-11-07 |
CN107330447B CN107330447B (en) | 2020-04-24 |
Family
ID=60194939
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710414874.5A Active CN107330447B (en) | 2017-06-05 | 2017-06-05 | Silhouette recognition system combining feedback type ICM neural network and FPF |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107330447B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110659666A (en) * | 2019-08-06 | 2020-01-07 | 广东工业大学 | Image classification method of multilayer pulse neural network based on interaction |
CN112312190A (en) * | 2020-10-10 | 2021-02-02 | 游艺星际(北京)科技有限公司 | Video picture display method and device, electronic equipment and storage medium |
CN116342912A (en) * | 2023-05-30 | 2023-06-27 | 中国铁路设计集团有限公司 | Heterogeneous remote sensing image matching method and system based on correlation peak analysis |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101577003A (en) * | 2009-06-05 | 2009-11-11 | 北京航空航天大学 | Image segmenting method based on improvement of intersecting visual cortical model |
CN102651132A (en) * | 2012-04-06 | 2012-08-29 | 华中科技大学 | Medical image registration method based on intersecting cortical model |
CN103295201A (en) * | 2013-05-31 | 2013-09-11 | 中国人民武装警察部队工程大学 | Multi-sensor image fusion method on basis of IICM (improved intersecting cortical model) in NSST (nonsubsampled shearlet transform) domain |
-
2017
- 2017-06-05 CN CN201710414874.5A patent/CN107330447B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101577003A (en) * | 2009-06-05 | 2009-11-11 | 北京航空航天大学 | Image segmenting method based on improvement of intersecting visual cortical model |
CN102651132A (en) * | 2012-04-06 | 2012-08-29 | 华中科技大学 | Medical image registration method based on intersecting cortical model |
CN103295201A (en) * | 2013-05-31 | 2013-09-11 | 中国人民武装警察部队工程大学 | Multi-sensor image fusion method on basis of IICM (improved intersecting cortical model) in NSST (nonsubsampled shearlet transform) domain |
Non-Patent Citations (4)
Title |
---|
JASON M.KINSER: ""Parameter estimation using dual fractional power filters"", 《PRECEEDINGS OF SPIE》 * |
ULF EKBLAD 等: ""The intersectin cortical model in image processing"", 《NUCLEAR INSTRUMENTS & METHODS IN PHYSICS RESEARCH SECTION A》 * |
ULF EKBLAD 等: ""Theoretical foundation of the intersecting cortical model and its use for change detection of aircraft,cars,and nuclear explosion tests"", 《SIGNAL PRECESSING》 * |
薛峰: ""脉冲耦合神经网络与小数幂指数滤波器在数字识别中的应用研究"", 《万方数据知识服务平台》 * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110659666A (en) * | 2019-08-06 | 2020-01-07 | 广东工业大学 | Image classification method of multilayer pulse neural network based on interaction |
CN110659666B (en) * | 2019-08-06 | 2022-05-13 | 广东工业大学 | Image classification method of multilayer pulse neural network based on interaction |
CN112312190A (en) * | 2020-10-10 | 2021-02-02 | 游艺星际(北京)科技有限公司 | Video picture display method and device, electronic equipment and storage medium |
CN116342912A (en) * | 2023-05-30 | 2023-06-27 | 中国铁路设计集团有限公司 | Heterogeneous remote sensing image matching method and system based on correlation peak analysis |
CN116342912B (en) * | 2023-05-30 | 2023-08-11 | 中国铁路设计集团有限公司 | Heterogeneous remote sensing image matching method and system based on correlation peak analysis |
Also Published As
Publication number | Publication date |
---|---|
CN107330447B (en) | 2020-04-24 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106023220B (en) | A kind of vehicle appearance image of component dividing method based on deep learning | |
CN107194404B (en) | Underwater target feature extraction method based on convolutional neural network | |
CN104268593B (en) | The face identification method of many rarefaction representations under a kind of Small Sample Size | |
CN105138993B (en) | Establish the method and device of human face recognition model | |
Kadir et al. | Neural network application on foliage plant identification | |
CN107292250A (en) | A kind of gait recognition method based on deep neural network | |
CN106485259B (en) | A kind of image classification method based on high constraint high dispersive principal component analysis network | |
CN107103338A (en) | Merge the SAR target identification methods of convolution feature and the integrated learning machine that transfinites | |
Aquino et al. | The effect of data augmentation on the performance of convolutional neural networks | |
CN107330447A (en) | The outline identifying system that a kind of reaction type ICM neutral nets and FPF are combined | |
CN107704822A (en) | The extraction of finger vein further feature and matching process and system based on incomplete completion | |
Gabella | Topology of learning in artificial neural networks | |
Khavalko et al. | Image classification and recognition on the base of autoassociative neural network usage | |
CN107239827A (en) | A kind of spatial information learning method based on artificial neural network | |
Ruiz-Garcia et al. | Deep learning for illumination invariant facial expression recognition | |
Kim et al. | A code based fruit recognition method via image convertion using multiple features | |
CN106780468A (en) | View-based access control model perceives the conspicuousness detection method of positive feedback | |
Mathur et al. | The K-means Clustering based fuzzy edge detection technique on MRI images | |
Bisri et al. | Machine learning algorithms based on sampling techniques for raisin grains classification | |
Zhang et al. | An analysis of CNN feature extractor based on KL divergence | |
Diqi et al. | Implementation of CNN for plant leaf classification | |
Ignat et al. | Gender recognition with Gabor filters | |
Haselhoff et al. | An evolutionary optimized vehicle tracker in collaboration with a detection system | |
Farias et al. | Adaboost classification of TJ-II Thomson Scattering images | |
Raghavendra et al. | Features subset selection using improved teaching learning based optimisation (ITLBO) algorithms for IRIS recognition |
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: 20231103 Address after: No. 57-5 Development Avenue, No. 6015, Yichang Area, China (Hubei) Free Trade Zone, Yichang City, Hubei Province, 443005 Patentee after: Hubei Jiugan Technology Co.,Ltd. Address before: 443002 No. 8, University Road, Yichang, Hubei Patentee before: CHINA THREE GORGES University |
|
TR01 | Transfer of patent right |