CN101976345B - Method for recognizing image scale invariant pattern under noise condition - Google Patents
Method for recognizing image scale invariant pattern under noise condition Download PDFInfo
- Publication number
- CN101976345B CN101976345B CN 201010297827 CN201010297827A CN101976345B CN 101976345 B CN101976345 B CN 101976345B CN 201010297827 CN201010297827 CN 201010297827 CN 201010297827 A CN201010297827 A CN 201010297827A CN 101976345 B CN101976345 B CN 101976345B
- Authority
- CN
- China
- Prior art keywords
- image
- noise
- correlation
- intrinsic mode
- radially
- 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.)
- Expired - Fee Related
Links
Images
Abstract
The invention discloses a method for recognizing an image scale invariant pattern under the noise condition. The method comprises two separated stages: a training stage of training the image of a specific object, namely a reference image, to obtain a correlation filter, and a correlation detection stage of recognizing the object to be detected by using the obtained correlation filter. In the method, noise estimation can change adaptively with the change of the signal to noise ratio, and high noise immunity is achieved. The main principle of the method is that: the image can be reconstructed by an intrinsic mode function correlated to the important structure of the image, thus the correlation filter can be built by some specific layers (the intrinsic mode function) to realize the scale invariant pattern recognition; moreover, as the noise distribution on the intrinsic mode function of each layer has a certain rule, the aim of eliminating noise can be fulfilled by the noise distribution rule.
Description
Technical field
What the present invention relates to is a kind of image processing method.A kind of method of utilizing computing machine in the image under the noise conditions, to detect, identify the different scale object particularly.
Background technology
A fundamental purpose of pattern-recognition is exactly the method (for example to input object insensitive detection method that distorts) that research can be carried out distortion invariant recognition.During image was processed, three very typical object distortion were translation, rotation and dimensional variation.People have carried out a large amount of research for these three kinds of distortion, have proposed some and have obtained the method for the constant detection of distortion.These method great majority are based on the thought of correlation filter.The translation of objects in images can utilize Fourier spectrum to detect easily.Objects in images rotates in the plane, and then its Fourier spectrum also can produce corresponding rotation and be take 2 π as the cycle, so can utilize circular harmonic function (circular resonant function) to detect.And the dimensional variation of objects in images, owing to do not have obvious feature (for example as the periodicity of rotating), so the constant detection of the yardstick of object has larger difficulty.
The constant related detecting method of yardstick based on correlation filter is the main research that carries out the constant detection of object yardstick, mainly comprises the radially mediation wave filter (Radialharmonic filter) based on plum forests RadialHarmonyFunction (Mellin radial harmonic function), only phase place radially is in harmonious proportion wave filter (Phase-only radial harmonic filter), only phase place is in harmonious proportion wave filter (Phase-only logarithmic radial harmonic filter) to number redial, the mark energy radially is in harmonious proportion wave filter (fractionalpower radial harmonic filter) etc.But the correlation peak of these method outputs is subjected to the impact of the dimensional variation factor also unstable, and very responsive to ground unrest, and the detection performance of these methods is very restricted.Radially mediation wave filter based on wavelet transformation can remove a part of high frequency noise, but its disadvantage is to specify in advance wavelet basis function, and different basis functions causes different testing results.
Summary of the invention
Estimation can adaptive change has the noise conditions hypograph yardstick invariant Pattern Recognition method of fine noise immunity with the variation of noise signal to noise ratio (S/N ratio) to the object of the present invention is to provide a kind of noise.
The object of the present invention is achieved like this:
Comprise that for the certain objects image be the training stage that reference picture training obtains correlation filter, utilize the stage of coherent detection stage two separation that the correlation filter that obtains identifies examining object;
(1) the described training stage comprises following steps:
1) utilizes empirical mode decomposition (the bidimensional empirical mode decomposition:BEMD) technology to decompose certain objects image f (x, y), obtain each layer intrinsic mode function F of object
i(intrinsic mode functions:IMF) and remaining components R;
2) be high fdrequency component F with ground floor IMF
1, carry out plum forests and radially be in harmonious proportion decomposition, obtain radially harmonic component;
3) choose radially harmonic component of certain single order, construct radially harmonic correlation wave filter;
(2) the described coherent detection stage comprises following steps:
1) inquery image g (x, y) is carried out empirical mode decomposition, the intrinsic mode function F after obtaining decomposing
i
The energy of the intrinsic mode functions at different levels of 2) asking, and come the content of noise in the estimated image with this;
3) try to achieve intrinsic mode function energy minimum the layer j;
4) remove intrinsic mode function less than j, utilize remaining intrinsic mode function to come reconstructed image;
5) utilize Laplce's Gauss method to ask the edge Edge of reconstructed image (Resid (g));
6) utilize radially that harmonic correlation wave filter and following formula carry out associative operation, obtain two-dimensional correlation output;
7) utilize relevant output peak value to carry out the judgement of yardstick invariant Pattern Recognition, take from correlation peak 60% as decision threshold, if cross-correlation peak value is greater than 60% of the autocorrelation peak of reference picture, then image g (x to be checked, y) be the dimensional variation version of known image f (x, y); Otherwise image g (x, y) is not the dimensional variation version of image f (x, y).
The present invention utilizes the high fdrequency component of object (image model) to construct correlation filter and detect pattern to be checked.Comprise a noise estimation procedure, can automatically select according to noisy amount (signal to noise ratio (S/N ratio)) in the image to be checked the number of the intrinsic mode function of removal, can adapt to noise and change.
Principle of the present invention is described as follows:
If image f (x, y) and its Fourier transform are W (ω
1, ω
2), then to look like be that g (x, y)=f (α x, α y) and its corresponding Fourier transform are to its dimensional variation domain
Scalogram as associative operation in, the contribution of low-frequency component is very little, medium-high frequency plays a major role.Similar with wavelet transformation, empirical mode decomposition can be regarded a bank of filters as, and each rank intrinsic mode function of its generation is the process of launching gradually from the high frequency to the low frequency, be that the ground floor intrinsic mode function comprises the high-frequency information in the image, last one deck intrinsic mode function comprises the low-limit frequency information in the image.Therefore the ground floor intrinsic mode function of f (x, y) is done radially mediation and construct radially harmonic correlation wave filter after decomposing.When detecting, get equally the ground floor intrinsic mode function of image to be checked as the input of correlation filter, utilize the intensity of the relevant peaks on the output plane that correlation filter produces and the position of relevant peaks, just can detect through the reference picture pattern of dimensional variation and the position in detected image.But generally speaking, usually all contain noise in the middle of the image to be checked, and radially the harmonic correlation wave filter is responsive to noise, therefore must carry out noise reduction process.Normally do not have noisy image, from the low frequency to the high frequency, energy reduces gradually generally speaking; The white Gaussian noise signal energy distributes just in time opposite, and the energy from the low frequency to the high frequency increases gradually.Therefore, if comprised noise in the image to be checked, can utilize the result of empirical mode decomposition, in the distribution of each mode energy, seek some j (the mode layer j that undergos mutation that energy is undergone mutation, layer noise contribution less than j is dominant, and is dominant greater than the tomographic image composition of j).In original image, deduct the layer that noise is dominant, the layer reconstructed image that utilizes iconic element to be dominant, just can reach the effect of denoising, and the number of plies of denoising relevant with the noisy amount in the image (affected by signal to noise ratio (S/N ratio)), this denoising process is a kind of adaptive process.After noisy image denoising reconstruct to be checked, utilize marginal information (radio-frequency component), as radially being in harmonious proportion the input of wave filter, the two-dimensional correlation of asking output utilizes relevant output peak value just can detect the image (pattern) of dimensional variation.
The present invention is the graphical rule invariant Pattern Recognition method under a kind of noise conditions.Overcome existing based on being in harmonious proportion radially that the relevant peaks of yardstick invariant Pattern Recognition algorithm output of wave filter is subjected to the impact of scale factor and to the susceptibility of noise.Estimation can adaptive change has fine noise immunity to the noise of the method with the variation of noise signal to noise ratio (S/N ratio).The principle of this method mainly is to utilize the intrinsic mode function relevant with the important structure of image to be reconstructed image, therefore can utilize some certain layer (intrinsic mode function) structure correlation filter to reach the yardstick invariant Pattern Recognition; Therefore and the distribution of noise on each layer intrinsic mode function has certain rule, can utilize this regularity of distribution of noise to reach the purpose of denoising.
Description of drawings
Fig. 1 is training process: produce the radially process flow diagram of harmonic correlation wave filter.
Fig. 2 is testing process: yardstick invariant Pattern Recognition method detects the process flow diagram of inquery image.
Fig. 3 is the process flow diagram of determining the noisy amount of inquery image and definite denoising number of plies and Image Reconstruction.
Embodiment
Below in conjunction with accompanying drawing the present invention is done more detailed description:
This method comprises the stage of two separation: (1) training obtains correlation filter for certain objects image (reference picture); (2) utilize the correlation filter that obtains that examining object is identified.
Such as Fig. 1, the training stage comprises following steps:
1) utilize empirical mode decomposition (the bidimensional empirical mode decomposition:BEMD) technology to decompose specific object f (x, y) (be commonly referred to reference picture), obtain each layer intrinsic mode function F of object
i(intrinsicmode functions:IMF) and remaining components R;
2) with ground floor IMF (high fdrequency component) F
1, carry out plum forests and radially be in harmonious proportion decomposition, obtain radially harmonic component;
(x in the formula
0, y
0) be the center of decomposing that radially is in harmonious proportion, the rank of m for decomposing.
3) choose radially harmonic component of certain single order M, construct radially harmonic correlation wave filter radial{F
1(f) }.
h(r,θ)=f
M(θ)r
i2πM-1
radial{F
1(f)}=h(r,θ)
Such as Fig. 2 and Fig. 3, the coherent detection stage comprises following steps:
1) inquery image g (x, y) is carried out empirical mode decomposition, the intrinsic mode function F after obtaining decomposing
i
The energy of the intrinsic mode functions at different levels of 2) asking, and come the content of noise in the estimated image with this;
Picture size is p * q.
3) try to achieve intrinsic mode function energy minimum the layer j;
4) remove intrinsic mode function less than j, utilize remaining intrinsic mode function to come reconstructed image;
5) utilize Laplce's Gauss method to ask the edge Edge of reconstructed image (Resid (g))
6) utilize and radially to be in harmonious proportion wave filter and to carry out associative operation with following formula, obtain two-dimensional correlation and export
7) utilize relevant output peak value to carry out the judgement of yardstick invariant Pattern Recognition, take from generally speaking correlation peak | C
Ff| 60% (be 0.6*|C
Ff|) as decision threshold.If cross-correlation peak value | C
Gf|>0.6*|C
Ff|, then image g to be checked (x, y) is the dimensional variation version of known image f (x, y); Otherwise image g (x, y) is not the dimensional variation version of image f (x, y).
The above is a kind of enforcement of the algorithm that proposes of the present invention, but on some step, can carry out appropriate change, to adapt to the demand of concrete condition.For example, in the step 3 of training stage) when constructing radially the harmonic correlation wave filter, can as required, carry out suitable adjustment to choosing of harmonic component M radially.For example, in coherent detection stage step 4) when utilizing the intrinsic mode function reconstructed image, the number of plies j that abandons can suitably adjust (as being adjusted into j=j-1 or j=j+1).
Claims (1)
1. noise conditions hypograph yardstick invariant Pattern Recognition method, comprise that for the certain objects image be the training stage that reference picture training obtains correlation filter, utilize the stage of coherent detection stage two separation that the correlation filter that obtains identifies examining object; It is characterized in that:
(1) the described training stage comprises following steps:
1) utilizes Empirical mode decomposition to decompose certain objects image f (x, y), obtain each layer intrinsic mode function F of object
iWith remaining components R,
2) be high fdrequency component F with ground floor IMF
1, carry out plum forests and radially be in harmonious proportion decomposition, obtain radially harmonic component,
3) choose radially harmonic component of certain single order M, construct radially harmonic correlation wave filter radial{F
1(f) },
h(r,θ)=f
M(θ)r
i2πM-1
radial{F
1(f)}=h(r,θ);
(2) the described coherent detection stage comprises following steps:
1) inquery image g (x, y) is carried out empirical mode decomposition, the intrinsic mode function F after obtaining decomposing
i,
The energy of the intrinsic mode functions at different levels of 2) asking, and come the content of noise in the estimated image with this,
Picture size is p * q;
4) remove intrinsic mode function less than j, utilize remaining intrinsic mode function to come reconstructed image,
5) utilize Laplce's Gauss method to ask the edge Edge of reconstructed image (Resid (g));
6) utilize radially that harmonic correlation wave filter and Edge (Resid (g)) carry out associative operation, obtain two-dimensional correlation output,
Symbol
Represent associative operation;
7) utilize relevant output peak value to carry out the judgement of yardstick invariant Pattern Recognition, take from correlation peak 60% as decision threshold, if cross-correlation peak value is greater than 60% of the autocorrelation peak of reference picture, then image g (x to be checked, y) be the dimensional variation version of known image f (x, y); Otherwise image g (x, y) is not the dimensional variation version of image f (x, y).
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN 201010297827 CN101976345B (en) | 2010-09-30 | 2010-09-30 | Method for recognizing image scale invariant pattern under noise condition |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN 201010297827 CN101976345B (en) | 2010-09-30 | 2010-09-30 | Method for recognizing image scale invariant pattern under noise condition |
Publications (2)
Publication Number | Publication Date |
---|---|
CN101976345A CN101976345A (en) | 2011-02-16 |
CN101976345B true CN101976345B (en) | 2013-02-06 |
Family
ID=43576229
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN 201010297827 Expired - Fee Related CN101976345B (en) | 2010-09-30 | 2010-09-30 | Method for recognizing image scale invariant pattern under noise condition |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN101976345B (en) |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101251898A (en) * | 2008-03-25 | 2008-08-27 | 腾讯科技(深圳)有限公司 | Skin color detection method and apparatus |
CN101398893A (en) * | 2008-10-10 | 2009-04-01 | 北京科技大学 | Adaboost arithmetic improved robust human ear detection method |
CN101655910A (en) * | 2008-08-21 | 2010-02-24 | 索尼(中国)有限公司 | Training system, training method and detection method |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7519201B2 (en) * | 2005-10-28 | 2009-04-14 | Honda Motor Co., Ltd. | Detecting humans via their pose |
US7778446B2 (en) * | 2006-12-06 | 2010-08-17 | Honda Motor Co., Ltd | Fast human pose estimation using appearance and motion via multi-dimensional boosting regression |
-
2010
- 2010-09-30 CN CN 201010297827 patent/CN101976345B/en not_active Expired - Fee Related
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101251898A (en) * | 2008-03-25 | 2008-08-27 | 腾讯科技(深圳)有限公司 | Skin color detection method and apparatus |
CN101655910A (en) * | 2008-08-21 | 2010-02-24 | 索尼(中国)有限公司 | Training system, training method and detection method |
CN101398893A (en) * | 2008-10-10 | 2009-04-01 | 北京科技大学 | Adaboost arithmetic improved robust human ear detection method |
Also Published As
Publication number | Publication date |
---|---|
CN101976345A (en) | 2011-02-16 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN104765979B (en) | A kind of sea clutter denoising method based on integrated empirical mode decomposition | |
CN102928517A (en) | Method for denoising acoustic testing data of porcelain insulator vibration based on wavelet decomposition threshold denoising | |
CN105913382B (en) | The high-fidelity anisotropic filtering method of threshold value optimizing | |
CN109871733A (en) | A kind of adaptive sea clutter signal antinoise method | |
CN106568589A (en) | Rubbing acoustic emission denoise method based on empirical wavelet transform | |
Zheng et al. | A novel edge detection for buried target extraction after SVD-2D wavelet processing | |
Yao et al. | An adaptive seismic signal denoising method based on variational mode decomposition | |
Vijay et al. | Evaluation of effectiveness of wavelet based denoising schemes using ANN and SVM for bearing condition classification | |
Kong et al. | Adaptive TQWT filter based feature extraction method and its application to detection of repetitive transients | |
CN103985092A (en) | Post-processing noise elimination method for performing ICA analysis of plural f MRI data | |
Hu et al. | Low quality fingerprint image enhancement based on Gabor filter | |
CN101976345B (en) | Method for recognizing image scale invariant pattern under noise condition | |
Li et al. | Magnetotelluric signal-noise separation method based on SVM–CEEMDWT | |
CN102368332B (en) | SAR (Synthetic Aperture Radar) image despeckling method based on local Gaussian model in NSCT (Non-subsampled Contourlet Transform) region | |
Zhou et al. | Sparse dictionary analysis via structure frequency response spectrum model for weak bearing fault diagnosis | |
CN104657983A (en) | Method for detecting densities of fabric hairballs based on Gabor filtering | |
Butt et al. | Multilateral filtering: A novel framework for generic similarity-based image denoising | |
Guo et al. | An Image Denoising Algorithm based on Kuwahara Filter | |
Jain et al. | Performance comparison of two image denoising algorithm at different noises | |
Guo et al. | Canny edge detection algorithm based on wavelet transform and RAMF | |
CN109117780A (en) | Noise cancellation method and system | |
Cheng et al. | Image denoising based on hybrid fourier and neighborhood wavelet coefficients | |
Lixia et al. | Analysis and calculation of wavelet correlation filtering denoise | |
CN113910005B (en) | Grinding system and signal processing method and device thereof | |
Clapson et al. | A gravitational wave burst search method based on the S transform |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
C14 | Grant of patent or utility model | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20130206 Termination date: 20180930 |