CN106780398B - A kind of image de-noising method based on noise prediction - Google Patents
A kind of image de-noising method based on noise prediction Download PDFInfo
- Publication number
- CN106780398B CN106780398B CN201710016951.1A CN201710016951A CN106780398B CN 106780398 B CN106780398 B CN 106780398B CN 201710016951 A CN201710016951 A CN 201710016951A CN 106780398 B CN106780398 B CN 106780398B
- Authority
- CN
- China
- Prior art keywords
- image
- noise
- information
- redundancy
- image sequence
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 28
- 238000004458 analytical method Methods 0.000 claims abstract description 6
- 230000006835 compression Effects 0.000 claims description 18
- 238000007906 compression Methods 0.000 claims description 18
- 230000009466 transformation Effects 0.000 claims description 3
- 239000000470 constituent Substances 0.000 claims description 2
- 238000010276 construction Methods 0.000 claims 1
- 238000012880 independent component analysis Methods 0.000 abstract description 4
- 230000009467 reduction Effects 0.000 abstract description 3
- 239000011159 matrix material Substances 0.000 description 8
- 238000000926 separation method Methods 0.000 description 3
- 238000009499 grossing Methods 0.000 description 2
- 238000003384 imaging method Methods 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 238000011946 reduction process Methods 0.000 description 2
- 230000000717 retained effect Effects 0.000 description 2
- 230000008859 change Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 150000003839 salts Chemical class 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 230000017105 transposition Effects 0.000 description 1
Classifications
-
- G06T5/70—
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
Abstract
The invention discloses a kind of image de-noising methods based on noise prediction, belong to the technical field of image procossing.The present invention passes through unnecessary redundancy in KSVD algorithm removal original image first, then non-average value processing is carried out to image redundancy information remaining after KSVD algorithm process by neighborhood window, avoid long-range prediction noise bring inaccuracy, the image after removing prediction noise is enabled to retain detailed information to a greater extent, it is again the observed image of independent composition analysis algorithm with the difference of image after original image and non-average value processing, the useful information component and noise component(s) for ensuring to isolate by independent component analysis are mutually indepedent, retain the useful information of image as far as possible while denoising, image after guaranteeing noise reduction is optimal.
Description
Technical field
The invention discloses a kind of image de-noising methods based on noise prediction, belong to the technical field of image procossing.
Background technique
With the development of the informationization technology based on computer network, requirement of the people to digital picture quality is more next
Higher, the noises such as thermal noise, shot noise and quantum noise of image output equipment and the generation of some photoelectric devices can seriously drop
The quality of the image of low final output.Digital picture noise can be divided into internal noise and external noise according to Producing reason,
Gaussian noise and Rayleigh noise are defined and can be divided into according to the shape that noise amplitude is distributed at any time, and Gaussian noise and the spiced salt are made an uproar
Sound is two kinds of noises common in digital picture.
Traditional denoising mode puts on an equal footing noise and useful information, it is believed that noise and image detail information are in some change
It changes in domain and is distributed in different sections, however, noise and certain useful information images are generally overlapped mutually in some region,
Useful detailed information can also be removed while removing noise, different denoising means can only specific noise signal range it
Interior effective, off-limits noisy image can large area distortion by processing.
Blind source separating is a kind of processing technique based on thermal signals such as statistics, and initial classic applications are " cocktail party ".
Blind source separating being gradually applied to image co-registration because can analyze multidimensional data, image enhancement, feature extraction, artifact elimination, mixing
In terms of closing image, eliminating the image procossings such as line at random, and image noise reduction needs are carried out by the priori of noise using blind source separating
Knowledge carries out sparse coding compression by way of the noisy acoustic image of training.
Summary of the invention
Goal of the invention of the invention is the deficiency for above-mentioned background technique, provides a kind of image based on noise prediction
Denoising method remains image detail information to the maximum extent and solves blind source by the gradually noise reduction process to original image
Separation algorithm must have the technical issues of multiple image signal sources.
The present invention adopts the technical scheme that for achieving the above object
The invention patent is to carry out the image of certain noise prediction in a kind of slave noisy image based on blind source separating to go
Method for de-noising needs to carry out a part of denoising to this image when gained image only has one, and first have to carry out is pressure
Contracting obtains the redundancy of image by compression, carries out non-average value processing to redundancy and obtain the different image of noise intensity,
This step is repeated, the image of multiple available noises containing varying strength carries out blind source point further according to independent composition analysis algorithm
From denoising image is obtained, does not need to shoot photo in large quantities and detailed information can be retained to the maximum extent.
This method is specifically: carrying out non-homogeneous processing to the redundancy through KSVD dictionary compression and obtains containing varying strength
The image of Gaussian noise, raw noise image subtract the image containing varying strength Gaussian noise and obtain a preliminary denoising figure
Picture, then blind source separating is carried out to preliminary denoising image, the specific method is as follows:
1, sparse compression raw noise image
The raw noise image that shooting obtains is handled using KSVD compression algorithm, that is, meets the contractible iteration condition of KSVD,
The objective function of KSVD algorithm indicates are as follows:
Wherein, D ∈ Rn×K y∈Rn x∈RK, D, Y, X respectively represent the rarefaction representation of dictionary, training signal, training signal
Vector, D ∈ Rn×K, Y ∈ Rn, X ∈ RK, Rn×KFor the vector space of n × K, RnFor n-dimensional vector space, RKFor K dimensional vector space,For the set of N number of training signal,For the solution vector set of Y, T0For zero non-in rarefaction representation coefficient
The upper limit of the number of amount, i.e., the maximum difference degree in coefficient vector.
After iterated transform, many noise informations in raw noise image are rejected together, and what is remained is superfluous
Remaining information but contains many image detail informations, that is, the noisy acoustic intelligence of the redundancy information crossed through KSVD compression processing and
A large amount of image detail information, non-average value processing is to these further rarefactions of excessive redundancy;
2, non-average value processing redundancy
Search window is using D pixel as window size, and neighborhood window is using d pixel as window size, due to KSVD compression processing
The redundancy generated afterwards also retains a large amount of image detail information while remaining partial information, according between neighborhood
Similitude determines the weight of pixel to estimate a relatively small noise signalBy between two neighborhood windows of calculating
Similarity degree is that pixel y is assigned to weightThe prediction noise of pixel xAre as follows:
Wherein, weightIndicate the similarity between pixel x and pixel y, its value is by with pixel x, pixel
Distance between rectangular neighborhood V (x) centered on point y, V (y) | | V (x)-V (y) | |2It determines:
Wherein,
Z (x) is normalization coefficient, and h is the smoothing parameter for controlling Gaussian function attenuation degree, and d is the size of neighborhood window,
Ds is transform domain, repeats the non-available image sequence being made of the image of multiple noises containing varying strength of average value processing, former
Beginning noise image X' and prediction noiseMake the difference obtain preliminary denoising image
3, blind source separating tentatively denoises image
Just become new observed image, according to blind source separating:
Subscript i is order of every frame observed image in image sequence, and the value 1 ..., m of i, m is observed image
Totalframes, aijIt is component mixed coefficint, S is indicated by picture content sjThe matrix of composition, S=[s1,s2,...,sj...,sn]T, under
Marking j indicates component serial number, the value 1 ..., n of j, and n is component sum;M, n is integer, and m >=n >=2, T representing matrix transposition;
Separation matrix W is obtained by fixed point independent component analysis FastICA, to obtain unlimited approximation component SjPoint
Measure yi, relational expression indicates as follows:
Y=WT=WAS → S (8)
In formula (8), Y indicates the matrix being made of the component isolated, Y=[y1,y2,...yn]T, yiIndicate image information
Component.
Since non-average value processing is carried out by neighborhood window, it is ensured that will not be made because of long-range prediction noise
At inaccuracy is calculated, the image after removal prediction noise signal is allowed to retain detailed information to a greater extent, to pass through
The difference of picture and raw noise image after non-average value processing, which carries out blind source separating as observed image, can obtain optimal go
It makes an uproar image.
For the image sequence of imaging device shooting, the present invention proposes a kind of image de-noising method based on noise prediction,
Gray proces are carried out to raw noise image sequence, the raw noise image sequence after gray proces is carried out by KSVD algorithm
Dictionary compression obtain it is sparse after redundant image sequence, to redundancy image sequence carry out neighborhood in similarity transformation obtain newly
Noisy image sequence, using in redundancy image noise information and image useful information as isolated component carry out blind source point
From the standard variance or variance of each isolated component being separated being counted, according to isolated component standard variance or the spy of variance
Property filters out image useful information isolated component.
Present invention employs above-mentioned technical proposal, have the advantages that first by KSVD remove image in need not
The redundancy wanted then carries out non-average value processing to remaining image redundancy information after KSVD processing by neighborhood window, keeps away
Long-range prediction noise bring inaccuracy is exempted from, the image after removing prediction noise is enabled to retain details to a greater extent
Information, then with after original image and non-average value processing image difference be independent composition analysis algorithm observed image, pass through independence
The useful information component and noise component(s) that constituent analysis ensures to isolate are mutually indepedent, retain image as far as possible while denoising
Useful information, the image after guaranteeing noise reduction are optimal.
Detailed description of the invention
Fig. 1 is the flow chart of image de-noising method of the present invention.
Specific embodiment
The technical solution of invention is described in detail with reference to the accompanying drawing.
An image sequence is shot using imaging device, captured target is in relative static conditions;Later by pair
Obtained gray level image carries out KSVD compression, retains image information to the maximum extent and obtains having comprising noise information and image information
With the redundancy image sequence of information, similarity transformation in neighborhood is carried out to redundancy image sequence and handles to obtain new contain
It makes an uproar image sequence;New noisy image sequence is divided using the independent component analysis method of higher order statistical in blind source separating
From, the standard variance (Standard Deviation, SD) or variance (Vanriation) of each component being separated are calculated,
Being worth the greater is required image useful information component, and the standard variance of multiple noise component(s)s is closer to each other, but relative to
The standard variance of image useful information is smaller.
Picture noise information and image useful information are considered as to the isolated component of composition image, different noise images is combined and obtains
The lesser standard variance (Standard Deviation, SD) or variance (Vanriation) arrived is roughly the same, and larger mark
Quasi- variance or the corresponding isolated component of variance indicate image information, export larger standard variance or the corresponding isolated component institute of variance
The image of composition, i.e. completion noise reduction process.
When image only has one, need to carry out a part of denoising to this image, first have to carry out is compression,
The redundancy of image is obtained by compression, non-average value processing is carried out to redundancy and obtains the different image of noise intensity, with
This analogizes, the image of multiple available noises containing varying strength, carries out blind source separating further according to independent composition analysis algorithm and obtains
To denoising image.It does not need largely to shoot photo and detailed information can be retained to greatest extent.
Its method carries out non-homogeneous processing to the redundancy of KSVD dictionary compression as shown in Figure 1: and obtains containing different strong
The image of Gaussian noise is spent, raw noise image subtracts the image containing varying strength Gaussian noise and obtains a preliminary denoising figure
Picture, then blind source separating is carried out to preliminary denoising image, the specific method is as follows:
1, sparse compression raw noise image
The raw noise image that shooting obtains is handled using KSVD compression algorithm, that is, meets the contractible iteration condition of KSVD,
The objective function of KSVD algorithm indicates are as follows:
Wherein, D ∈ Rn×K y∈Rn x∈RK, D, Y, X respectively represent the rarefaction representation of dictionary, training signal, training signal
Vector, D ∈ Rn×K, Y ∈ Rn, X ∈ RK, Rn×KFor the vector space of n × K, RnFor n-dimensional vector space, RKFor K dimensional vector space,For the set of N number of training signal,For the solution vector set of Y, T0For zero non-in rarefaction representation coefficient
The upper limit of the number of amount, i.e., the maximum difference degree in coefficient vector.
After iterated transform, many noise informations are rejected together in raw noise image, the redundancy remained
Information but contains the detailed information of many images, that is, the noisy acoustic intelligence of the redundancy information crossed through KSVD compression processing and
A large amount of image detail information, non-average value processing is to these further rarefactions of excessive redundancy.
2, non-average value processing redundancy
Search window is using D pixel as window size, and neighborhood window is using d pixel as window size, due to KSVD compression processing
The redundancy generated afterwards also retains a large amount of image detail information while remaining partial noise information, according to neighborhood
Between similitude determine the weight of pixel to estimate a relatively small noise signal.By calculating two neighborhood windows
Similarity degree between mouthful is that pixel y is assigned to weightThe prediction noise of pixel xAre as follows:
Wherein, weightIndicate the similarity between pixel x and pixel y, its value is by with pixel x, pixel
Distance between rectangular neighborhood V (x) centered on point y, V (y) | | V (x)-V (y) | |2It determines:
Wherein,
Z (x) is normalization coefficient, and h is the smoothing parameter for controlling Gaussian function attenuation degree, and d is the size of neighborhood window,
Ds is transform domain, repeats the non-available image sequence being made of the image of multiple noises containing varying strength of average value processing, former
Beginning noise image X' and prediction noiseMake the difference obtain preliminary denoising image
3, blind source separating tentatively denoises image
Just become new observed image, according to blind source separating:
Subscript i is order of every frame observed image in image sequence, and the value 1 ..., m of i, m is observed image
Totalframes, aijIt is component mixed coefficint, S is indicated by picture content sjThe matrix of composition, S=[s1,s2,...,sj,...,sn]T,
Subscript j indicates component serial number, the value 1 ..., n of j, and n is component sum;M, n is integer, and m >=n >=2, T representing matrix turn
It sets;
Separation matrix W is obtained by fixed point independent component analysis FastICA, to obtain unlimited approximation component SjPoint
Measure yi, relational expression indicates as follows:
Y=WT=WAS → S (8)
In formula (8), Y indicates the matrix being made of the component isolated, Y=[y1,y2,...,yi,...,yn]T, yiIt indicates
Image information component.
Since non-average value processing is carried out by neighborhood window, it is ensured that will not be made because of long-range prediction noise
At the inaccuracy of calculating, the image after removal prediction noise signal is allowed to retain detailed information to a greater extent, with warp
The difference of picture and raw noise image after crossing non-average value processing can obtain optimal as observed image progress blind source separating
Denoise image.
Claims (6)
1. a kind of image de-noising method based on noise prediction, which is characterized in that compressed to obtain to individual original image superfluous
Remaining information carries out heterogeneous processing to redundancy and obtains the image of multiple noise informations containing varying strength, to raw noise figure
The image of picture and each noise information containing varying strength takes difference to obtain a series of preliminary denoising images, a series of preliminary denoisings
Image construction observed image sequence carries out blind source separating to observed image sequence and obtains denoising image to the end.
2. a kind of image de-noising method based on noise prediction according to claim 1, which is characterized in that calculated by KSVD
Method to raw noise image carry out dictionary compression obtain it is sparse after redundancy, the redundancy contains noise information and figure
As detailed information.
3. a kind of image de-noising method based on noise prediction according to claim 1, which is characterized in that only using fixed point
Vertical constituent analysis FastICA algorithm carries out blind source separating to observed image sequence and obtains denoising image to the end.
4. a kind of image de-noising method based on noise prediction, which is characterized in that compress to raw noise image sequence
To redundancy image sequence, similarity transformation in neighborhood is carried out to redundancy image sequence and obtains new noisy image sequence
Column, are separated from new noisy image sequence using the method for higher order statistical isolated component standard variance or variance in blind source separating
Image information out.
5. a kind of image de-noising method based on noise prediction according to claim 4, which is characterized in that utilize blind source separating
The method of middle higher order statistical isolated component standard variance or variance isolates image information from new noisy image sequence, specifically
Method are as follows: using in redundancy image noise information and image useful information as isolated component carry out blind source separating, count quilt
The standard variance or variance for each isolated component isolated, the characteristic according to isolated component standard variance or variance filter out figure
As useful information isolated component.
6. a kind of image de-noising method based on noise prediction according to claim 4, which is characterized in that raw noise figure
As sequence is compressed the method for obtaining redundancy image sequence are as follows: firstly, being carried out at gray scale to raw noise image sequence
Reason, then, by KSVD algorithm to after gray proces raw noise image sequence carry out dictionary compression obtain it is sparse after it is superfluous
Remaining image sequence.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710016951.1A CN106780398B (en) | 2017-01-10 | 2017-01-10 | A kind of image de-noising method based on noise prediction |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710016951.1A CN106780398B (en) | 2017-01-10 | 2017-01-10 | A kind of image de-noising method based on noise prediction |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106780398A CN106780398A (en) | 2017-05-31 |
CN106780398B true CN106780398B (en) | 2019-07-23 |
Family
ID=58948846
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710016951.1A Active CN106780398B (en) | 2017-01-10 | 2017-01-10 | A kind of image de-noising method based on noise prediction |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106780398B (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111428596A (en) * | 2020-03-16 | 2020-07-17 | 重庆邮电大学 | Grinding sound signal detection method based on three sound pickups |
Family Cites Families (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7474756B2 (en) * | 2002-12-18 | 2009-01-06 | Siemens Corporate Research, Inc. | System and method for non-square blind source separation under coherent noise by beamforming and time-frequency masking |
TWI397057B (en) * | 2009-08-03 | 2013-05-21 | Univ Nat Chiao Tung | Audio-separating apparatus and operation method thereof |
CN103400358A (en) * | 2013-08-27 | 2013-11-20 | 中国兵器工业导航与控制技术研究所 | Wavelet information redundancy based infrared image denoising algorithm |
CN106204461B (en) * | 2015-05-04 | 2019-03-05 | 南京邮电大学 | In conjunction with the compound regularized image denoising method of non local priori |
CN106251315B (en) * | 2016-08-23 | 2018-12-18 | 南京邮电大学 | A kind of image de-noising method based on full variation |
-
2017
- 2017-01-10 CN CN201710016951.1A patent/CN106780398B/en active Active
Also Published As
Publication number | Publication date |
---|---|
CN106780398A (en) | 2017-05-31 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Gnanadurai et al. | An efficient adaptive thresholding technique for wavelet based image denoising | |
CN110992292B (en) | Enhanced low-rank sparse decomposition model medical CT image denoising method | |
Li et al. | Multifocus image fusion using wavelet-domain-based deep CNN | |
CN109685730B (en) | Wavelet denoising method based on self-adaptive non-local mean value | |
CN110930327B (en) | Video denoising method based on cascade depth residual error network | |
CN110533614B (en) | Underwater image enhancement method combining frequency domain and airspace | |
CN107292316B (en) | Sparse representation-based method for improving image definition | |
Gopinathan et al. | Wavelet and FFT Based Image Denoising Using Non-Linear Filters. | |
CN107133590A (en) | A kind of identification system based on facial image | |
CN111915518B (en) | Hyperspectral image denoising method based on triple low-rank model | |
CN104135664A (en) | Method for digital processing of medical image | |
CN109003247B (en) | Method for removing color image mixed noise | |
CN106780398B (en) | A kind of image de-noising method based on noise prediction | |
CN110349101A (en) | A kind of multiple dimensioned united Wavelet image denoising method | |
CN111353982B (en) | Depth camera image sequence screening method and device | |
Alsayyh et al. | A Novel Fused Image Compression Technique Using DFT, DWT, and DCT. | |
CN110163489B (en) | Method for evaluating rehabilitation exercise effect | |
Abbas et al. | Face Recognition using DWT with HMM | |
Budhiraja et al. | Effect of pre-processing on MST based infrared and visible image fusion | |
CN110490819A (en) | Image de-noising method based on NSST and NLM filtering and hard -threshold technology | |
Li et al. | Remote sensing image enhancement based on adaptive thresholding in NSCT domain | |
Zhang et al. | Detection of foreign fibers in cotton on the basis of wavelet | |
Kanchana et al. | Texture classification using discrete shearlet transform | |
Han | Artificial Neural Network for Processing Fingerprint Image Noise | |
Talebi et al. | Asymptotic performance of global denoising |
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 | ||
EE01 | Entry into force of recordation of patent licensing contract |
Application publication date: 20170531 Assignee: Anhui Sutong Intelligent Technology Co.,Ltd. Assignor: NANJING University OF POSTS AND TELECOMMUNICATIONS Contract record no.: X2021980006185 Denomination of invention: An image denoising method based on noise prediction Granted publication date: 20190723 License type: Common License Record date: 20210716 |
|
EE01 | Entry into force of recordation of patent licensing contract |