CN104240187A - Image denoising device and image denoising method - Google Patents
Image denoising device and image denoising method Download PDFInfo
- Publication number
- CN104240187A CN104240187A CN201310236003.0A CN201310236003A CN104240187A CN 104240187 A CN104240187 A CN 104240187A CN 201310236003 A CN201310236003 A CN 201310236003A CN 104240187 A CN104240187 A CN 104240187A
- Authority
- CN
- China
- Prior art keywords
- image
- feature
- unit
- denoising
- block
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Abstract
The invention provides an image denoising device and an image denoising method. The image denoising device comprises an input unit, a partitioning unit, a feature extraction unit, a clustering unit, a filtering unit and an output unit, wherein the input unit is used for inputting image data of an image, the partitioning unit is used for dividing the image into a plurality of blocks according to the input image data, the feature extraction unit is used for extracting features of all the blocks in the image, the clustering unit is used for carrying out clustering according to the extracted features, and the blocks in the image are grouped to obtain block sets. The filtering unit is used for denoising the block sets according to a multi-linear subspace learning filter. The output unit is used for outputting the denoised image data. According to the image denoising device, the blocks in the whole image can be clustered and grouped, and therefore the similar blocks can be extracted from the whole image. The image denoising device can not only denoise a two-dimensional image and can also denoise a three-dimensional image and even an image with the higher dimensionality, and therefore the image denoising device can denoise medical images and other images with any dimensionality.
Description
Technical field
The present invention relates to image denoising device and image de-noising method, the image particularly related to for Arbitrary Dimensions such as medical images carries out image denoising device and the image de-noising method of denoising.
Background technology
Image denoising is very important preprocessing means.It can help improve image quality, improves the degree of accuracy of Iamge Segmentation, image registration etc.For medical image, vision-based detection intuitively can also be carried out by assist physician.But image denoising is also faced with huge challenge.Reason is all features needing to retain as far as possible original image removing noise while.
In recent years, emerge large quantities of noise-removed technology for two dimensional image, BM3D (block-matching and3D filtering: Block-matching and the three-dimensional filtering) technology of wherein relatively the most effectively non-patent literature 1 proposition.In BM3D technology, for a width still image, the three-dimensional module for wherein similar two-dimensional block composition carries out space transforming, and carries out inverse transformation after utilizing shrinkage method to extract coefficient of efficiency, thus reaches denoising object.
But, in above-mentioned BM3D technology, still there are some problems.
First, in Block-matching process, above-mentioned BM3D technology finds similar piece of reference block in specific scope.But such as medical image, similar piece is present in whole image.Therefore, similar piece that searches out in specific scope is very limited.And above-mentioned BM3D technology finds similar piece of reference block based on the brightness value of image.But brightness value is easy to by noise, similar piece search directly can be had influence on.
In addition, along with the development of modern technologies, the kind of image is also various all the more, not only has the natural image of two dimension, and also has three-dimensional medical image, high-definition picture etc.Above-mentioned BM3D technology only with two dimensional image as denoising object, can not carry out denoising for three-dimensional and dimensional images.And above-mentioned BM3D technology can not judge the dimension of input picture, thus different modes can not be used to carry out denoising for different images.
Moreover carrying out in space transforming process, the conversion substrate that above-mentioned BM3D Technology application one group is fixed solves the Denoising Problems of all images.Change because this group conversion substrate does not rely on the change of input picture, be difficult to the specific characteristic of corresponding different images.And, the every one dimension of above-mentioned BM3D technology for the multidimensional module after composition is changed respectively, does not consider the crosscorrelation contact between every one dimension.
In addition, as the direct technology processed high dimensional data, in non-patent literature 2, GND-PCA technology is proposed.But, in non-patent literature 2, do not relate to and how GND-PCA technology be used for image denoising.
Non-patent literature 1: " Image denoising by sparse3D transform-domain collaborative filtering ", the work such as Kostadin Dabov, 2007
Non-patent literature 2: " Generalized N-dimensional principal component analysis (GND-PCA) and its application on construction of statistical appearance models for medical volumes with fewer samples ", the work such as Rui Xu, 2009
As mentioned above, in existing Image Denoising Technology, in Block-matching process, space transforming process etc., there are problems.And, for the multidimensional image extensively utilized as medical image etc., still lack effective Image Denoising Technology.
Summary of the invention
The present invention is directed to the problems referred to above of the prior art, its object is to, a kind of image denoising device and the image de-noising method that effectively can carry out denoising for the image of the Arbitrary Dimensions such as medical image are provided.
In order to solve the aforementioned problems in the prior, the invention provides a kind of image denoising device, for carrying out denoising to image, it is characterized in that, possessing: input block, inputting the view data of described image; Blocking unit, according to the view data of input, is divided into multiple pieces by described image; Feature extraction unit, extracts the feature of each block in described image; Cluster cell, carries out cluster based on the feature extracted, and is divided into groups and obtains block group by described multiple pieces in described image thus; Filter unit, utilizes polyteny sub-space learning wave filter to carry out denoising to block group; And output unit, export the view data after denoising.
According to image denoising device of the present invention, the image of input is divided into multiple pieces, extracts the feature of each block in image, and carry out cluster based on the feature extracted, thus by multiple pieces of groupings in image.Thereby, it is possible to carry out cluster to block and divide into groups in whole image, thus similar piece can be extracted in whole image.And, by the polyteny sub-space learning wave filter such as based on GND-PCA, GND-ICA, MPCA or MICA is used for image denoising, can not only to two dimensional image denoising, also can to three-dimensional even more dimensional images carry out denoising, thus denoising can be carried out to the image of Arbitrary Dimensions.
In image denoising device of the present invention, also can be that described feature extraction unit extracts multiple features of each block in described image; Described image denoising device also possesses feature selection unit, and this feature selection unit selects a part of feature from described multiple feature; Described cluster cell carries out cluster based on the feature selected by described feature selection unit.
Wherein, also can be that described feature extraction unit also combines extracted multiple features, obtains union feature thus; Described feature selection unit carries out feature selecting to described union feature, selects a part of feature thus from described multiple feature.
Wherein, also can be that described feature selection unit selects a part of feature from described multiple feature, reduce to make the computing time needed for the cluster of described cluster cell.
Thus, extract multiple feature from each block image, with in prior art only based on the brightness value of image mode compared with, the accuracy of grouping can be improved, improve the similarity between the block that is grouped.And, from multiple feature, such as utilize PCA, ICA to select a part of feature for cluster, the processing time of cluster can not only be reduced, can also effectively avoid noise on the impact of feature.
In image denoising device of the present invention, also can be that described feature extraction unit, based at least one in brightness, color, shape, texture, SIFT descriptor and scale invariant feature conversion described symbol, extracts the feature of each block in described image.
At this, list the several preferred feature for cluster, carry out cluster by least one utilizing in these features, the accuracy of grouping can be improved further.
In image denoising device of the present invention, also can be that described image denoising device also possesses dimension detecting unit, this dimension detecting unit detects the dimension of the view data inputted by described input block.
Wherein, also can be that, when the dimension that described dimension detecting unit is detected as the described view data inputted by described input block is greater than 2, described filter unit utilizes polyteny sub-space learning wave filter to carry out denoising to block group.
Wherein, also can be that described image denoising device also possesses pre-filtering processing unit; When the dimension that described dimension detecting unit is detected as the described view data inputted by described input block equals 2, described pre-filtering processing unit carries out pre-filtering process to this view data, and the view data after pre-filtering process is supplied to described blocking unit.
Thus, by judging the dimension of the image of input, the denoising of the image being more suitable for this dimension can be selected.Such as, when the dimension of image is greater than 2, denoising is carried out by utilizing polyteny sub-space learning wave filter, denoising can not only be carried out to the multidimensional image that cannot carry out denoising in the past, and the information of input picture can be made full use of, the filtering of application polyteny sub-space learning, considers the mutual relationship of image inside, improves denoising effect.In addition, when the dimension of image equals 2, by first carrying out pre-filtering process, more described above denoising being carried out to image, denoising effect can be improved further.
In image denoising device of the present invention, also can be that described image denoising device also possesses polymerized unit, the block in the block group after described filter unit denoising, by polymerization, be reverted to the original position of this block in described image by this polymerized unit; Described output unit exports the view data after being reduced by described polymerized unit.
Thus, by the block after denoising is reverted to the original position of this block, and export the image after reduction, the general image after denoising can be provided to user.
In image denoising device of the present invention, also can be that described image is divided into equal and opposite in direction and overlapped multiple pieces by described blocking unit; When the dimension of the view data inputted is A, the dimension of described piece is A, and described piece of group shows as the module of dimension A+1, and wherein A is the integer of more than 2.
At this, by the image of input is divided into equal and opposite in direction and overlapped multiple pieces, can effectively in whole image, carries out cluster to block and divide into groups, thus extract similar piece in whole image.In addition, by multiple pieces of groupings being carried out the block group of constituent ratio block height one dimension, in subsequent treatment, denoising can be carried out to this block group.
In image denoising device of the present invention, also can be that described cluster cell utilizes when cluster and allows soft cluster overlapping between cluster.
Thus, by utilizing the soft cluster modes such as KFCM_S to carry out cluster in whole image range, both the overall situation can consider the information of entire image, and also can, according to the characteristic of every block, make gained divide into groups there is common factor.
In order to solve the aforementioned problems in the prior, the present invention also provides a kind of image de-noising method, for carrying out denoising to image, it is characterized in that, comprising: input step, inputs the view data of described image; Blocking step, according to the view data of input, is divided into multiple pieces by described image; Characteristic extraction step, extracts the feature of each block in described image; Sorting procedure, carries out cluster based on the feature extracted, and is divided into groups and obtains block group by described multiple pieces in described image thus; Filter step, utilizes polyteny sub-space learning wave filter to carry out denoising to block group; And output step, export the view data after denoising.
According to image de-noising method of the present invention, the image of input is divided into multiple pieces, extracts the feature of each block in image, and carry out cluster based on the feature extracted, thus by multiple pieces of groupings in image.Thereby, it is possible to carry out cluster to block and divide into groups in whole image, thus similar piece can be extracted in whole image.And, by the polyteny sub-space learning wave filter such as based on GND-PCA, GND-ICA, MPCA or MICA is used for image denoising, can not only to two dimensional image denoising, also can to three-dimensional even more dimensional images carry out denoising, thus denoising can be carried out to the image of Arbitrary Dimensions.
According to image denoising device of the present invention and image de-noising method, effectively denoising can be carried out for the image of the Arbitrary Dimensions such as medical image.Wherein, the present invention is not limited to mode listed above.Such as, also can mutually combine at above-mentioned each optimal way of image denoising device of the present invention, and these optimal ways and combination certainly also can be applicable to image de-noising method of the present invention.
Accompanying drawing explanation
Fig. 1 is the schematic diagram using MRI Medical Instruments entirety of the present invention.
Fig. 2 is the structured flowchart of the image denoising device of first embodiment of the invention.
Fig. 3 is the process flow diagram of the image de-noising method of first embodiment of the invention.
Fig. 4 is the structured flowchart of the image denoising device of second embodiment of the invention.
Fig. 5 is the process flow diagram of the image denoising process of one embodiment of the present of invention.
Fig. 6 is the process flow diagram of the image denoising process of an alternative embodiment of the invention.
Fig. 7 is the process flow diagram of the image denoising process of prior art.
Embodiment
Below in conjunction with accompanying drawing, embodiment and embodiment, the present invention will be described in more detail.In addition, be accompanied by same Reference numeral to same or considerable part in the accompanying drawings, the repetitive description thereof will be omitted.
First, the one-piece construction as the MRI Medical Instruments using image denoising device of the present invention and image de-noising method is described.At this, be applicable to MRI Medical Instruments for image denoising device of the present invention and image de-noising method be described, but image denoising device of the present invention and image de-noising method are not limited to be applicable to this, can also be applicable to the various image denoising process beyond other various Medical Instruments such as CT and medical application.
Fig. 1 is the schematic diagram using MRI Medical Instruments entirety of the present invention.According to Fig. 1, common MRI system consists of the following components: main magnet; Gradient Unit (gradient coil, gradient pulse program); Radio frequency unit (radio-frequency coil, pulse protocol, receiving cable, transmission channel); Computer system and spectrometer (computing machine, storer, display); Other utility appliance (shielding, sick bed).The present invention is directed to computer system and spectrometer improves, as shown in " image denoising " in Fig. 1 dotted line frame.Other parts such as main magnet, Gradient Unit, radio frequency unit and other utility appliance can utilize prior art, do not repeat at this.
Below describe image denoising device of the present invention and image de-noising method in detail.In order to understand the present invention better, the first image denoising process of schematic illustration prior art.Fig. 7 is the process flow diagram of the image denoising process of prior art.As shown in Figure 7, in the image denoising process of prior art, first two-dimentional noise image is inputted, then by the grouped element mated based on two-dimensional block, Block-matching is carried out to two-dimentional noise image, then by orthogonal transformation filter unit, filtering is carried out to the block group obtained by Block-matching, then revert to original position by two-dimentional polymerized unit by filtered piece, finally export the two dimensional image after denoising.In the image denoising process of prior art, as mentioned above, not only in Block-matching process, space transforming process etc., there are problems, and three-dimensional above multidimensional image cannot be applicable to.
For the problems referred to above of the image denoising process of prior art, image denoising device provided by the invention and image de-noising method can carry out denoising for the image of the Arbitrary Dimensions such as medical image effectively.
First, image denoising device and the image de-noising method of first embodiment of the invention is illustrated.Fig. 2 is the structured flowchart of the image denoising device of first embodiment of the invention.As shown in Figure 2, image denoising device 1, for carrying out denoising to the image of input, both can perform image denoising program by general computing machine and realize, also can realize based on special hardware configuration.When image denoising device 1 is realized by special hardware configuration, such as, can be realized by the special IC based on FPGA technology.Image denoising device 1 possesses input block 11, blocking unit 12, feature extraction unit 13, cluster cell 14, filter unit 15 and output unit 16.
As shown in Figure 2, the view data of input block 11 input picture, the form of the image wherein inputted is not limit, and can be the arbitrary formats such as DICOM, bmp, jpg, tiff, gif, raw.Image, according to the view data of input, is divided into multiple pieces by blocking unit 12.Feature extraction unit 13 extracts the feature of each block in image.Cluster cell 14 carries out cluster based on the feature extracted, and is divided into groups by multiple pieces in image thus and obtains block group.Filter unit 15 utilizes polyteny sub-space learning wave filter to carry out denoising to block group.Output unit 16 exports the view data after denoising.
Then, the image de-noising method that the image denoising device of first embodiment of the invention performs is described.Fig. 3 is the process flow diagram of the image de-noising method of first embodiment of the invention.As shown in Figure 3, image de-noising method comprises input step S1, blocking step S2, characteristic extraction step S3, sorting procedure S4, filter step S5 and exports step S6.
In input step S1, the view data of input block 11 input picture.In the present invention, the dimension for input picture does not limit, and is likely the image of two dimension, is also likely image that is three-dimensional or more higher-dimension.Wherein, dimension that is three-dimensional or the more image of higher-dimension refers to the length of this image, sample number and time etc.
In blocking step S2, image, according to the view data of input, is divided into multiple pieces by blocking unit 12.Such as, input picture is divided into equal and opposite in direction and overlapped multiple pieces by blocking unit 12, or gets equal-sized multiple pieces at random.
In characteristic extraction step S3, feature extraction unit 13 extracts the feature of each block in image.Such as, each piece in image is extracted to its proper vector, as brightness, color, shape, texture, SIFT descriptor etc.
In sorting procedure S4, sorting procedure 14 carries out cluster based on the feature extracted, and is divided into groups by multiple pieces in image thus and obtains block group.Such as, according to the feature of each block, utilize soft clustering method, find the block with similar features, and by similar piece of grouping, obtain module and the block group of a higher one dimension.
In filter step S5, filter unit 15 utilizes polyteny sub-space learning wave filter to carry out denoising to block group, namely mainly adopts for the module of above-mentioned higher one dimension and carries out denoising based on polyteny sub-space learning wave filter.Wherein, refer to according to input picture based on polyteny sub-space learning wave filter, by considering influencing each other between every one dimension, system automatic learning goes out at the bottom of a transform-based, transformation space is changed because of the difference of input picture, thus reaches the object of auto adapted filtering.Of the present invention is wave filter based on GND-PCA (Generalized N Dimensional Principal Component Analysis) based on polyteny sub-space learning wave filter, and be not limited to this kind of polyteny sub-space learning wave filter, other also can adopt as GND-ICA (Generalized N Dimensional Independent Component Analysis), MPCA or MICA etc.By using polyteny sub-space learning wave filter, thus make to become possibility to the denoising of 3-D view.
In output step S6, output unit 16 exports the view data after denoising.
According to the image denoising device of first embodiment of the invention, the image of input is divided into multiple pieces, extracts the feature of each block in image, and carry out cluster based on the feature extracted, thus by multiple pieces of groupings in image.Thereby, it is possible to carry out cluster to block and divide into groups in whole image, thus similar piece can be extracted in whole image.And, by the polyteny sub-space learning wave filter such as based on GND-PCA, GND-ICA, MPCA or MICA is used for image denoising, can not only to two dimensional image denoising, also can to three-dimensional even more dimensional images carry out denoising, thus denoising can be carried out to the image of Arbitrary Dimensions.
Then, image denoising device and the image de-noising method of second embodiment of the invention is illustrated.The image denoising device of the second embodiment of the present invention and image de-noising method are the further improvement to the first embodiment of the present invention, omit the description for the structure same or equivalent with the first embodiment and step.
Fig. 4 is the structured flowchart of the image denoising device of second embodiment of the invention.As shown in Figure 4, the image denoising device 10 of second embodiment of the invention, except the entire infrastructure of image denoising device 1 possessing the first embodiment, also possesses dimension detecting unit 17, pretreatment unit 18, feature selection unit 19 and polymerized unit 20.Below only these structures are stressed.
Dimension detecting unit 17 detects the dimension of the view data inputted by input block 11.As mentioned above, in the present invention, the dimension of input picture does not limit, and is likely the image of two dimension, is also likely image that is three-dimensional or more higher-dimension.At this, first dimension detecting unit 17 carries out dimension judgement to input data, according to the judged result of dimension detecting unit 17, determines whether carry out pre-service by pretreatment unit 18 pairs of view data.Specifically, when the dimension that dimension detecting unit 17 is detected as view data is greater than 2, not by pretreatment unit 18, and directly view data is supplied to blocking unit 12; When the dimension that dimension detecting unit 17 is detected as view data equals 2, view data is supplied to pretreatment unit 18, carries out pre-service by pretreatment unit 18 pairs of view data.
Pretreatment unit 18 pairs of view data carry out pre-filtering process, and the view data after pre-filtering process is supplied to blocking unit 12.
Multiple features that feature selection unit 19 extracts from feature extraction unit 13 each block, select a part of feature.Such as, when feature extraction unit 13 to extracted multiple features (proper vector) combine obtain union feature (one-dimensional vector) thus, feature selection unit 19 pairs of union features carry out feature selecting, select a part of feature thus from multiple feature.At this, feature selection unit 19 such as by the degree of functioning of multiple feature, extracts more efficiently a part of feature in multiple feature.In addition, feature selection unit 19, by selecting a part of feature from multiple feature, reduces to make the computing time needed for the cluster of cluster cell 14.Specifically, when forming an one-dimensional vector when being combined by the vector of feature extraction unit 13 to multiple features that each block extracts, this one-dimensional vector of multiple pieces in image forms a matrix, by calculating transition matrix by feature selection unit 19 to this matrix, thus extract the major component of proper vector, and use it for the cluster of cluster cell 14, the operand of cluster cell 14 can be reduced, reduce the computing time needed for cluster.
Block in block group after filter unit 15 denoising, by polymerization, is reverted to this block original position in the picture by polymerized unit 20, and the image after reduction is supplied to output unit 16, exports for output unit 16.
The image de-noising method of second embodiment of the invention adds the dimension detecting step, pre-treatment step, feature selection step and the polymerization procedure that are performed respectively by above-mentioned dimension detecting unit 17, pretreatment unit 18, feature selection unit 19 and polymerized unit 20 equally and realizes on the basis of the image de-noising method of the first embodiment.Image denoising device second embodiment of the invention and image de-noising method, can not only obtain the effect identical with the first embodiment, but also can obtain additional effect as above.
Wherein, in second embodiment of the invention, describe the example adding above-mentioned dimension detecting unit 17, pretreatment unit 18, feature selection unit 19 and polymerized unit 20 on the basis of the image denoising device 1 of the first embodiment, but also only can add dimension detecting unit 17 and pretreatment unit 18, or only add feature selection unit 19, or only add polymerized unit 20.Equally, also on the basis of the image de-noising method of the first embodiment, only can add dimension detecting step and pre-treatment step, or only add feature selection step, or only add polymerization procedure.
Below, illustrate two of the present invention and represent embodiment.At this, following embodiment is only for understanding the present invention better, not as a limitation of the invention.
Fig. 5 is the process flow diagram of the image denoising process of one embodiment of the present of invention.As shown in Figure 5, first, system acceptance is to a new input picture.In the present embodiment, no matter the dimension of input picture is two dimension or higher-dimension (more than three-dimensional), all performs identical process.Therefore, both can judge that input picture was two dimension or higher-dimension, also can not have been judged.Be that the situation of two dimensional image is described for example below with input picture.
Then, input picture is divided into equal and opposite in direction and overlapped block, B={b
1, b
2..., b
n, wherein n is the quantity of block.Such as, a width pixel is the image of 505 × 505, can be divided into 17 × 17 and coincident pixel be 8 3844 blocks.Further, the value of the pixel of image, the size of block, coincident pixel the numerical value be not limited in citing, the as required size of adjustable numerical value.In addition, also equal-sized piece can be got at random in entire image.
Then, each piece is extracted to its proper vector, as brightness, color, shape, texture, SIFT descriptor etc.Wherein, color characteristic is as color histogram (color histogram), color scale (scalable color), colour moment (color moment) etc.; Textural characteristics is as Tamura feature, overall Texture descriptor (global texture descriptor), gabor histogram, gabor vector etc.; Shape facility is as edge histogram (edge histogram) etc.SIFT descriptor is also the method for effectively extraction block feature in addition.Because all features extracted are all by vector representation, the present invention is by all r proper vector q
1, q
2..., q
rall be linked to be an one-dimensional vector y=[q
1, q
2..., q
r]
tcarry out subsequent calculations.
Then, feature selecting is carried out to one-dimensional vector, not only accelerate the computing time of follow-up cluster, reduce the impact of noise on the feature extracted simultaneously.Vector x=A that the one-tenth that feature selecting goes out is grouped into
ty, wherein A is the transition matrix that the methods such as Based PC A, ICA learn out; The element number of x is less than or equal to the element number of y.
Then, cluster is carried out to the feature after selection.Cluster be according to object between similarity carry out the process distinguishing and classify, be a kind of unsupervised classification.Here the cluster mode taked can be hard clustering method, as synthesized cluster (Hierarchical agglomerative clustering), K-means cluster, also can be soft clustering method Fuzzy c-means, fuzzy-possibilistic c-means, Expectation Maximization based clustering.Hard clustering method is strictly divided into each pending object in certain class, and its degree of membership is not 1 is exactly 0.Soft clustering method allows between cluster overlapping, allows fuzzy border.That is, a block can be assigned in several different class simultaneously and go.Below for fuzzy-possibilistic c-means cluster, the method for soft cluster is described.One group of data be not labeled
wherein, n is the number of the block be extracted in piece image, x
kthe proper vector of the block obtained after being through feature selecting, x
k∈ R
p, p is x
kthe number of middle element.The objective function of fuzzy-possibilistic c-means is
Wherein
A subordinated-degree matrix,
C cluster centre;
represent x
ktypicalness in i-th class or x
kbelong to the probability of i-th class, it only depends on x
kwith v
idistance, and to have nothing to do with the position at other class centers, and
Be distance matrix, A-norm is defined as
Constant m > 1, η > 1.
In order to minimize objective function, must meet:
Be subordinate to angle value:
Typicalness value:
Cluster centre:
Here, u
ikbe subject to the impact of all c cluster centre; And t
ikonly by the impact of i-th cluster centre.
Now sum up fuzzy-possibilistic c-means clustering processing as following table 1.
[table 1]
According to content and the actual needs of T, definite threshold carries out soft cluster to each proper vector.
Then, in two dimensional image
(I
i, i=1,2 is the parts number on every one dimension), find corresponding two-dimensional block according to proper vectors all in each class after cluster
(q is the number of the corresponding block found), grouping obtains a three-dimensional module
Then, right
module application carries out denoising based on polyteny sub-space learning wave filter.Its process is as follows: first, on every one dimension, by high dimensional data matrixing, calculates eigenvectors matrix.And according to eigenwert size and threshold size, former the proper vectors that keeping characteristics value is relatively maximum composition characteristic vector matrix.Secondly, by alternative manner, find optimized eigenvectors matrix U on corresponding every one dimension
k, 1≤k≤3.Note, in the third dimension, namely in sample dimension, retain all proper vectors.Again, by the module of three-dimensional
project to the eigenvectors matrix after optimization to produce spatially:
.Finally, by rebuilding, the three-dimensional module after denoising is obtained:
Then, by the method for polymerization, polymerized unit reverts to position original in image the block after denoising, obtains denoising output image.
Finally, the view data after denoising is exported.
Fig. 6 is the process flow diagram of the image denoising process of an alternative embodiment of the invention.As shown in Figure 6, the present embodiment is with the difference of the upper embodiment illustrated with reference to Fig. 5, judges that the image of input is two dimension or higher-dimension (more than three-dimensional), and carries out different process accordingly.
That is, when judging that input picture is two dimensional image, pre-filtering process is carried out to input picture.Wave filter now can be any one image denoising wave filter.Mainly contain three major types.(1) Fourier transformation method: DCT, wavelet transformation; (2) Bayesian Estimation method; (3) based on the filtering method of similar piece: BM3D, NLM etc.Using pretreated image as new input data.Process is afterwards identical with the upper embodiment illustrated with reference to Fig. 5 before.
When judging the dimension of input picture higher than two dimension, perform following process.
First, dimensional images will be inputted
(wherein N is the dimension of image, N > 2, I
i, i=1,2 ..., N is the parts number on every one dimension) be divided into equal-sized overlapped higher-dimension block,
j
i< I
i, wherein n is the quantity of block.Wherein, the value of the pixel of image, the size of block, coincident pixel the numerical value be not limited in citing, the as required size of adjustable numerical value.In addition, equal-sized piece can be got at random in entire image.
Then, each higher-dimension block is extracted to its proper vector, as brightness, color, shape, texture, SIFT descriptor etc.Because all features extracted are all by vector representation, the present invention is by all r proper vector q
1, q
2..., q
rall be linked to be an one-dimensional vector y=[q
1, q
2..., q
r]
tcarry out subsequent calculations.
Then, feature selecting is carried out to one-dimensional vector, not only accelerate the computing time of follow-up cluster, reduce the impact of noise on the feature extracted simultaneously.Vector x=A that the one-tenth that feature selecting goes out is grouped into
ty, wherein A is the transition matrix that the methods such as Based PC A, ICA learn out; The element number of x is less than or equal to the element number of y.
Then, cluster is carried out to the feature after selection.Clustering method is consistent with the upper embodiment illustrated with reference to Fig. 5 before.
Then, at dimensional images
in, find corresponding higher-dimension block according to proper vectors all in each class after cluster
(m is the number of the similar higher-dimension block found), grouping obtains the module of a higher one dimension
Then, right
module application carries out denoising based on polyteny sub-space learning wave filter.Its process is as follows: first, on every one dimension, by high dimensional data matrixing, calculates eigenvectors matrix.And according to eigenwert size and threshold size, former the characteristic of correspondence vectors also composition characteristic vector matrix that keeping characteristics value is relatively large.Secondly, by alternative manner, find optimized eigenvectors matrix U on corresponding every one dimension
k, 1≤k≤N+1.Note, in the end one dimension, namely in sample dimension, retain all proper vectors.Again, by the module of higher one dimension
project to the eigenvectors matrix after optimization to produce spatially:
.Finally, by rebuilding, the higher-dimension module after denoising is obtained:
Then, the method for being polymerized by higher-dimension, being reverted to position original in image obtaining higher-dimension module after denoising, obtaining denoising output image.The method of this higher-dimension polymerization is the expansion of the polymerization applied in prior art, makes it can be polymerized equally high dimensional data.
Finally, the view data after denoising is exported.
Above with reference to the accompanying drawings of the specific embodiment of the present invention and embodiment.Wherein, embodiment described above and embodiment are only object lessons of the present invention, for understanding the present invention, and are not used in restriction scope of the present invention.Those skilled in the art can carry out the reasonable omission of various distortion, combination and key element to embodiment and embodiment based on technological thought of the present invention, the mode obtained thus is also included within scope of the present invention.
Claims (12)
1. an image denoising device, for carrying out denoising to image, is characterized in that, possessing:
Input block, inputs the view data of described image;
Blocking unit, according to the view data of input, is divided into multiple pieces by described image;
Feature extraction unit, extracts the feature of each block in described image;
Cluster cell, carries out cluster based on the feature extracted, and is divided into groups and obtains block group by described multiple pieces in described image thus;
Filter unit, utilizes polyteny sub-space learning wave filter to carry out denoising to block group; And
Output unit, exports the view data after denoising.
2. image denoising device as claimed in claim 1, is characterized in that,
Described feature extraction unit extracts multiple features of each block in described image;
Described image denoising device also possesses feature selection unit, and this feature selection unit selects a part of feature from described multiple feature;
Described cluster cell carries out cluster based on the feature selected by described feature selection unit.
3. image denoising device as claimed in claim 2, is characterized in that,
Described feature extraction unit also combines extracted multiple features, obtains union feature thus;
Described feature selection unit carries out feature selecting to described union feature, selects a part of feature thus from described multiple feature.
4. image denoising device as claimed in claim 1, is characterized in that,
Described feature extraction unit, based at least one in brightness, color, shape, texture, SIFT descriptor and scale invariant feature conversion described symbol, extracts the feature of each block in described image.
5. image denoising device as claimed in claim 2, is characterized in that,
Described feature selection unit selects a part of feature from described multiple feature, reduces to make the computing time needed for the cluster of described cluster cell.
6. image denoising device as claimed in claim 1, is characterized in that,
Described image denoising device also possesses dimension detecting unit, and this dimension detecting unit detects the dimension of the view data inputted by described input block.
7. image denoising device as claimed in claim 6, is characterized in that,
When the dimension that described dimension detecting unit is detected as the described view data inputted by described input block is greater than 2, described filter unit utilizes polyteny sub-space learning wave filter to carry out denoising to block group.
8. image denoising device as claimed in claim 6, is characterized in that,
Described image denoising device also possesses pre-filtering processing unit;
When the dimension that described dimension detecting unit is detected as the described view data inputted by described input block equals 2, described pre-filtering processing unit carries out pre-filtering process to this view data, and the view data after pre-filtering process is supplied to described blocking unit.
9. image denoising device as claimed in claim 1, is characterized in that,
Described image denoising device also possesses polymerized unit, and the block in the block group after described filter unit denoising, by polymerization, is reverted to the original position of this block in described image by this polymerized unit;
Described output unit exports the view data after being reduced by described polymerized unit.
10. image denoising device as claimed in claim 1, is characterized in that,
Described image is divided into equal and opposite in direction and overlapped multiple pieces by described blocking unit;
When the dimension of the view data inputted is A, the dimension of described piece is A, and described piece of group shows as the module of dimension A+1, and wherein A is the integer of more than 2.
11. image denoising devices as claimed in claim 1, is characterized in that,
Described cluster cell utilizes when cluster and allows soft cluster overlapping between cluster.
12. 1 kinds of image de-noising methods, for carrying out denoising to image, is characterized in that, comprising:
Input step, inputs the view data of described image;
Blocking step, according to the view data of input, is divided into multiple pieces by described image;
Characteristic extraction step, extracts the feature of each block in described image;
Sorting procedure, carries out cluster based on the feature extracted, and is divided into groups and obtains block group by described multiple pieces in described image thus;
Filter step, utilizes polyteny sub-space learning wave filter to carry out denoising to block group; And
Export step, export the view data after denoising.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310236003.0A CN104240187A (en) | 2013-06-08 | 2013-06-08 | Image denoising device and image denoising method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310236003.0A CN104240187A (en) | 2013-06-08 | 2013-06-08 | Image denoising device and image denoising method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN104240187A true CN104240187A (en) | 2014-12-24 |
Family
ID=52228197
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201310236003.0A Pending CN104240187A (en) | 2013-06-08 | 2013-06-08 | Image denoising device and image denoising method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104240187A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106875280A (en) * | 2017-03-02 | 2017-06-20 | 深圳大图科创技术开发有限公司 | Integrated community service platform |
| CN106909946A (en) * | 2017-03-02 | 2017-06-30 | 深圳明创自控技术有限公司 | A kind of picking system of multi-modal fusion |
| CN106919954A (en) * | 2017-03-02 | 2017-07-04 | 深圳明创自控技术有限公司 | A kind of cloud computing system for commodity classification |
| CN111724317A (en) * | 2020-05-20 | 2020-09-29 | 天津大学 | Method for constructing Raw domain video denoising supervision data set |
| CN105117740B (en) * | 2015-08-21 | 2021-06-15 | 北京旷视科技有限公司 | Font identification method and apparatus |
| CN113436093A (en) * | 2021-06-16 | 2021-09-24 | 中国电子科技集团公司第五十四研究所 | Remote sensing image denoising method based on energy domain transformation |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20070110294A1 (en) * | 2005-11-17 | 2007-05-17 | Michiel Schaap | Image enhancement using anisotropic noise filtering |
| CN102663686A (en) * | 2012-03-19 | 2012-09-12 | 西安电子科技大学 | Image denoising method based on Treelet switch and Gaussian scale mixture model |
| CN103093433A (en) * | 2013-01-25 | 2013-05-08 | 西安电子科技大学 | Natural image denoising method based on regionalism and dictionary learning |
-
2013
- 2013-06-08 CN CN201310236003.0A patent/CN104240187A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20070110294A1 (en) * | 2005-11-17 | 2007-05-17 | Michiel Schaap | Image enhancement using anisotropic noise filtering |
| CN102663686A (en) * | 2012-03-19 | 2012-09-12 | 西安电子科技大学 | Image denoising method based on Treelet switch and Gaussian scale mixture model |
| CN103093433A (en) * | 2013-01-25 | 2013-05-08 | 西安电子科技大学 | Natural image denoising method based on regionalism and dictionary learning |
Non-Patent Citations (1)
| Title |
|---|
| 李彬等: "非局部降噪快速模糊C-均值聚类算法", 《计算机工程与应用》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105117740B (en) * | 2015-08-21 | 2021-06-15 | 北京旷视科技有限公司 | Font identification method and apparatus |
| CN106875280A (en) * | 2017-03-02 | 2017-06-20 | 深圳大图科创技术开发有限公司 | Integrated community service platform |
| CN106909946A (en) * | 2017-03-02 | 2017-06-30 | 深圳明创自控技术有限公司 | A kind of picking system of multi-modal fusion |
| CN106919954A (en) * | 2017-03-02 | 2017-07-04 | 深圳明创自控技术有限公司 | A kind of cloud computing system for commodity classification |
| CN111724317A (en) * | 2020-05-20 | 2020-09-29 | 天津大学 | Method for constructing Raw domain video denoising supervision data set |
| CN113436093A (en) * | 2021-06-16 | 2021-09-24 | 中国电子科技集团公司第五十四研究所 | Remote sensing image denoising method based on energy domain transformation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wang et al. | Breast cancer image classification via multi-network features and dual-network orthogonal low-rank learning | |
| WO2018045626A1 (en) | Super-pixel level information fusion-based hyperspectral image classification method and system | |
| Shao et al. | Improved color texture descriptors for remote sensing image retrieval | |
| Zheng et al. | Discriminative dictionary learning via Fisher discrimination K-SVD algorithm | |
| CN104240187A (en) | Image denoising device and image denoising method | |
| Jian et al. | Content-based image retrieval via a hierarchical-local-feature extraction scheme | |
| CN104778476A (en) | Image classification method | |
| Yu et al. | A content-based goods image recommendation system | |
| Panda et al. | Automated brain tumor detection using discriminative clustering based MRI segmentation | |
| Syam et al. | An effective similarity measure via genetic algorithm for content based image retrieval with extensive features. | |
| Fang et al. | Image classification with an RGB-channel nonsubsampled contourlet transform and a convolutional neural network | |
| CN104281856A (en) | Image preprocessing method and system for brain medical image classification | |
| Bai et al. | K-means based histogram using multiresolution feature vectors for color texture database retrieval | |
| Liu et al. | Improving image retrieval by integrating shape and texture features | |
| Wang et al. | A new Gabor based approach for wood recognition | |
| Etemad et al. | Color texture image retrieval based on Copula multivariate modeling in the Shearlet domain | |
| CN116310452B (en) | Multi-view clustering method and system | |
| Dandotiya et al. | Image retrieval using edge detection, RLBP, color moment method for YCbCr and HSV color space | |
| Bi et al. | A robust color edge detection algorithm based on the quaternion Hardy filter | |
| CN102306179B (en) | Image content retrieval method based on hierarchical color distribution descriptor | |
| Mohite et al. | 3D local circular difference patterns for biomedical image retrieval | |
| Akbari et al. | Feature fusion based on joint sparse representations and wavelets for multiview classification | |
| Guo et al. | Breast cancer detection via wavelet energy and support vector machine | |
| Qiao et al. | Lung nodule classification using curvelet transform, LDA algorithm and BAT-SVM algorithm | |
| Lin et al. | Bayesian information criterion based feature filtering for the fusion of multiple features in high-spatial-resolution satellite scene classification |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| AD01 | Patent right deemed abandoned | ||
| AD01 | Patent right deemed abandoned |
Effective date of abandoning: 20180302 |