CN106157240A - Remote sensing image super resolution method based on dictionary learning - Google Patents

Abstract

The present invention provides a kind of Remote Sensing Image Super Resolution method based on dictionary learning, and the method includes: step 1: for the image library of different types of ground objects, carry out the dictionary learning of corresponding type of ground objects；Step 2, type of ground objects identification is carried out to original image；Step 3, use species type-word allusion quotation accordingly, carry out image super-resolution restructuring procedure.The present invention utilizes the thought of terrain classification, the super-resolution method based on classifying dictionary of proposition.Owing to the number of atom in the dictionary that restructuring procedure uses is greatly reduced, dictionary scale is greatly reduced, and effective atomic number purpose ratio is significantly increased in the classifying dictionary using, so using effective atomic number mesh in dictionary almost identical with general dictionary at restructuring procedure, while ensureing reconstruction quality, reconstructed velocity is greatly improved.This process relates only to the regulation of parameter, and computational efficiency is high, and the reconstruction quality for remote sensing images is high.

Description

Remote Sensing Image Super Resolution method based on dictionary learning

Technical field

The present invention relates to remotely sensed image technical field, in particular to a kind of Remote Sensing Image Super Resolution method based on dictionary learning.

Background technology

With the development of computer science and space science, remote sensing technology particularly remotely sensed image has obtained rapid development.Image spatial resolution is a critical index of remote sensing image quality evaluation, is also a parameter very important in remotely sensed image application, most important in the acquisition and application of image.Compared with traditional middle low resolution remote sensing images, high-resolution remote sensing image can clearly express feature distribution and the space correlation of ground object target, distinguishable go out ground object target internal more detailed structure composition, provide good condition and basis for interpretation analysis.In recent years, the advantage of high-resolution remote sensing image becomes apparent from, and oneself, through being widely used in surveying and mapping, transport development, resource environment, farming and forestry, the military every field such as interpretation, disaster monitoring, becomes indispensable important means.

But, in digital picture acquisition process, there is several factors can cause the decline of image resolution ratio, the anamorphose that for example relative motion between imaging device and subject causes, it is image blurring that atmospheric perturbation, imaging device optics cause, the frequency alias that causes of sampling, and imaging, the noise etc. introducing in transmitting procedure.In order to overcome the restriction of manufacturing process and cost, the method utilizing Digital Signal Processing to improve image resolution ratio increasingly receives significant attention, and related method is referred to as super-resolution (Super Resolution is called for short SR) Image Reconstruction.

Existing super-resolution method generally comprises three kinds: based on method, the method based on the method rebuild with based on study of interpolation.Method based on method, the method for neighborhood embedded (neighbor embedding), support vector regression method (SVR) and rarefaction representation that the super-resolution method of study has based on sample (Example-based) than more typical.But as Image Reconstruction is more and more higher to the requirement of resolution ratio, the reconstruct of the remote sensing images in, texture information more rich, different type of ground objects geomorphic feature bigger to observation scope for these existing methods can not meet requirement.

These three super-resolution method all uses general dictionary to be reconstructed.Owing to the observation scope of remote sensing images is bigger, texture information type of ground objects geomorphic feature more rich, different is enriched, so for reaching certain quality reconstruction, the scale of dictionary will increase to ensure the number of effective atom, causes reconstructed velocity excessively slow.

Content of the invention

For solving defect or the deficiency that prior art exists, it is contemplated that for the feature that remote sensing images observation scope is bigger, texture information type of ground objects geomorphic feature more rich, different is abundant, it is proposed that a kind of Remote Sensing Image Super Resolution method based on different type of ground objects classification learning dictionaries.

The above-mentioned purpose of the present invention is realized by the technical characteristic of independent claims, and dependent claims develops the technical characteristic of independent claims in the way of selecting else or be favourable.

For reaching above-mentioned purpose, the technical solution adopted in the present invention is as follows:

Remote Sensing Image Super Resolution method based on dictionary learning, comprising:

Step 1: for the image library of different types of ground objects, carry out the dictionary learning of corresponding type of ground objects；

Step 2, type of ground objects identification is carried out to original image；And

Step 3, use species type-word allusion quotation accordingly, carry out image super-resolution restructuring procedure.

In further embodiment, the dictionary learning of described step 1 comprises the steps:

11) the high-resolution training image of input is pre-processed；

12) carry out feature extraction to pretreated image, build low-resolution image block collection；

13) the low-resolution image block collection of each classification is trained, obtains low-resolution dictionary and the coefficient vector of corresponding training image blocks rarefaction representation, and obtain high-resolution dictionary accordingly；

In further embodiment, described step 11) the high-resolution training image inputting is pre-processed, its realization includes:

111) for input high-resolution training imageObtain corresponding low-resolution image according to following formula (1)This process is Interpolation Process:

$y_l={\mathrm{Qz}}_l=Q({\mathrm{SHy}}_h+v)=L^{\mathrm{all}}{y}_h+\tilde{v}---\left(1\right)$

WhereinFor bicubic interpolation operator,For low pass filter,For down-sampling operator, v～N (0, σ²I) representing that v belongs to average is 0, and variance is σ²The Gaussian Profile of I；

112) according to the output in (1) formulaRemove low-frequency information by formula (2):

$e_h^j=y_h^j-y_l^j---\left(2\right)$

113) pretreatment terminates.

In further embodiment, aforementioned 12) feature extraction is carried out to pretreated image, build low-resolution image block collection, its realization includes:

121) utilize following formula (3), use R high-pass filter, extract in image the corresponding local feature with high-frequency information so that each low-resolution imageAll obtain a filtering image set:

$\{f_r*y_l^j\},r=\mathrm{1,2},...,R---\left(3\right)$

Wherein * is convolution；

122) data set that topography's block is constituted is obtainedBy formula (4) to the low-resolution image block use principal component analysis method dimensionality reduction inputting:

$p_l^k=B{\tilde{p}}_l^k---\left(4\right)$

WhereinFor projection operator；

Wherein, this step 122) include step in detail below:

The data set that topography's block is constitutedIn,Obtained by (2) formulaIn position, k extracts, It is by filtering imageExtract in same position, and the size of high-resolution block covers high-definition picture.

In further embodiment, aforementioned 13) being trained the low-resolution image block collection of each classification, obtaining low-resolution dictionary and the coefficient vector of corresponding training image blocks rarefaction representation, and obtain high-resolution dictionary accordingly, it implements and includes:

131) the low-resolution image block collection that the training set of images of each classification is pre-processed and obtains after feature extraction by following formula (5) is utilizedCarry out K-SVD training, obtain low-resolution dictionary D_LWith corresponding training image blocksCoefficient vector { the α of rarefaction representation^k}_k:

$D_L,\left\{{\mathrm{\α}}^k\right\}=\arg \min \underset{k}{\mathrm{\Σ}}{\left|\right|p_l^k-D_L{\mathrm{\α}}^k\left|\right|}_2^2---\left(5\right)$

Wherein α^k ₀≤ L,It is two norm operators；

132) following formula (6) is utilized to obtain high-resolution dictionary D_H:

$D_H=\arg \min \underset{k}{\mathrm{\Σ}}{\left|\right|p_h^k-D_H{\mathrm{\α}}^k\left|\right|}_2^2---\left(6\right)$

133) it is derived from the associating dictionary of corresponding classification to { D_H,D_L}_k, wherein k is corresponding classification.

In further embodiment, the original image type of ground objects identification of described step, comprise the following steps:

21) the LBP feature of every piece image in each classification image library is extracted；

22) the LBP feature to all extractions for the support vector machine method is used to carry out classification based training；

23) extract the LBP feature of the interpolation image of pending low-resolution image, put into and grader carries out Classification and Identification, determine the type of ground objects of image.

In further embodiment, the image super-resolution reconstruct of described step 3, comprise the following steps:

31) according to step 23) in obtain the type of ground objects of pending image after, obtain corresponding associating dictionary to { D_H,D_L}_k, wherein k represents dictionary to affiliated classification；

32) according to abovementioned steps 122) method extract image block collection

33) use orthogonal matching pursuit method to choose L atom, build rarefaction representation vector, estimate high-resolution block；

34) high-definition picture estimating to reconstruct is obtained.

The present invention utilizes the thought of terrain classification, it is proposed that use the super-resolution method based on classifying dictionary.Owing to the number of atom in the dictionary that restructuring procedure uses is greatly reduced, so dictionary scale is greatly reduced.Because effective atomic number purpose ratio is significantly increased in the classifying dictionary using, dictionary scale is greatly reduced simultaneously, so using effective atomic number mesh in dictionary almost identical with general dictionary at restructuring procedure, it is ensured that while reconstruction quality, reconstructed velocity is greatly improved.This process relates only to the regulation of parameter, and computational efficiency is high, and the reconstruction quality for remote sensing images is high.Use PCA dimensionality reduction and KSVD method training dictionary simultaneously during dictionary learning, improve the efficiency of dictionary learning；Use the greedy iteration of OMP algorithm in restructuring procedure, improve reconstruction quality and efficiency.

As long as it should be appreciated that all parts combining the subject matter that can be viewed as the disclosure in the case that such design is not conflicting of aforementioned concepts and the extra design describing in greater detail below.In addition, all combinations of theme required for protection are considered as a part for the subject matter of the disclosure.

Foregoing and other aspect, embodiment and the feature that present invention teach that can be more fully appreciated with in conjunction with accompanying drawing from the following description.The feature of other additional aspect such as illustrative embodiments of the present invention and/or beneficial effect will be obvious in the following description, or learn in the practice according to the detailed description of the invention that present invention teach that.

Brief description

Accompanying drawing is not intended to drawn to scale.In the accompanying drawings, each illustrating in each figure is identical or approximately uniform part can be indicated by the same numeral.For clarity, in each figure, not each part is all labeled.Now, by by example the embodiment that various aspects of the invention are described in reference to the drawings, wherein:

Fig. 1 is some embodiment illustrating according to the present invention, based on the flow chart of the Remote Sensing Image Super Resolution method of classifying dictionary.

Fig. 2 is some embodiment illustrating according to the present invention, the flow chart of dictionary learning.

Fig. 3 is some embodiment illustrating according to the present invention, the flow chart of Image Reconstruction.

Fig. 4 is some embodiment illustrating according to the present invention, based on a schematic diagram implementing example of the Remote Sensing Image Super Resolution method of classifying dictionary.

Detailed description of the invention

In order to know more about the technology contents of the present invention, especially exemplified by specific embodiment and coordinate institute's accompanying drawings to be described as follows.

Each side with reference to the accompanying drawings to describe the present invention in the disclosure, the embodiment illustrating shown in the drawings of many.Embodiment of the disclosure and must not be intended to include all aspects of the invention.Should be understood, multiple design presented hereinbefore and embodiment, and describe in more detail below those design and embodiment can in many ways in any one is implemented, this is because disclosed in this invention design and embodiment be not limited to any embodiment.In addition, some aspects disclosed by the invention can be used alone, or otherwise any appropriately combined use with disclosed by the invention.

The disclosure utilizes the thought of terrain classification, proposes a kind of super-resolution method using based on classifying dictionary, and in conjunction with shown in Fig. 1, it implements and includes:

Step 1: for the image library of different types of ground objects, carry out the dictionary learning of corresponding type of ground objects；

Step 2, type of ground objects identification is carried out to original image；And

Step 3, use species type-word allusion quotation accordingly, carry out image super-resolution restructuring procedure.

By selecting different dictionary targetedly, owing to the ratio of atom effective in these dictionaries is relatively large, dictionary scale is less simultaneously, this ensures that theres quality reconstruction, and reconstructed velocity is greatly improved simultaneously.

2-3 illustrates the exemplary realization of each step aforementioned below in conjunction with the accompanying drawings.

In order to make it easy to understand, the character/parameter declaration in the disclosure is as follows:

The remote sensing images training set of inputWherein k=1,2 ... N represents picture number in image set, h represent that image is high-definition picture, the corresponding low-resolution image collection obtainingWherein l represents that image is low-resolution image.The data set that topography's block is constitutedWhereinIt is the high-resolution block of j-th image,It is the low resolution block of j-th image.Represent j-th image remove low-frequency information after remaining high-frequency information.{f_rIt is high-pass filter set, wherein r is the numbering of high-pass filter.{q^k}_kBlock rarefaction representation for correspondence image block position k is vectorial,High-resolution block dictionary for correspondence image block position k.The high-resolution block estimated for correspondence image block position k.

In conjunction with shown in Fig. 2, the dictionary learning of described step 1 comprises the steps:

11) the high-resolution training image of input is pre-processed；

12) carry out feature extraction to pretreated image, build low-resolution image block collection；

13) the low-resolution image block collection of each classification is trained, obtains low-resolution dictionary and the coefficient vector of corresponding training image blocks rarefaction representation, and obtain high-resolution dictionary accordingly.

Specifically, described step 11) the high-resolution training image inputting is pre-processed, its realization includes:

111) for input high-resolution training imageObtain corresponding low-resolution image according to following formula (1)This process is Interpolation Process (this process is referred to as Interpolation Process):

$y_l={\mathrm{Qz}}_l=Q({\mathrm{SHy}}_h+v)=L^{\mathrm{all}}{y}_h+\tilde{v}---\left(1\right)$

WhereinFor bicubic interpolation operator,For low pass filter,For down-sampling operator, v～N (0, σ²I) representing that v belongs to average is 0, and variance is σ²The Gaussian Profile of I；

112) according to the output in (1) formulaRemove low-frequency information by formula (2):

$e_h^j=y_h^j-y_l^j---\left(2\right)$

113) pretreatment terminates.

Preferably, aforementioned 12) carrying out feature extraction to pretreated image, building low-resolution image block collection, its realization includes:

121) utilize following formula (3), use R high-pass filter, extract in image the corresponding local feature with high-frequency information so that each low-resolution imageAll obtain a filtering image set:

$\{f_r*y_l^j\},r=\mathrm{1,2},...,R---\left(3\right)$

Wherein * is convolution；

122) data set that topography's block is constituted is obtainedBy formula (4) to the low-resolution image block use principal component analysis method dimensionality reduction inputting:

$p_l^k=B{\tilde{p}}_l^k---\left(4\right)$

WhereinFor projection operator；

Wherein, this step 122) include step in detail below:

The data set that topography's block is constitutedIn,Obtained by (2) formulaIn position, k extracts, It is by filtering imageExtract in same position, and the size of high-resolution block covers high-definition picture.Here high-resolution block to be ensured is sized to cover high-definition picture.

In further embodiment, aforementioned 13) being trained the low-resolution image block collection of each classification, obtaining low-resolution dictionary and the coefficient vector of corresponding training image blocks rarefaction representation, and obtain high-resolution dictionary accordingly, it implements and includes:

131) the low-resolution image block collection that the training set of images of each classification is pre-processed and obtains after feature extraction by following formula (5) is utilizedCarry out K-SVD training, obtain low-resolution dictionary D_LWith corresponding training image blocksCoefficient vector { the α of rarefaction representation^k}_k:

$D_L,\left\{{\mathrm{\α}}^k\right\}=\arg \min \underset{k}{\mathrm{\Σ}}{\left|\right|p_l^k-D_L{\mathrm{\α}}^k\left|\right|}_2^2---\left(5\right)$

Wherein α^k ₀≤ L,It is two norm operators；

132) following formula (6) is utilized to obtain high-resolution dictionary D_H:

$D_H=\arg \min \underset{k}{\mathrm{\Σ}}{\left|\right|p_h^k-D_H{\mathrm{\α}}^k\left|\right|}_2^2---\left(6\right)$

133) it is derived from the associating dictionary of corresponding classification to { D_H,D_L}_k, wherein k is corresponding classification.

In conjunction with shown in Fig. 3, the original image type of ground objects identification of described step, comprise the following steps:

21) the LBP feature of every piece image in each classification image library is extracted；

22) the LBP feature to all extractions for the support vector machine method (SVM) is used to carry out classification based training；

23) extract the LBP feature of the interpolation image of pending low-resolution image, put into and grader carries out Classification and Identification, determine the type of ground objects of image.

The image super-resolution reconstruct of described step 3, comprises the following steps:

31) according to step 23) in obtain the type of ground objects of pending image after, obtain corresponding associating dictionary to { D_H,D_L}_k, wherein k represents dictionary to affiliated classification；

32) according to abovementioned steps 122) method extract image block collection

33) use orthogonal matching pursuit method to choose L atom, build rarefaction representation vector, estimate high-resolution block；

34) high-definition picture estimating to reconstruct is obtained.

Fig. 4 illustrates the concrete implementation process utilizing preceding method to carry out Remote Sensing Image Super Resolution method.Table 1 below has been listed and has been embodied as parameter in some embodiments.Visible, owing to the number of atom in the dictionary that restructuring procedure uses is greatly reduced, so dictionary scale is greatly reduced.Because effective atomic number purpose ratio is significantly increased in the classifying dictionary using, dictionary scale is greatly reduced simultaneously, so using effective atomic number mesh in dictionary almost identical with general dictionary at restructuring procedure, it is ensured that while reconstruction quality, reconstructed velocity is greatly improved.This process relates only to the regulation of parameter, and computational efficiency is high, and the reconstruction quality for remote sensing images is high.

Table 1

Although the present invention is disclosed above with preferred embodiment, so it is not limited to the present invention.Persond having ordinary knowledge in the technical field of the present invention, without departing from the spirit and scope of the present invention, when can be used for a variety of modifications and variations.Therefore, protection scope of the present invention ought be as the criterion depending on those as defined in claim.

Claims (7)

1. the Remote Sensing Image Super Resolution method based on dictionary learning, it is characterised in that this Remote Sensing Image Super Resolution method includes:

Step 1: for the image library of different types of ground objects, carry out the dictionary learning of corresponding type of ground objects；

Step 2, type of ground objects identification is carried out to original image；And

Step 3, use species type-word allusion quotation accordingly, carry out image super-resolution restructuring procedure.

2. the Remote Sensing Image Super Resolution method based on dictionary learning according to claim 1, it is characterised in that the dictionary learning of described step 1 comprises the steps:

11) the high-resolution training image of input is pre-processed；

12) carry out feature extraction to pretreated image, build low-resolution image block collection；

13) the low-resolution image block collection of each classification is trained, obtains low-resolution dictionary and the coefficient vector of corresponding training image blocks rarefaction representation, and obtain high-resolution dictionary accordingly.

3. the Remote Sensing Image Super Resolution method based on dictionary learning according to claim 2, it is characterised in that described step 11) the high-resolution training image inputting is pre-processed, its realization includes:

111) for input high-resolution training imageObtain corresponding low-resolution image according to following formula (1)This process is Interpolation Process:

WhereinFor bicubic interpolation operator,For low pass filter,For down-sampling operator, v～N (0, σ²I) representing that v belongs to average is 0, and variance is σ²The Gaussian Profile of I；

112) according to the output in (1) formulaRemove low-frequency information by formula (2):

113) pretreatment terminates.

4. the Remote Sensing Image Super Resolution method based on dictionary learning according to claim 3, it is characterised in that aforementioned 12) feature extraction is carried out to pretreated image, build low-resolution image block collection, its realization includes:

121) utilize following formula (3), use R high-pass filter, extract in image the corresponding local feature with high-frequency information so that each low-resolution imageAll obtain a filtering image set:

Wherein * is convolution；

122) data set that topography's block is constituted is obtainedBy formula (4) to the low-resolution image block use principal component analysis method dimensionality reduction inputting:

WhereinFor projection operator；

Wherein, this step 122) include step in detail below:

The data set that topography's block is constitutedIn,Obtained by (2) formulaIn position, k extracts,It is by filtering imageExtract in same position, and the size of high-resolution block covers high-definition picture.

5. the Remote Sensing Image Super Resolution method based on dictionary learning according to claim 4, it is characterized in that, aforementioned 13) the low-resolution image block collection of each classification is trained, obtain low-resolution dictionary and the coefficient vector of corresponding training image blocks rarefaction representation, and obtain high-resolution dictionary accordingly, it implements and includes:

131) the low-resolution image block collection that the training set of images of each classification is pre-processed and obtains after feature extraction by following formula (5) is utilizedCarry out K-SVD training, obtain low-resolution dictionary D_LWith corresponding training image blocksCoefficient vector { the α of rarefaction representation^k}_k:

Wherein α^k ₀≤ L,It is two norm operators；

132) following formula (6) is utilized to obtain high-resolution dictionary D_H:

133) it is derived from the associating dictionary of corresponding classification to { D_H,D_L}_k, wherein k is corresponding classification.

6. the Remote Sensing Image Super Resolution method based on dictionary learning according to claim 5, it is characterised in that the original image type of ground objects identification of described step, comprises the following steps:

21) the LBP feature of every piece image in each classification image library is extracted；

22) the LBP feature to all extractions for the support vector machine method is used to carry out classification based training；

23) extract the LBP feature of the interpolation image of pending low-resolution image, put into and grader carries out Classification and Identification, determine the type of ground objects of image.

7. the Remote Sensing Image Super Resolution method based on dictionary learning according to claim 6, it is characterised in that the image super-resolution reconstruct of described step 3, comprises the following steps:

31) according to step 23) in obtain the type of ground objects of pending image after, obtain corresponding associating dictionary to { D_H,D_L}_k, wherein k represents dictionary to affiliated classification；

32) according to abovementioned steps 122) method extract image block collection

33) use orthogonal matching pursuit method to choose L atom, build rarefaction representation vector, estimate high-resolution block；

34) high-definition picture estimating to reconstruct is obtained.