CN104580829A - Terahertz image enhancing method and system - Google Patents

Abstract

The invention relates to a terahertz image enhancing method and system. The method comprises steps as follows: median filtering and noise reduction, non-local filtering, frequency-domain high-pass filtering, edge processing and overlapping processing are performed on a terahertz image, the adopted non-local filtering is different from domain filtering adopted in some methods, speckle noises can be filtered out to a certain extent through domain filtering, but edge information is more fuzzy, the non-local filtering is more resistant to the noises, and the filtered-out part comprises little geometrical structure information.

Description

A kind of Terahertz image enchancing method and system

Technical field

The present invention relates to a kind of Terahertz image enchancing method and system, particularly relate to a kind of disposal route and system of Terahertz image enhaucament.

Background technology

The reason affecting Terahertz image resolution ratio and sharpness has a lot, and one of them is exactly because laser power shake causes terahertz emission intensity to change, thus causes the shake that Terahertz gradation of image distributes, and namely there is obvious speck.And simultaneously may be comparatively dark along with image background, the problem that contrast is not strong.The noise reduction of terahertz imaging and enhancing have many methods.Conventional have based on the noise reduction of wavelet transformation, rim detection and enhancing etc.But these methods are usually directed to more complicated mathematical operation, lack versatility and intuitive.

Summary of the invention

The technical matters that the present invention solves is: build a kind of Terahertz image enchancing method and system, overcome prior art Terahertz image enhaucament and be usually directed to more complicated mathematical operation, lacks versatility and intuitive.

Technical scheme of the present invention is: provide a kind of Terahertz image enchancing method, comprise the steps:

Medium filtering noise reduction: first medium filtering is carried out to Terahertz original image, then linear gradation stretching is done in the tonal range of 0-255 to image;

Non local filtering: the estimated value of being tried to achieve pixel by the weighted mean value of total space territory pixel, obtains the similarity between two pixels, is then weighted on average to it;

Image enhaucament: the covariance matrix and the orthogonal matrix that build image vector, the image vector that discrete principal component analysis is enhanced is carried out to image vector, specifically comprise: discrete principal component analysis method is by the proper vector corresponding to the larger eigenwert of part, then carries out discrete principal component analysis (PCA) inverse transformation to described feature especially vector and carry out image enhaucament;

Edge treated: adopt horizontal and vertical operator to carry out edge treated to the image after non-local filtering process;

Overlap-add procedure: superposed by the image of the image after edge treated with second order high-pass filtering process, carries out image sharpening by the Terahertz image after superposition, is finally processed image.

Further technical scheme of the present invention is: in non local filter step, comprises and determines search window, similarity window and filtering depth parameter.

Further technical scheme of the present invention is: in non local filter step, and the similarity between two pixels is according to the similar retrieval between gray scale vector.

Further technical scheme of the present invention is: the similarity between gray scale vector is represented by the decreasing function of weighted euclidean distance.

Further technical scheme of the present invention is: similarity window is centered by pixel, the square field of fixed size.

Technical scheme of the present invention is: build a kind of Terahertz Image Intensified System, comprise medium filtering noise reduction module, non local filtration module, image enhancement module, image edge processing module, imaging importing module, described medium filtering noise reduction module carries out medium filtering to Terahertz original image, again to image 0 ?255 tonal range in do linear gradation stretch, described non local filtration module tries to achieve the estimated value of pixel by the weighted mean value of total space territory pixel, obtain the similarity between two pixels, then be weighted on average to it, described image enhancement module builds covariance matrix and the orthogonal matrix of image vector, discrete principal component analysis is carried out by the proper vector corresponding to the larger eigenwert of part, again discrete principal component analysis (PCA) inverse transformation is carried out to described feature especially vector and carry out image enhaucament, described image edge processing module adopts horizontal and vertical operator to carry out edge treated to the image after non-local filtering process, the image of image after edge treated with second order high-pass filtering process superposes by described imaging importing module, Terahertz image after superposition is carried out image sharpening, is finally processed image.

Further technical scheme of the present invention is: comprise weight factor determination module, and described weight factor determination module is by the pixel determination weight in the vectorial similar gray scale field of gray scale.

Further technical scheme of the present invention is: the Euclidean distance comprising the Euclidean distance expectation value between the noise pixel point obtaining image expects module.

Further technical scheme of the present invention is: described horizontal and vertical operator comprise in Roberts, Prewitt or Sobel operator one or more.

Further technical scheme of the present invention is: described in carry out image sharpening operator comprise in Roberts, Prewitt or Sobel operator one or more.

Technique effect of the present invention is: build a kind of Terahertz image enchancing method and system, comprise medium filtering noise reduction: first carry out medium filtering to Terahertz original image, then in the tonal range of 0-255, does linear gradation stretching to image; Non local filtering: the estimated value of being tried to achieve pixel by the weighted mean value of total space territory pixel, obtains the similarity between two pixels, is then weighted on average to it; Image enhaucament: the covariance matrix and the orthogonal matrix that build image vector, the image vector that discrete principal component analysis is enhanced is carried out to image vector, specifically comprise: discrete principal component analysis method is by the proper vector corresponding to the larger eigenwert of part, then carries out discrete principal component analysis (PCA) inverse transformation to described feature especially vector and carry out image enhaucament; Edge treated: adopt horizontal and vertical operator to carry out edge treated to the image after non-local filtering process; Overlap-add procedure: superposed by the image of the image after edge treated with second order high-pass filtering process, carries out image sharpening by the Terahertz image after superposition, is finally processed image.Terahertz image enchancing method of the present invention and system, adopt the image enchancing method of non-local filtering and principal component analysis (PCA), the non-local filtering adopted is different from the field filtering adopted in some method, though field filtering can filtering speckle noise to a certain extent, but marginal information is fuzzyyer, but not part filter has more repellence to noise, the geometry information contained in filtering part is less.

Accompanying drawing explanation

Fig. 1 is process flow diagram of the present invention.

Fig. 2 is structural representation of the present invention.

Embodiment

Below in conjunction with specific embodiment, technical solution of the present invention is further illustrated.

As shown in Figure 1, the specific embodiment of the present invention is: provide a kind of Terahertz image enchancing method, comprise the steps:

Medium filtering noise reduction: first medium filtering is carried out to Terahertz original image, then linear gradation stretching is done in the tonal range of 0-255 to image.

Specific implementation process is as follows: medium filtering is a kind of conventional nonlinear smoothing filtering, and its ultimate principle is that the Mesophyticum of each point value in a field of this point of value of any in digital picture is replaced.If f (x, y) is the gray-scale value of image slices vegetarian refreshments, filter window is that the medium filtering of A is defined as:

f^(x,y)＝MED{f(x,y)}(x,y)∈A (1)

In the tonal range of 0-255, do linear gradation afterwards again stretch, obtain the image of contrast strengthen.

Non local filtering: the estimated value of being tried to achieve pixel by the weighted mean value of total space territory pixel, obtains the similarity between two pixels, is then weighted on average it.

Specific implementation process is as follows: refer to that the gray-scale value of current pixel point is obtained by the gray-scale value weighted mean of the total space territory pixel similar to its structure, weight depends on structural similarity degree.Suppose given discrete by digital picture v={v (the i) ∣ i ∈ I} of noise pollution, can be tried to achieve by the weighted mean of total space territory pixel the estimated value NL [v] (i) of pixel i:

NL[v](i)＝Σw(i，j)v(j) (2)

Weight { w (i, j) } _jdepend on the similarity of pixel i and j, and meet:

0≤w(i，j)≤1；

Σ _jw(i，j)＝1. (3)

Similarity between two pixel i and j depends on gray scale vector v (N _i) and v (N _j) between similarity.N _krepresent the square field being centrally located at the fixed size of k.This similarity is by weighted euclidean distance ‖ v (N _i)-v (N _j) ‖ ² _{2, a}decreasing function represent.Wherein a is the standard deviation of gaussian kernel.Euclidean distance expectation value between the noise pixel point of image can be tried to achieve by following formula:

$E{\left|\right|v\left(N_i\right)-v\left(N_j\right)\left|\right|}_{2,a}^2={\left|\right|u\left(N_i\right)-u\left(N_j\right)\left|\right|}_{2,a}^2+2{\mathrm{\σ}}^2:---\left(4\right)$

The pass of v and u is: v=u+n, v are image pixel observed readings, and u is image actual value, and n is the noise of superposition.σ is the standard deviation of the spacing of two gray scale vectors.The expectation of this Euclidean distance maintains the similarity between different pixels point.With v (N _i) pixel in similar gray scale field has larger weight generally, defined by following formula:

$w(i,j)=\frac{1}{Z\left(i\right)}{e}^{-\frac{{\left|\right|v\left(N_i\right)-v\left(N_j\right)\left|\right|}_{2,a}^2}{h^2}}---\left(5\right)$

Normaliztion constant factor Z (i) is defined as:

$Z\left(i\right)=\underset{j}{\mathrm{\Σ}}{e}^{-\frac{{\left|\right|v\left(N_i\right)-v\left(N_j\right)\left|\right|}_{2,a}^2}{h^2}}---\left(6\right)$

Wherein h represents filter strength, the decay of control characteristic function, or the rate of decay of the further control weight factor.

Be generally convenience of calculation, N _iget centered by pixel i, the square field of fixed size (2m+1) × (2m+1), w (i, j) and Z (i) can be expressed as:

$w(i,j)=\frac{1}{G\left(i\right)}\exp \left[\frac{\underset{n_i\∈N_i,n_j\∈N_j,k_i\∈k}{\mathrm{\Σ}}{k}_i{(n_i-n_j)}^2}{h^2}\right]---\left(7\right)$

$Z\left(i\right)=\underset{j}{\mathrm{\Σ}}\exp \left[\frac{\underset{n_i\∈N_i,n_j\∈N_j,k_i\∈k}{\mathrm{\Σ}}{k}_i{(n_i-n_j)}^2}{h^2}\right]---\left(8\right)$

$k_i=\frac{1}{m}\underset{d=d_i}{\overset{m}{\mathrm{\Σ}}}\frac{1}{{(2d+1)}^2}---\left(9\right)$

The geometry in what non local filtering was compared the is whole field of two single-points, so have more repellence to noise, and the part leached contains less geometry information.

Image enhaucament: the covariance matrix and the orthogonal matrix that build image vector, the image vector that discrete principal component analysis is enhanced is carried out to image vector, specifically comprise: discrete principal component analysis method is by the proper vector corresponding to the larger eigenwert of part, then carries out discrete principal component analysis (PCA) inverse transformation to described feature especially vector and carry out image enhaucament.

Specific implementation process is as follows: principal component analysis (PCA) (Principal Component Analysis, principal component analysis (PCA), be called for short " PCA ") be a kind of image conversion, pass through principal component analysis, most important element and structure can be gone out from extracting data confusing in a large number, thus remove noise.The covariance matrix of the image vector i that filter process is crossed is defined as:

c ₁＝E[(1-m ₁)(1-m ₁) ^T]

λ ₁>=λ ₂>=...>=λ _n ²covariance matrix C _ieigenwert, characteristic of correspondence vector is b _i, constitute N ²× N ²orthogonal matrix B:

$B=\left[\begin{array}{c}{b}_1^T\\ b_2^T\\ \·\\ \·\\ \·\\ b_{N^2}^T\end{array}\right]=\left[\begin{array}{cccccc}{b}_{11}& b_{12}& \·& \·& \·& b_{1N^2}\\ b_{21}& b_{22}& \·& \·& \·& b_{2N^2}\\ \·& \·& & & & \·\\ \·& \·& & & & \·\\ \·& \·& & & & \·\\ b_{N^{2}1}& b_{N^{2}2}& \·& \·& \·& b_{N^2{N}^2}\end{array}\right]$

The mathematic(al) representation of discrete PCA conversion:

The mathematic(al) representation of discrete PCA inverse transformation:

A is Orthogonal Symmetric transformation matrix, is the normalized form of matrix B.I, g are the image vector after the image crossed of filter process and conversion respectively.M _iit is the mean vector of i.

Be discrete PCA to i to convert, retain the individual larger eigenvalue λ of k above ₁>=λ ₂>=...>=λ _k, N after removal ²the individual less eigenwert of-k try to achieve the proper vector corresponding to eigenwert of front k, do discrete PCA inverse transformation, one that just obtains image vector i is similar to removed information, in less eigenwert characteristic of correspondence vector, does not exist in. it is the primary structure after original image vector i simplifies.Noise is normally represented by less eigenwert, and removing less eigenwert can reach except effect of making an uproar, but edge and texture are also contained in less eigenwert characteristic of correspondence vector equally, so keeping edge and reducing the balance will gone between noise.

Edge treated: adopt Roberts edge detection operator to realize horizontal and vertical direction and edge treated is carried out to the image after non-local filtering process.If f (x, y) is gradation of image distribution function, then its Reberts edge detection operator is

$g(x,y)={\{{[\sqrt{f(x,y)}-\sqrt{f(x+1,y+1)}]}^2+{[\sqrt{f(x+1,y)}-\sqrt{f(x,y+1)}]}^2\}}^{\frac{1}{2}}$

Roberts edge detection operator realizes the rim detection in horizontal and vertical direction respectively, and operational form is:

$\left\{\begin{array}{c}{\mathrm{\Δ}}_{x}f(x,y)=f(x,y)-f(x-1,y-1)\\ {\mathrm{\Δ}}_{y}f(x,y)=f(x-1,y)-f(x,y-1)\end{array}\right.$

Overlap-add procedure: the image of the image after edge treated with second order high-pass filtering process superposes by the image registration based on half-tone information method, concrete grammar is definition benchmark image I (x, y) with template image T (x, y), make template image move on benchmark image, and calculate similarity degree between the two, namely the place that peak value occurs is registration position, calculating formula of similarity on each displacement point (i, j) determined is

$D(i,j)=\frac{{\mathrm{\Σ}}_x{\mathrm{\Σ}}_{y}T(x,y)I(x-i,y-j)}{\sqrt{{\mathrm{\Σ}}_x{\mathrm{\Σ}}_y{I}^2(x-i,y-j)}}$

Image sharpening: the Terahertz image after superposition is carried out image sharpening, and utilize Roberts operator to carry out sharpening, Roberts operator template is the template of a 2*2, and for current pending pixel f (x, y), Roberts operator definitions is as follows:

$\▿f=|f(x+1.y+1)-f(x,y)|+|f(x+1,y)-f(x,y+1)|$

Specifically being expressed as of template

$D_1=\left[\begin{array}{cc}-1& 0\\ 0& 1\end{array}\right]D_2=\left[\begin{array}{cc}0& -1\\ 1& 0\end{array}\right]$

ξ ₁＝D ₁(f(x,y)) ξ ₂＝D ₂(f(x,y))

$\▿f(x,y)=\left|{\mathrm{\ξ}}_1\right|+\left|{\mathrm{\ξ}}_2\right|$

Finally processed image.

As shown in Figure 2, the specific embodiment of the present invention is: build a kind of Terahertz Image Intensified System, comprise medium filtering noise reduction module 1, non local filtration module 2, image enhancement module 3, image edge processing module 4, imaging importing module 5, described medium filtering noise reduction module 1 pair of Terahertz original image carries out medium filtering, again to image 0 ?255 tonal range in do linear gradation stretch, described non local filtration module 2 tries to achieve the estimated value of pixel by the weighted mean value of total space territory pixel, obtain the similarity between two pixels, then be weighted on average to it, image enhancement module 3 builds covariance matrix and the orthogonal matrix of image vector, discrete principal component analysis is carried out by the proper vector corresponding to the larger eigenwert of part, again discrete principal component analysis (PCA) inverse transformation is carried out to described feature especially vector and carry out image enhaucament, described image edge processing module 4 adopts horizontal and vertical operator to carry out edge treated to the image after non-local filtering process, the image of image after edge treated with second order high-pass filtering process superposes by described imaging importing module 5, Terahertz image after superposition is carried out image sharpening, is finally processed image.

As shown in Figure 2, specific embodiment of the invention process is: described medium filtering noise reduction module 1 first carries out medium filtering to Terahertz original image, then in the tonal range of 0-255, does linear gradation stretching to image.

Specific implementation process is as follows: medium filtering is a kind of conventional nonlinear smoothing filtering, and its ultimate principle is that the Mesophyticum of each point value in a field of this point of value of any in digital picture is replaced.If f (x, y) is the gray-scale value of image slices vegetarian refreshments, filter window is that the medium filtering of A is defined as:

f^(x,y)＝MED{f(x,y)}(x,y)∈A (1)

In the tonal range of 0-255, do linear gradation afterwards again stretch, obtain the image of contrast strengthen.

Non local filtration module 2 tries to achieve the estimated value of pixel by the weighted mean value of total space territory pixel, obtains the similarity between two pixels, is then weighted on average it.

Specific implementation process is as follows: refer to that the gray-scale value of current pixel point is obtained by the gray-scale value weighted mean of the total space territory pixel similar to its structure, weight depends on structural similarity degree.Suppose given discrete by digital picture v={v (the i) ∣ i ∈ I} of noise pollution, can be tried to achieve by the weighted mean of total space territory pixel the estimated value NL [v] (i) of pixel i:

NL[v](i)＝Σw(i，j)v(j) (2)

Weight { w (i, j) } _jdepend on the similarity of pixel i and j, and meet:

0≤w(i，j)≤1；

Σ _jw(i，j)＝1. (3)

Similarity between two pixel i and j depends on gray scale vector v (N _i) and v (N _j) between similarity.N _krepresent the square field being centrally located at the fixed size of k.This similarity is by weighted euclidean distance ‖ v (N _i)-v (N _j) ‖ ² _{2, a}decreasing function represent.Wherein a is the standard deviation of gaussian kernel.Euclidean distance expectation value between the noise pixel point of image can be tried to achieve by following formula:

$E{\left|\right|v\left(N_i\right)-v\left(N_j\right)\left|\right|}_{2,a}^2={\left|\right|u\left(N_i\right)-u\left(N_j\right)\left|\right|}_{2,a}^2+2{\mathrm{\σ}}^2:---\left(4\right)$

The pass of v and u is: v=u+n, v are image pixel observed readings, and u is image actual value, and n is the noise of superposition.σ is the standard deviation of the spacing of two gray scale vectors.The expectation of this Euclidean distance maintains the similarity between different pixels point.With v (N _i) pixel in similar gray scale field has larger weight generally, defined by following formula:

$w(i,j)=\frac{1}{Z\left(i\right)}{e}^{-\frac{{\left|\right|v\left(N_i\right)-v\left(N_j\right)\left|\right|}_{2,a}^2}{h^2}}---\left(5\right)$

Normaliztion constant factor Z (i) is defined as:

$Z\left(i\right)=\underset{j}{\mathrm{\Σ}}{e}^{-\frac{{\left|\right|v\left(N_i\right)-v\left(N_j\right)\left|\right|}_{2,a}^2}{h^2}}---\left(6\right)$

Wherein h represents filter strength, the decay of control characteristic function, or the rate of decay of the further control weight factor.

Be generally convenience of calculation, N _iget centered by pixel i, the square field of fixed size (2m+1) × (2m+1), comprise weight factor determination module 6, described weight factor determination module 6 is by the pixel determination weight in the vectorial similar gray scale field of gray scale, w (i, j) and Z (i) can be expressed as:

$w(i,j)=\frac{1}{G\left(i\right)}\exp \left[\frac{\underset{n_i\∈N_i,n_j\∈N_j,k_i\∈k}{\mathrm{\Σ}}{k}_i{(n_i-n_j)}^2}{h^2}\right]---\left(7\right)$

$Z\left(i\right)=\underset{j}{\mathrm{\Σ}}\exp \left[\frac{\underset{n_i\∈N_i,n_j\∈N_j,k_i\∈k}{\mathrm{\Σ}}{k}_i{(n_i-n_j)}^2}{h^2}\right]---\left(8\right)$

$k_i=\frac{1}{m}\underset{d=d_i}{\overset{m}{\mathrm{\Σ}}}\frac{1}{{(2d+1)}^2}---\left(9\right)$

The geometry in what non local filtering was compared the is whole field of two single-points, so have more repellence to noise, and the part leached contains less geometry information.

Image enhancement module 3 builds covariance matrix and the orthogonal matrix of image vector, the image vector that discrete principal component analysis is enhanced is carried out to image vector, specifically comprise: discrete principal component analysis method is by the proper vector corresponding to the larger eigenwert of part, then carries out discrete principal component analysis (PCA) inverse transformation to described feature especially vector and carry out image enhaucament.

Specific implementation process is as follows: principal component analysis (PCA) (Principal Component Analysis, principal component analysis (PCA), be called for short " PCA ") be a kind of image conversion, pass through principal component analysis, most important element and structure can be gone out from extracting data confusing in a large number, thus remove noise.The covariance matrix of the image vector i that filter process is crossed is defined as:

C ₁＝E[(1-m ₁)(1-m ₁) ^T]

λ ₁>=λ ₂>=...>=λ _n ²covariance matrix C _ieigenwert, characteristic of correspondence vector is b _i, constitute N ²× N ²orthogonal matrix B:

$B=\left[\begin{array}{c}{b}_1^T\\ b_2^T\\ \·\\ \·\\ \·\\ b_{N^2}^T\end{array}\right]=\left[\begin{array}{cccccc}{b}_{11}& b_{12}& \·& \·& \·& b_{1N^2}\\ b_{21}& b_{22}& \·& \·& \·& b_{2N^2}\\ \·& \·& & & & \·\\ \·& \·& & & & \·\\ \·& \·& & & & \·\\ b_{N^{2}1}& b_{N^{2}2}& \·& \·& \·& b_{N^2{N}^2}\end{array}\right]$

The mathematic(al) representation of discrete PCA conversion:

The mathematic(al) representation of discrete PCA inverse transformation:

A is Orthogonal Symmetric transformation matrix, is the normalized form of matrix B.I, g are the image vector after the image crossed of filter process and conversion respectively.M _iit is the mean vector of i.

Be discrete PCA to i to convert, retain the individual larger eigenvalue λ of k above ₁>=λ ₂>=...>=λ _k, N after removal ²the individual less eigenwert of-k try to achieve the proper vector corresponding to eigenwert of front k, do discrete PCA inverse transformation, one that just obtains image vector i is similar to removed information, in less eigenwert characteristic of correspondence vector, does not exist in. it is the primary structure after original image vector i simplifies.Noise is normally represented by less eigenwert, and removing less eigenwert can reach except effect of making an uproar, but edge and texture are also contained in less eigenwert characteristic of correspondence vector equally, so keeping edge and reducing the balance will gone between noise.

Image edge processing module 4 adopts horizontal and vertical operator to carry out edge treated to the image after non-local filtering process.

Edge treated: adopt Roberts edge detection operator to realize horizontal and vertical direction and edge treated is carried out to the image after non-local filtering process.If f (x, y) is gradation of image distribution function, then its Reberts edge detection operator is

$g(x,y)={\{{[\sqrt{f(x,y)}-\sqrt{f(x+1,y+1)}]}^2+{[\sqrt{f(x+1,y)}-\sqrt{f(x,y+1)}]}^2\}}^{\frac{1}{2}}$

Roberts edge detection operator realizes the rim detection in horizontal and vertical direction respectively, and operational form is:

$\left\{\begin{array}{c}{\mathrm{\Δ}}_{x}f(x,y)=f(x,y)-f(x-1,y-1)\\ {\mathrm{\Δ}}_{y}f(x,y)=f(x-1,y)-f(x,y-1)\end{array}\right.$

The image of image after edge treated with second order high-pass filtering process superposes by imaging importing module 5, the Terahertz image after superposition is carried out image sharpening, is finally processed image.

Overlap-add procedure: the image of the image after edge treated with second order high-pass filtering process superposes by the image registration based on half-tone information method, concrete grammar is definition benchmark image I (x, y) with template image T (x, y), make template image move on benchmark image, and calculate similarity degree between the two, namely the place that peak value occurs is registration position, calculating formula of similarity on each displacement point (i, j) determined is

$D(i,j)=\frac{{\mathrm{\Σ}}_x{\mathrm{\Σ}}_{y}T(x,y)I(x-i,y-j)}{\sqrt{{\mathrm{\Σ}}_x{\mathrm{\Σ}}_y{I}^2(x-i,y-j)}}$

Image sharpening: the Terahertz image after superposition is carried out image sharpening, and utilize Roberts operator to carry out sharpening, Roberts operator template is the template of a 2*2, and for current pending pixel f (x, y), Roberts operator definitions is as follows:

$\▿f=|f(x+1.y+1)-f(x,y)|+|f(x+1,y)-f(x,y+1)|$

Specifically being expressed as of template

$D_1=\left[\begin{array}{cc}-1& 0\\ 0& 1\end{array}\right]D_2=\left[\begin{array}{cc}0& -1\\ 1& 0\end{array}\right]$

ξ ₁＝D ₁(f(x,y)) ξ ₂＝D ₂(f(x,y))

$\▿f(x,y)=\left|{\mathrm{\ξ}}_1\right|+\left|{\mathrm{\ξ}}_2\right|$

Finally processed image.

Technique effect of the present invention is: build a kind of Terahertz image enchancing method and system, comprise medium filtering noise reduction: first carry out medium filtering to Terahertz original image, then in the tonal range of 0-255, does linear gradation stretching to image; Non local filtering: the estimated value of being tried to achieve pixel by the weighted mean value of total space territory pixel, obtains the similarity between two pixels, is then weighted on average to it; Image enhaucament: the covariance matrix and the orthogonal matrix that build image vector, the image vector that discrete principal component analysis is enhanced is carried out to image vector, specifically comprise: discrete principal component analysis method is by the proper vector corresponding to the larger eigenwert of part, then carries out discrete principal component analysis (PCA) inverse transformation to described feature especially vector and carry out image enhaucament; Edge treated: adopt horizontal and vertical operator to carry out edge treated to the image after non-local filtering process; Overlap-add procedure: superposed by the image of the image after edge treated with second order high-pass filtering process, carries out image sharpening by the Terahertz image after superposition, is finally processed image.Terahertz image enchancing method of the present invention and system, the non-local filtering adopted is different from the field filtering adopted in some method, though field filtering can filtering speckle noise to a certain extent, but marginal information is fuzzyyer, but not part filter has more repellence to noise, the geometry information contained in filtering part is less.

Above content is in conjunction with concrete preferred implementation further description made for the present invention, can not assert that specific embodiment of the invention is confined to these explanations.For general technical staff of the technical field of the invention, without departing from the inventive concept of the premise, some simple deduction or replace can also be made, all should be considered as belonging to protection scope of the present invention.

Claims (10)

1. a Terahertz image enchancing method, comprises the steps:

Medium filtering noise reduction: first medium filtering is carried out to Terahertz original image, then linear gradation stretching is done in the tonal range of 0-255 to image;

Non local filtering: the estimated value of being tried to achieve pixel by the weighted mean value of total space territory pixel, obtains the similarity between two pixels, is then weighted on average to it;

Image enhaucament: the covariance matrix and the orthogonal matrix that build image vector, the image vector that discrete principal component analysis is enhanced is carried out to image vector, specifically comprise: discrete principal component analysis method is by the proper vector corresponding to the larger eigenwert of part, then carries out discrete principal component analysis (PCA) inverse transformation to described feature especially vector and carry out image enhaucament;

Edge treated: adopt horizontal and vertical operator to carry out edge treated to the image after non-local filtering process;

Overlap-add procedure: superposed by the image of the image after edge treated with second order high-pass filtering process, carries out image sharpening by the Terahertz image after superposition, is finally processed image.

2. Terahertz image enchancing method according to claim 1, is characterized in that, in non local filter step, comprises and determines search window, similarity window and filtering depth parameter.

3. Terahertz image enchancing method according to claim 1, is characterized in that, in non local filter step, the similarity between two pixels is according to the similar retrieval between gray scale vector.

4. according to the Terahertz image enchancing method that claim 3 is stated, it is characterized in that, the similarity between gray scale vector is represented by the decreasing function of weighted euclidean distance.

5. Terahertz image enchancing method according to claim 1, it is characterized in that, the method of image vector being carried out to the image vector that discrete principal component analysis is enhanced comprises: retain multiple larger eigenwert above, multiple less eigenwert is left after removal, try to achieve described multiple proper vector corresponding to eigenwert above, do discrete principal component analysis (PCA) inverse transformation, obtain an approximate value of image vector.

6. a Terahertz Image Intensified System, it is characterized in that, comprise medium filtering noise reduction module, non local filtration module, image enhancement module, image edge processing module, imaging importing module, described medium filtering noise reduction module carries out medium filtering to Terahertz original image, in the tonal range of 0-255, do linear gradation to image again to stretch, described non local filtration module tries to achieve the estimated value of pixel by the weighted mean value of total space territory pixel, obtain the similarity between two pixels, then be weighted on average to it, described image enhancement module builds covariance matrix and the orthogonal matrix of image vector, discrete principal component analysis is carried out by the proper vector corresponding to the larger eigenwert of part, again discrete principal component analysis (PCA) inverse transformation is carried out to described feature especially vector and carry out image enhaucament, described image edge processing module adopts horizontal and vertical operator to carry out edge treated to the image after non-local filtering process, the image of image after edge treated with second order high-pass filtering process superposes by described imaging importing module, Terahertz image after superposition is carried out image sharpening, is finally processed image.

7. Terahertz Image Intensified System according to claim 6, is characterized in that, comprise weight factor determination module, and described weight factor determination module is by the pixel determination weight in the similar gray scale field of gray scale vector.

8. Terahertz Image Intensified System according to claim 6, is characterized in that, the Euclidean distance comprising the Euclidean distance expectation value between the noise pixel point obtaining image expects module.

9. Terahertz Image Intensified System according to claim 6, is characterized in that, described horizontal and vertical operator comprise in Roberts, Prewitt or Sobel operator one or more.

10. Terahertz Image Intensified System according to claim 6, is characterized in that, described in carry out image sharpening operator comprise in Roberts, Prewitt or Sobel operator one or more.