CN105426912A - Blind separation method for replacement aliasing image - Google Patents

Abstract

The invention relates to a blind separation method for a replacement aliasing image. A replacement region constructing the replacement aliasing image contains noise; a feature of an image sequence is extracted by utilizing a noise reduction automatic encoder, and then the replacement aliasing image is reconstructed by utilizing the extracted feature to obtain a reconstructed image sequence; and the reconstructed image sequence and the image sequence are subjected to quotient operation and then the reconstructed image sequence is subjected to thresholding operation, so that a replacement region image is separated out. According to the method, the shortcoming of difficulty in selecting a feature domain in the prior art is overcome, the phenomenon of false detection is avoided, and the accuracy of a blind separation result of the aliasing image is improved.

Description

A kind of blind separating method of replacing aliased image

Technical field

The present invention relates to a kind of blind separating method of replacing aliased image, belong to the field of blind source separating method processing digital images.

Background technology

Blind source separating (BBS) is a tradition and have challenging problem in signal transacting.Namely blind source separating is comformed when the source signal of observation signal all cannot be known with mixture model in polyhybird signal and is separated by different source signals, in the widespread use of distorted image context of detection.Displacement aliased image is the image of a kind of special single channel hybrid mode in tampered image, and certain part in image is by certain aliquot replacement in another piece image.Different from traditional superposition vision-mix, in this type of displacement aliased image, replaced region content of original image has loss, and the position of replacement areas, size, number are all unknown.

Existing displacement aliased image blind separation theory is the detection that feature analysis by having tampered image itself realizes replacement areas, realized the detection in image transform region by blind source separating, or realize the detection of replacement areas by the rarefaction representation of replacement image.And these theories also exist some deficiency following: (1) existing blind separation theory is the replacement image for single source.(2) existing blind separation theory is the replacement image for same mode process.(3) existing blind separation theory realizes being separated for replacement image specific features.The image sources of society is various, form is different, and existing blind separating method can not accurately be separated displacement aliased image.Shown by research, affecting the most important factor of aliased image blind separation effect is choosing property field.Therefore, how choosing the feature in aliased image accurately, is a problem demanding prompt solution.

Summary of the invention

The object of the invention is to overcome the deficiencies in the prior art, propose a kind of method of replacing aliased image blind separation, solve and not easily choose due to aliased image property field and affect the problem of blind separation accuracy.

The present invention is achieved by following scheme:

Replace a blind separating method for aliased image, step is as follows:

Step 1), build displacement aliased image I, the replacement areas of this displacement aliased image I contains noise;

Step 2), displacement aliased image I is divided into m equal-sized sub-block, wherein, m>=1; Set up an array, the pixel value of every behavior sub-block of array, this array is designated as image sequence X; Image sequence X is normalized and obtains new image sequence X ₁;

Step 3), utilize noise reduction autocoder to adopt forward direction algorithm to image sequence X ₁coding, namely extracts image sequence X ₁feature, recycling backpropagation and Gradient Descent to extract image sequence X ₁feature be reconstructed, obtain reconstruct image sequence Y; The image sequence Y of reconstruct and image sequence X ₁do business, obtain quotient images L;

Step 4), according to the mode dividing aliased image I, quotient images L is also divided into m equal-sized sub-block, wherein, m >=1; The threshold value of the pixel value and setting of choosing each sub-block compares, and realizes the replacement areas in displacement aliased image I to be separated with replaced area image according to the fiducial value obtained.

Further, step 2) described in image sequence X is normalized, method for normalizing expression formula is as follows:

X ₁＝double(X)/255

Further, step 3) described in noise reduction autocoder be three-layer neural network structure: input layer, hidden layer and output layer, adopt 0 Matrix cover input data of certain probability distribution, the input value expression formula of a hidden layer jth node is as follows:

$z_j=\underset{i=1}{\overset{m}{\Σ}}{w}_{ji}^{\left(1\right)}{x}_i+b_j^{\left(1\right)}$

Wherein, represent the weight between input layer i-th node and a hidden layer jth node; the bias of hidden layer jth node; x _iit is the input layer value of i-th node;

The expression way of output layer value y is as follows:

$y=\underset{i-1}{\overset{m}{\Σ}}{w}_{ij}^{\left(2\right)}{a}_j+b_i^{\left(2\right)}$

Wherein, a _jfor the output valve of each hidden node, by the non-linear expression of sigmoid activation function f (z), i.e. f (z)=1/ (1+e ^-z), a _j=f (z _j); for the weight between hidden layer i-th node and an output layer jth node; for the bias of output layer jth node.

Further, step 3) described in the expression formula of back-propagation method be:

$J(w,b,x,y)=\frac{1}{2}\left|\right|y_i-x_i|{|}^2$

$J(w,b)\[1/m\underset{i=1}{\overset{m}{\Σ}}J(w,b,x^{\left(i\right)},y^{\left(i\right)})\]+\mathrm{\λ}/2\underset{l=1}{\overset{nl-1}{\Σ}}\underset{i=1}{\overset{sl}{\Σ}}\underset{j=1}{\overset{sl+1}{\Σ}}{\left(w_{ji}^{\left(l\right)}\right)}^2$

Wherein, w is weighted value; B is biased; J (w, b, x, y) is the error of single sample; y _iit is the value of i-th node output layer; x _iit is the value of the input layer of i-th node; it is the square error of m node; it is the weight attenuation term of whole cost function;

Utilize gradient descent method to finely tune parameter w and b, expression formula is as follows:

$w_{ij}^{\left(l\right)}=w_{ij}^{\left(l\right)}-\mathrm{\α}\frac{\∂}{\∂w_{ij}^{\left(l\right)}}J(w,b)$

$b_i^{\left(l\right)}=b_i^{\left(l\right)}-\mathrm{\α}\frac{\∂}{\∂b_i^{\left(l\right)}}J(w,b)$

Wherein, α is learning rate; for the partial derivative of weight; for biased partial derivative

Further, step 3) described in the expression formula of quotient images L as follows:

L＝X ₁/Y

Wherein, X ₁for to the image sequence X after image sequence X normalization place ₁; Y is the image sequence to reconstruct.

Further, described step 4) in, the threshold value of the minimum pixel value and setting of choosing each sub-block of quotient images L compares, according to comparative result, binary conversion treatment is carried out to the pixel value of each sub-block, by the binary image that obtains and aliased image I by the mode of dot product, i.e. separable replacement areas and replaced area image.

The present invention's beneficial effect is compared to the prior art:

In the blind separating method of displacement aliased image, particularly important to choosing of its feature.The present invention proposes a kind of blind separating method of replacing aliased image.Noise reduction autocoder is used for the displacement aliased image of training Noise, extracts the feature of aliased image, namely displacement aliased image is encoded, the feature reconstruction displacement aliased image that recycling is extracted, be decode procedure.According to the difference that decoded image and original image exist, after doing business, adopt thresholding to operate, isolate replacement areas image.Autocoder has without supervision characteristic in degree of depth study, in the tasks such as Images Classification, have good effect.The present invention realizes choosing automatically aliasing graphic feature, not only save the time of Feature Selection, decrease the time complexity of retrieval, also overcome the shortcoming that property field in prior art is not easily chosen, avoid the phenomenon causing flase drop, improve the accuracy of separating resulting.

Adopt aliased image blind separating method of the present invention, be not subject to the restriction of image sources and processing mode, the displacement aliased image different for the replacement areas position of Noise, size, number and noise variance all can effectively isolate replacement areas image, has very strong adaptability.

And, in the process of training, adopt 0 Matrix cover original input data of certain probability distribution, Partial Feature is lost.The weight Noise trained due to the data after 0 Matrix cover is less, can reduce the noise of the aliased image after training, thus makes the feature of extraction have more robustness.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the displacement aliased image blind separating method that the present invention is based on noise reduction autocoder;

Fig. 2 to be replacement image Noise variance of the present invention be 0.2 displacement aliased image;

Fig. 3 image that to be replacement image Noise variance of the present invention be after the coding of the displacement aliased image of 0.2;

Fig. 4 to be replacement image Noise variance of the present invention be 0.2 the decoded image of displacement aliased image;

Fig. 5 to be replacement image Noise variance of the present invention be 0.2 displacement aliased image make the result after business;

Fig. 6 to be replacement image Noise variance of the present invention be 0.2 displacement aliased image blind separation design sketch;

Fig. 7 is the displacement aliased image that replacement image Noise type of the present invention is white Gaussian noise, variance is 0.1;

Fig. 8 is the displacement aliased image that replacement image Noise type of the present invention is salt-pepper noise, variance is 0.2;

Fig. 9 is the displacement aliased image that replacement image Noise type of the present invention is white Gaussian noise, variance is 0.3;

Figure 10 is the displacement aliased image that replacement image Noise type of the present invention is salt-pepper noise, variance is 0.4;

Figure 11 is the displacement aliased image that replacement image Noise type of the present invention is white Gaussian noise, variance is 0.5;

Figure 12-16 is the blind separation result figure of Fig. 7-11 successively;

Figure 17-18 is the displacement aliased image containing the different replacement image of size, position and noise variance;

Figure 19-20 is the blind separation result figure of Figure 17-18 successively;

Figure 21 is the structural drawing of DAE neural network algorithm of the present invention.

Embodiment

Below in conjunction with drawings and Examples, the present invention will be further described in detail.

Replace a blind separating method for aliased image, step is as follows:

Step (one), structure size are the displacement aliased image I of M × M, and the replacement areas of this image contains noise, then the mathematical model expression formula of replacing aliased image I is as follows:

I＝A⊙P+A _T1⊙P _T1+...+A _Tm⊙P _Tm+...+A _Tn⊙P _Tn(1)

Wherein, ⊙ represents that Hadamard amasss; P is the source images of replacement image; A is all 1's matrix; A _ti(i=1,2 ..., n) be permutation matrix; P _ti(i=1,2 ..., n) be the source images of Noise.

In formula (1), A _t1, A _t2... A _tnconcrete representation as follows:

$A_{T1}=\left\{\begin{array}{cc}1& (i,j)\∈U_{T1}\\ & .\\ & .\\ & .\\ 0& (i,j)\∈U_{Ti}\\ & .\\ & .\\ & .\\ 0& (i,j)\∈U_{Tn}\end{array}\right.,\mathrm{...},A_{Ti}=\left\{\begin{array}{cc}0& (i,j)\∈U_{T1}\\ & .\\ & .\\ & .\\ 1& (i,j)\∈U_{Ti}\\ & .\\ & .\\ & .\\ 0& (i,j)\∈U_{Tn}\end{array}\right.,\mathrm{...},A_{Tn}=\left\{\begin{array}{cc}0& (i,j)\∈U_{T1}\\ & .\\ & .\\ & .\\ 0& (i,j)\∈U_{Ti}\\ & .\\ & .\\ & .\\ 1& (i,j)\∈U_{Tn}\end{array}\right.---\left(2\right)$

Wherein, U _t1, U _t2... U _tnfor activation interval,

Step (two), aliased image I will be replaced according to 8 × 8 points of block sizes, piecemeal will be carried out to it, obtain m sub-block, wherein, m >=1, m sub-block is preserved into an array, and every a line of array represents the pixel value of a wherein fritter, and this array is designated as image sequence X.Using image sequence X as input data set, because the grey scale pixel value in image sequence X differs greatly, if directly utilize its training of noise reduction automatic coding can produce comparatively big error, therefore need to be normalized image sequence X, thus obtain new image sequence X ₁.Then normalization expression formula is as follows:

X ₁＝double(X)/255(3)

Step (three), structure noise reduction autocoder (DEA) neural network, utilize the image sequence X after normalization ₁it is trained, utilizes propagated forward algorithm first to image sequence X ₁carry out own coding operation, namely to image sequence, it carries out feature extraction.The concrete mode building DAE neural network is as follows:

(1) set the initial parameter value of DAE neural network, build DAE neural network structure, initializes weights w and biased b.Wherein, InputZeroMaskedFraction in DEA neural network produces 0 matrix of certain probability distribution, high dimensional feature (i.e. noise characteristic) due to 0 Matrix cover image reaches the effect of Data Dimensionality Reduction, by the view data of the data after the Data Dimensionality Reduction of 0 Matrix cover closer to not Noise, so the weight Noise that the data after 0 Matrix cover train is less, thus reach Noise Reduction.Its value is set to 0.5, adopts default value for other parameter values.

(2) DEA model as shown in figure 21, and DEA neural network is divided into three-layer network to comprise, input layer, hidden layer and output layer, the weight matrix on limit between adjacent two layers node with $w^{\left(2\right)}=\{w_{11}^{\left(2\right)},w_{12}^{\left(2\right)},\mathrm{...},w_{nm}^{\left(2\right)}\}$ Represent, the bias term of hidden layer and output layer is used $b^{\left(1\right)}=\{b_1^{\left(1\right)},b_2^{\left(1\right)},\mathrm{...},b_n^{\left(1\right)}\}$ With therefore, the input value expression formula of a hidden layer jth node is as follows:

$z_j=\underset{i=1}{\overset{m}{\Σ}}{w}_{ji}^{\left(1\right)}{x}_i+b_j^{\left(1\right)}---\left(4\right)$

Wherein, represent the weight between input layer i-th node and a hidden layer jth node; the bias of hidden layer jth node; x _iit is the input layer value of i-th node.

The expression way of output layer value y is as follows:

$y=\underset{i-1}{\overset{m}{\Σ}}{w}_{ij}^{\left(2\right)}{a}_j+b_i^{\left(2\right)}---\left(5\right)$

Wherein, a _jfor the output valve of each hidden node, by the non-linear expression of sigmoid activation function f (z), i.e. f (z)=1/ (1+e ^-z), a _j=f (z _j); for the weight between hidden layer i-th node and an output layer jth node; for the bias of output layer jth node.

The feature reconstruction displacement aliased image that step (four), utilization are extracted.By backpropagation and Gradient Descent adjustment weight w and biased b, then make reconstructed error minimum according to the feature reconstruction displacement aliased image extracted.Utilize the network trained, obtain the image sequence Y of reconstruct, adopt back-propagation algorithm to adjust it.Concrete mode is as follows:

(1) the cost function expression formula for single sample is as follows:

$J(w,b,x,y)=\frac{1}{2}\left|\right|y_i-x_i|{|}^2---\left(6\right)$

Wherein, w is weighted value; B is biased; J (w, b, x, y) represents the error of single sample; y _iit is the value of i-th node output layer; x _iit is the value of the input layer of i-th node.

Overall cost function expression formula for m sample set is as follows:

$\begin{array}{c}J(w,b)=\[1/m\underset{i=1}{\overset{m}{\Σ}}J(w,b;x^{\left(i\right)},y^{\left(i\right)})\]+\mathrm{\λ}/2\underset{l=1}{\overset{nl-1}{\Σ}}\underset{i=1}{\overset{sl}{\Σ}}\underset{j=1}{\overset{sl+1}{\Σ}}{\left(w_{ji}^{\left(l\right)}\right)}^2\\ =\[1/m\underset{i=1}{\overset{m}{\Σ}}1/2\left|\right|y_i-x_i|{|}^2\]+\mathrm{\λ}/2\underset{l=1}{\overset{nl-1}{\Σ}}\underset{i-1}{\overset{sl}{\Σ}}\underset{j=1}{\overset{sl+1}{\Σ}}{\left(w_{ji}^{\left(l\right)}\right)}^2\end{array}---\left(7\right)$

Wherein, $1/m\underset{i=1}{\overset{m}{\Σ}}J(w,b,x^{\left(i\right)},y^{\left(i\right)})$ It is the square error of m node; $\mathrm{\λ}/2\underset{l=1}{\overset{nl-1}{\Σ}}\underset{i=1}{\overset{sl}{\Σ}}\underset{j=1}{\overset{sl+1}{\Σ}}{\left(w_{ji}^{\left(l\right)}\right)}^2$ It is the weight attenuation term of whole cost function.

(2) according to the error that backpropagation feeds back, utilize gradient descent method to finely tune parameter w and b, expression formula is as follows:

$w_{ij}^{\left(l\right)}=w_{ij}^{\left(l\right)}-\mathrm{\α}\frac{\∂}{\∂w_{ij}^{\left(l\right)}}J(w,b)---\left(8\right)$

$b_i^{\left(l\right)}=b_i^{\left(l\right)}-\mathrm{\α}\frac{\∂}{\∂b_i^{\left(l\right)}}J(w,b)---\left(9\right)$

Wherein, α is learning rate; for the partial derivative of weight; for biased partial derivative.

(3) the solving of partial derivative:

Carry out propagated forward calculating according to initialized weight matrix with biased, and then obtain l ₂, l ₃..., l _nlthe activation value of layer.

Calculate the residual error of each node of each layer; Wherein, for l _nlthe computing formula of the residual error of each cell node i of layer is as follows:

$\begin{array}{c}{\mathrm{\δ}}_i^n=\frac{\∂}{\∂z_i^{nl}}J(\mathrm{\ω},b;x,y)=\frac{\∂}{\∂z_i^{nl}}\frac{1}{2}\left|\right|y-h_{\mathrm{\ω},b}\left(x\right)|{|}^2\\ =-(y_i-a_i^{\left(nl\right)})\·f^{\′}\left(z_i^{\left(nl\right)}\right)\end{array}---\left(10\right)$

For l=nl-1, nl-2 ... the residual computations formula of i-th node of each layer of 2 is as follows:

$\begin{array}{c}{\mathrm{\δ}}_i^{\left(l\right)}=\left(\underset{j=1}{\overset{s_{l+1}}{\Σ}}{\mathrm{\ω}}_{ji}^{\left(l\right)}{\mathrm{\δ}}_j^{(l+1)}\right)f^{\′}\left(z_i^{\left(l\right)}\right)\\ {\mathrm{\δ}}_i^{nl-1}=\frac{\∂}{\∂z_i^{nl-1}}J(\mathrm{\ω},b;x,y)=\frac{\∂}{\∂z_i^{nl-1}}\frac{1}{2}\left|\right|y-h_{\mathrm{\ω},b}\left(x\right)|{|}^2\\ =\left(\underset{j=1}{\overset{S_i-1}{\Σ}}{\mathrm{\ω}}_{ji}^{nl-1}{\mathrm{\δ}}_i^{n1}\right)f^{\′}\left(z_i^{nl-1}\right)\end{array}---\left(11\right)$

According to the partial derivative of required residual computations DAE neural network, expression formula is as follows:

$\begin{array}{c}\frac{\∂}{\∂{\mathrm{\ω}}_{ij}^{\left(l\right)}}J(\mathrm{\ω},b;x,y)=a_j^l{\mathrm{\δ}}_i^{l+1}\\ \frac{\∂}{\∂b_i^{\left(l\right)}}J(\mathrm{\ω},b;x,y)={\mathrm{\δ}}_i^{l+1}\end{array}---\left(12\right)$

Step (five), by reconstruct figure sequence Y and original image normalization after image sequence X ₁do business can obtain:

L＝X ₁/Y(13)

M equal-sized sub-block is divided into by making the image L after business, wherein, m >=1.Carry out analysis discovery by making the image L after business, the image factor sequence value of Noise is very little, and the image factor sequence value of Noise is obviously very not large.Therefore, extract the minimum pixel value of each sub-block, and by it stored in array N.

Step (six), to array N set threshold value, if the numerical value in array N is less than threshold value, be namely considered to the region of Noise, then the gray-scale value of the pixel of its corresponding blocks be set to 1, otherwise be set to 0.Thus the binaryzation realized image.Then, the image of original image and binaryzation is carried out dot product, replacement areas image can be separated with replaced area image.

MATLAB emulation experiment:

The CPU of to be dominant frequency the be 3.20GHz that experiment porch adopts, in save as 4GB, the PC of 64 Win7 operating systems, programmed by MATLABR2012b software.For guaranteeing accuracy of the present invention, the image of employing is all at the conventional standard picture of image processing field experiment.

The noise type that this experiment adopted is white Gaussian noise and salt-pepper noise, in order to verify the accuracy that this experiment is separated, choose 5 width experiment picture boat (as shown in Fig. 7,8,9,10,11) and carry out blind separation experiment, white Gaussian noise, salt-pepper noise, white Gaussian noise, salt-pepper noise, white Gaussian noise is divided into containing noise type in 5 width figure, noise variance is respectively 0.1,0.2,0.3,0.4,0.5, and separating resulting is as shown in Figure 12,13,14,15,16.

By drawing separating resulting analysis, algorithm of the present invention is to the image containing different noise variance and effectively can isolate replacement areas containing the image of white Gaussian noise or salt-pepper noise, demonstrates the separation accuracy of we.

Because the replacement areas position of image, size, number and noise variance all may be different, these factors all may affect image analogy accuracy.So this experiment also carries out blind separation for the image (as shown in Figure 17,18) containing these factors, experimental result is as shown in Figure 19,20.Can draw to the result of image analogy, the inventive method, to displacement regional location, size, image that number is different with noise variance, effectively can isolate replacement areas.Also demonstrate the present invention and there is well separation accuracy and robustness.

Adopt linear mode to be normalized image in above-described embodiment, also can choose other image normalization methods as other embodiments, as z-score method for normalizing, intermediate value method for normalizing etc.

Above-described embodiment when carrying out blind separation to aliased image, the blind separation that to be binarization method realize aliased image in conjunction with the mode of dot product of utilization.The method be in fact separated is not limited thereto, and also can adopt other image procossing mode step (six).

Under the thinking that the present invention provides; the mode easily expected to those skilled in the art is adopted to convert the technological means in above-described embodiment, replace, revise; and the effect played goal of the invention that is substantially identical with the relevant art means in the present invention, that realize is also substantially identical; the technical scheme of such formation is carried out fine setting to above-described embodiment and is formed, and this technical scheme still falls within the scope of protection of the present invention.

Claims (6)

1. replace a blind separating method for aliased image, it is characterized in that, step is as follows:

Step 1), build displacement aliased image I, the replacement areas of this displacement aliased image I contains noise;

Step 2), displacement aliased image I is divided into m equal-sized sub-block, wherein, m>=1; Set up an array, the pixel value of every behavior sub-block of array, this array is designated as image sequence X; Image sequence X is normalized and obtains new image sequence X ₁;

Step 3), utilize noise reduction autocoder to adopt forward direction algorithm to image sequence X ₁coding, namely extracts image sequence X ₁feature, recycling backpropagation and Gradient Descent to extract image sequence X ₁feature be reconstructed, obtain reconstruct image sequence Y; The image sequence Y of reconstruct and image sequence X ₁do business, obtain quotient images L;

Step 4), according to the mode dividing aliased image I, quotient images L is also divided into m equal-sized sub-block, wherein, m >=1; The threshold value of the pixel value and setting of choosing each sub-block compares, and realizes the replacement areas in displacement aliased image I to be separated with replaced area image according to the fiducial value obtained.

2. a kind of blind separating method of replacing aliased image according to claim 1, is characterized in that, step 2) described in image sequence X is normalized, method for normalizing expression formula is as follows:

X ₁＝double(X)/255。

3. a kind of blind separating method of replacing aliased image according to claim 1, it is characterized in that, step 3) described in noise reduction autocoder be three-layer neural network structure: input layer, hidden layer and output layer, adopt 0 Matrix cover input data of certain probability distribution, the input value expression formula of a hidden layer jth node is as follows:

$z_j=\underset{i=1}{\overset{m}{\Σ}}{w}_{ji}^{\left(1\right)}{x}_i+b_j^{\left(1\right)}$

Wherein, represent the weight between input layer i-th node and a hidden layer jth node; the bias of hidden layer jth node; x _iit is the input layer value of i-th node;

The expression way of output layer value y is as follows:

$y=\underset{i-1}{\overset{m}{\Σ}}{w}_{ij}^{\left(2\right)}{a}_j+b_i^{\left(2\right)}$

Wherein, a _jfor the output valve of each hidden node, by the non-linear expression of sigmoid activation function f (z), i.e. f (z)=1/ (1+e ^-z), a _j=f (z _j); for the weight between hidden layer i-th node and an output layer jth node; for the bias of output layer jth node.

4. a kind of blind separating method of replacing aliased image according to claim 3, is characterized in that, step 3) described in the expression formula of back-propagation method be:

$J(w,b,x,y)=\frac{1}{2}\left|\right|y_i-x_i|{|}^2$

$J(w,b)=\[1/m\underset{i=1}{\overset{m}{\Σ}}J(w,b,x^{\left(i\right)},y^{\left(i\right)})\]+\mathrm{\λ}/2\underset{l=1}{\overset{nl-1}{\Σ}}\underset{i=1}{\overset{sl}{\Σ}}\underset{j=1}{\overset{sl+1}{\Σ}}{\left(w_{ji}^{\left(l\right)}\right)}^2$

Wherein, w is weighted value; B is biased; J (w, b, x, y) is the error of single sample; y _iit is the value of i-th node output layer; x _iit is the value of the input layer of i-th node; it is the square error of m node; it is the weight attenuation term of whole cost function;

Utilize gradient descent method to finely tune parameter w and b, expression formula is as follows:

$w_{ij}^{\left(l\right)}=w_{ij}^{\left(l\right)}-\mathrm{\α}\frac{\∂}{\∂w_{ij}^{\left(l\right)}}J(w,b)$

$b_i^{\left(l\right)}=b_i^{\left(l\right)}-\mathrm{\α}\frac{\∂}{\∂b_i^{\left(l\right)}}J(w,b)$

Wherein, α is learning rate; for the partial derivative of weight; for biased partial derivative.

5. a kind of blind separating method of replacing aliased image according to claim 2, is characterized in that, step 3) described in the expression formula of quotient images L as follows:

L＝X ¹/Y

Wherein, X ₁for to the image sequence X after image sequence X normalization place ₁; Y is the image sequence to reconstruct.

6. a kind of blind separating method of replacing aliased image according to claim 1, it is characterized in that, described step 4) in, the threshold value of the minimum pixel value and setting of choosing each sub-block of quotient images L compares, according to comparative result, binary conversion treatment is carried out to the pixel value of each sub-block, by the binary image that obtains and aliased image I by the mode of dot product, i.e. separable replacement areas and replaced area image.