CN108364258A - A kind of method and system improving image resolution ratio - Google Patents

Abstract

The invention discloses a kind of method and system improving image resolution ratio.Method includes：Original image is decomposed using opposite total variance method, obtains cartoon subgraph and texture subgraph；The cartoon subgraph is amplified using interpolation method, obtains amplified cartoon subgraph；The texture subgraph is amplified using homotopy Method, obtains amplified texture subgraph；The amplified cartoon subgraph and amplified texture subgraph are synthesized, high-definition picture is obtained.The present invention is amplified, then synthesized using different methods by the cartoon subgraph and texture subgraph different to amplification demand, obtains clearer high-definition picture, while improving high-definition picture quality, reduce processing time respectively.

Description

A method and system for improving image resolution

technical field

The invention relates to the field of image enlargement, in particular to a method and system for improving image resolution.

Background technique

The super-resolution reconstruction technology of a single image refers to the technology of using a low-resolution image to enlarge the image to obtain a high-resolution image, so as to improve the resolution and clarity of the image. At present, this technology has been widely used in various instruments for image and video detection and display, such as medical imaging, video surveillance, remote sensing images, high-definition television and other fields.

For this reason, experts and scholars have proposed many single image super-resolution methods. These single-image super-resolution reconstruction methods are mainly divided into pixel-point interpolation-based, image-projection-based, and learning-based methods. Among them, interpolation-based methods mainly include bilinear interpolation and bicubic interpolation. This kind of method has a small amount of calculation and can meet the requirements of real-time image enlargement, but the quality of the high-resolution image reconstructed by this kind of method is low, the details are not rich, and the image is relatively blurred. Image projection-based methods include iterative back-projection method and convex set projection method POCS (projection onconvex set), etc. This type of method generally requires multiple iterative projections of the image to obtain a better high-resolution image. The image super-resolution method is easily disturbed by image noise, the obtained image is unstable, and there is also the problem of partial loss of image details. The super-resolution method based on learning is mainly to train a large number of high- and low-resolution image samples to finally obtain the prior knowledge of the image, and then perform super-resolution reconstruction on the image. The classic method has the maximum a posteriori probability estimation method MAP (maximum a posterior), domain embedding method NE (NeighborEmbedding), a method based on example learning; recently, Yang et al. used linear programming to solve the sparse representation of low-resolution image blocks, and linearly combined the representation coefficients with high-resolution dictionaries to obtain high-resolution Resolution image blocks; the disadvantage of this type of method is that training takes a long time and process, and the storage space required for training is large, and its advantage is that the quality of the obtained high-resolution image is generally better than the above two types of methods . Therefore, how to reduce the processing time while improving the quality of high-resolution images has become an urgent technical problem to be solved.

Contents of the invention

The object of the present invention is to provide a method and system for increasing image resolution, so as to reduce processing time while improving the quality of high-resolution images.

To achieve the above object, the present invention provides the following scheme:

A method for improving image resolution, said method comprising the steps of:

The original image is decomposed by relative total variation method to obtain cartoon sub-image and texture sub-image;

Using an interpolation method to enlarge the cartoon sub-image to obtain the enlarged cartoon sub-image;

Using a homotopy method to amplify the texture sub-image to obtain the amplified texture sub-image;

Combining the enlarged cartoon sub-image and the enlarged texture sub-image to obtain a high-resolution image.

Optionally, the relative total variation method is used to decompose the image to obtain cartoon sub-images and texture sub-images; specifically including:

Segment the original image to obtain the segmented image;

The segmented image is decomposed to obtain cartoon sub-image and texture sub-image by using the method of image total variation with minimum window function.

Optionally, the segmenting the original image to obtain the segmented image specifically includes:

template and Carry out convolution with original image, obtain the gradient image I _x of x direction and the gradient image I _y of y direction;

Calculating the gradient image modulus I _g of the gradient image in the x direction and the gradient image in the y direction;

Among them, (i, j) is the pixel coordinates;

Taking the pixels whose gradient image modulus is greater than the first threshold as edge pixels, and the pixels whose gradient image modulus is greater than the second threshold as candidate edge pixels;

Select a plurality of edge pixels, take each edge pixel as a starting point, connect the edge pixels or candidate edge pixels directly adjacent to the edge pixel in eight directions, obtain several closed areas, and The small closed area located in the large closed area is deleted, and several closed areas without inclusion relationship are obtained;

Make a grayscale histogram for each closed area that has no inclusion relationship, and use the K-means method to segment the grayscale histogram to obtain a segmented image.

Optionally, the window function is:

Point (x _p , y _p ) is the current center pixel point, (x _q , y _q ) is the total variation pixel point of (x _p , y _p ), σ ² _p,q is the variance of the pixel values in the object where (x _q ,y _q ) and (x _p ,y _p ) are located, c _p,q is the multiplication factor, and T _g is the third threshold.

Optionally, the interpolation method is used to enlarge the cartoon sub-image to obtain the enlarged cartoon sub-image, which specifically includes:

The pixels of the cartoon sub-image are interpolated using an interpolation template function to obtain the enlarged cartoon sub-image; the interpolation template function is:

L(x,y)＝N ₁ *o(x,y)*d(x,y)*sinc(α*x/σ ² )

*sinc(α*y/σ ² ) ^* exp(β ^* (x ² +y ² )/2σ ² ) (3)

Among them, o(x, y) indicates whether to use the value of the current pixel to estimate the value of the unknown pixel. If the current pixel and the unknown pixel are in the same object, then o(x, y)=1, that is, use The current pixel; otherwise, o(x,y)=0, that is, the current pixel is not used, and d(x,y) represents the distance between the current pixel and the center point of the interpolation template; the function sinc(x)=(sin(x )/x); σ ² is the mean square error of all pixels involved in the interpolation of unknown pixels; α and β are constants.

Optionally, the amplifying the texture sub-image by using a homotopy method to obtain the amplified texture sub-image specifically includes:

Using the K-SVD dictionary learning method (the K-SVD dictionary learning method is a method combining the K-means classification method and the singular value decomposition svd method) to obtain the image block dictionary of the texture sub-image offline, the image block dictionary includes high-resolution High-resolution image block dictionary and low-resolution image block dictionary;

For the texture sub-image, use the image block dictionary and the orthogonal matching tracking method to perform online amplification to obtain an initial high-resolution image;

performing nearest-neighbor edge processing on the initial high-resolution image; obtaining a high-resolution image after edge processing;

performing the first homotopy processing on the high-resolution image after the edge processing to obtain the first high-resolution image;

The second homotopy processing is performed on the first high-resolution image to obtain an enlarged high-resolution texture sub-image.

Optionally, said synthesizing the enlarged cartoon sub-image and the enlarged texture sub-image to obtain a high-resolution image specifically includes:

Adopt formula (4) to synthesize the cartoon sub-image after the enlargement and the texture sub-image after the enlargement, obtain a high-resolution image;

f _H ＝f _c +f _t +λ ₁ *G(f _t ) (4)

Among them, f _H is a high-resolution image, f _t is an enlarged texture image, f _c is an enlarged cartoon image, G( _ft ) is the modulus value of Robert's gradient for image f _t , and λ ₁ is a constant.

A system for increasing image resolution, the system comprising:

The image decomposition module is used to decompose the original image using the relative total variation method to obtain cartoon sub-images and texture sub-images;

The cartoon sub-image enlargement module is used to amplify the cartoon sub-image by interpolation to obtain the enlarged cartoon sub-image;

A texture sub-image enlarging module, configured to amplify the texture sub-image using a homotopy method to obtain an enlarged texture sub-image;

The image synthesis module is used to synthesize the enlarged cartoon sub-image and the enlarged texture sub-image to obtain a high-resolution image.

Optionally, the image decomposition module specifically includes:

The image segmentation sub-module is used to segment the original image to obtain the segmented image;

The image decomposition sub-module is used to decompose the segmented image to obtain cartoon sub-images and texture sub-images by using the image total variation method of minimizing the window function.

Optionally, the image segmentation submodule specifically includes:

Convolutional units for exploiting templates separately and Carry out convolution to original image, obtain the gradient image I _x of x direction and the gradient image I _y of y direction;

A gradient image modulus calculation unit, configured to calculate the gradient image modulus I _g of the gradient image in the x direction and the gradient image in the y direction;

Among them, (i, j) is the pixel coordinates;

A pixel point selection unit, configured to use pixels whose gradient image modulus is greater than the first threshold as edge pixels, and use pixels whose gradient image modulus is greater than the second threshold as candidate edge pixels;

The closed area acquisition unit is used to select a plurality of edge pixel points, and respectively use each edge pixel point as a starting point to connect edge pixel points or candidate edge pixel points directly adjacent to the edge pixel point in eight directions, Obtain several closed areas, delete the small closed areas located in the large closed area, and obtain several closed areas without inclusion relationship;

The image segmentation unit is configured to make a grayscale histogram for each closed area that has no inclusion relationship, and use the K-means method to segment the grayscale histogram to obtain a segmented image.

According to the specific embodiments provided by the invention, the invention discloses the following technical effects:

The invention discloses a method and system for improving image resolution. The method includes: using the relative total variation method to decompose the original image to obtain a cartoon sub-image and a texture sub-image; using an interpolation method to amplify the cartoon sub-image to obtain an enlarged cartoon sub-image; using a homotopy method to obtain a cartoon sub-image The texture sub-image is enlarged to obtain an enlarged texture sub-image; the enlarged cartoon sub-image and the enlarged texture sub-image are synthesized to obtain a high-resolution image. The present invention uses different methods to amplify cartoon sub-images and texture sub-images with different amplifying requirements, and then synthesizes them to obtain clearer high-resolution images, which reduces the processing time while improving the quality of high-resolution images .

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required in the embodiments. Obviously, the accompanying drawings in the following description are only some of the present invention. Embodiments, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without paying creative labor.

Fig. 1 is a flow chart of a method for improving image resolution provided by the present invention;

Fig. 2 is that a kind of method for improving image resolution provided by the present invention is carrying out the initial image block of homotopy processing for the first time;

Fig. 3 is that a kind of method for improving image resolution provided by the present invention is carrying out the initial image block of the second homotopy processing;

Fig. 4 is a structural block diagram of a system for improving image resolution provided by the present invention;

Fig. 5 is a comparison diagram of the effects of various different processing methods for character images provided by the present invention;

Fig. 6 is a comparison diagram of effects of various different processing methods of butterfly images provided by the present invention;

Wherein, (a) in Fig. 5 is the original image; (b) is the bicubic interpolation image; (c) is the image of the method proposed by Yang; (d) is the image of the method proposed by Zedye; (e) is the present invention The image of the method; (a) in Figure 6 is the original image; (b) is the bicubic interpolation image; (c) is the image of the method proposed by Yang; (d) is the image of the method proposed by Zedye; (e) is an image of the method of the present invention.

Detailed ways

The object of the present invention is to provide a method and system for increasing image resolution, so as to reduce processing time while improving the quality of high-resolution images.

In order to make the above objects, features and advantages of the present invention more comprehensible, the invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

The method based on interpolation has a better effect on the enlargement of the smooth area, and the method based on learning has a better effect on the enlargement of the detailed part of the image. In the method proposed by the present invention, the two are organically combined to obtain better performance. First, an improved image decomposition method is used to decompose the image into cartoon and texture parts. Among them, the cartoon part extracts the structural information of the image, and the pixel value only has a large change at the junction of the object, and the change of the pixel value inside the object is small, and the image is smooth. The texture part extracts the detail part of the image, where the pixel value changes greatly. Then, for the cartoon part, use the method of interpolation to enlarge; for the texture part, use the method of homotopy to enlarge. Finally, the processing results of the above two parts are better synthesized to obtain a clearer high-resolution image. In the following, each detail of the proposed method will be explained in more detail.

As shown in Figure 1, the present invention provides a kind of method that improves image resolution, and described method comprises the following steps:

Step 101, using relative total variation method to decompose the original image to obtain cartoon sub-image and texture sub-image. In the original cartoon and texture decomposition method RTV, the cartoon part is obtained by minimizing the total variation of the image with the window function. This method has the following disadvantages: 1. The total variation at the edge of the object of the image should not be counted into the objective function that needs to be minimized. The change of this pixel value is a normal change that different objects generally have different pixel values; 2. The window The size and range of the function value should change adaptively with the pixel value of the object, so as to improve the effect of cartoon part extraction. For example, when the pixel value changes quickly in the object, the window function should take a larger range and a larger value. For this reason, the relative total variation method proposed by the present invention decomposes the original image to obtain cartoon sub-images and texture sub-images, specifically including:

First, segment the original image to obtain the segmented image, including:

template and Carry out convolution with original image, obtain the gradient image I _x of x direction and the gradient image I _y of y direction;

Calculating the gradient image modulus I _g of the gradient image in the x direction and the gradient image in the y direction;

Among them, (i, j) is the pixel coordinates;

Taking the pixels whose gradient image modulus is greater than the first threshold as edge pixels, and the pixels whose gradient image modulus is greater than the second threshold as candidate edge pixels;

Select a plurality of edge pixels, take each edge pixel as a starting point, connect the edge pixels or candidate edge pixels directly adjacent to the edge pixel in eight directions, obtain several closed areas, and The small closed area located in the large closed area is deleted, and several closed areas without inclusion relationship are obtained;

Make a grayscale histogram for each closed area that has no inclusion relationship, and use the K-means method to segment the grayscale histogram to obtain a segmented image.

Second, the segmented image is decomposed to obtain the cartoon sub-image and the texture sub-image by using the method of image total variation that minimizes the window function. The window function is:

Point (x _p , y _p ) is the current center pixel point, (x _q , y _q ) is the total variation pixel point of (x _p , y _p ), σ ² _p,q is the variance of the pixel values in the object where (x _q ,y _q ) and (x _p ,y _p ) are located, c _p,q is the multiplication factor, which is equal to the detected edge pixels in the object Points are proportional, T _g is the third threshold.

Step 102, using an interpolation method to enlarge the cartoon sub-image to obtain an enlarged cartoon sub-image. The method of using interpolation proposed by the present invention to complete the estimation and calculation of unknown pixels in the high-resolution image of the cartoon sub-image is different from previous methods. The interpolation method of the present invention will follow the following two principles in the interpolation process: 1. Only use pixels in the same object as unknown pixels for interpolation to avoid blurring of pixels after interpolation. 2. Adaptively adjust the size of the interpolation template. For the smooth part of the image, a template with a relatively large range will be used to allow more pixels to participate in the interpolation. For the part with large changes in the image, a template with a smaller range will be used , and only use pixels with a large correlation with the current pixel for interpolation. In this way, the blurring of pixels after interpolation can also be avoided.

Specifically comprising, interpolating the pixels of the cartoon sub-image using an interpolation template function to obtain an enlarged cartoon sub-image; the interpolation template function is:

L(x,y)＝N ₁ *o(x,y)*d(x,y)*sinc(α*x/σ ² )

^* sinc(α ^* y/σ ² ) ^* exp(β ^* (x ² +y ² )/2σ ² ) (3)

Among them, o(x, y) indicates whether to use the value of the current pixel to estimate the value of the unknown pixel. If the current pixel and the unknown pixel are in the same object, then o(x, y)=1, that is, use The current pixel; otherwise, o(x,y)=0, that is, the current pixel is not used, and d(x,y) represents the distance between the current pixel and the center point of the interpolation template; the function sinc(x)=(sin(x )/x); σ ² is the mean square error of all pixels involved in the interpolation of unknown pixels; α and β are constants. Since the interpolation template function needs to be truncated for use in practical applications, only pixels with L(x,y)>T ₁ (T ₁ is a function that needs to be determined in experiments) are selected to participate in the interpolation. N ₁ is a normalization constant, so that the cumulative value of all the coefficients participating in the interpolation is 1.

Step 103, using a homotopy method to enlarge the texture sub-image to obtain an enlarged texture sub-image. In the method for texture sub-image processing proposed by the present invention, the learning process of off-line high-resolution dictionary and low-resolution dictionary adopts the K-SVD dictionary learning method to first obtain the high-resolution dictionary D _h and Low resolution dictionary D _l . For the online enlargement process of the texture sub-image, firstly, the obtained image block dictionary and the Orthogonal Matching Pursuit (OMP) method are used for enlargement to obtain an initial high-resolution image. Next, the present invention performs nearest neighbor edge processing on the initial high-resolution image, and then runs the homotopy method twice for online recovery. In the process of adding edges to the image, the unknown pixel value is determined by the value of the nearest pixel point, the height of the added edge is M ₁ , the width is N ₁ , and M ₁ ×N ₁ is for each high-resolution texture image block The size of the processed image blocks. In the first homotopy process, the initial high-resolution image block shown in Fig. 2 is used for updating. In the second homotopy processing, the initial high-resolution image block shown in Fig. 3 is used for updating. Wherein, FIG. 2 shows a schematic diagram of an image block to be enlarged in a high-resolution image, where the shaded part is the enlarged region, and the white part is the part where pixel value estimation is required. In Fig. 3, an image block in the second homotopy processing is shown, where the shaded part is considered to be a part with known pixel values, and the blank part is considered to be a part that needs to be updated again.

Specifically include: using the K-SVD dictionary learning method to obtain the image block dictionary of the texture sub-image offline, the image block dictionary includes a high-resolution image block dictionary and a low-resolution image block dictionary; for the texture sub-image, Using the image block dictionary and the orthogonal matching tracking method to perform online amplification to obtain an initial high-resolution image; perform nearest-neighbor edge processing on the initial high-resolution image; obtain a high-resolution image after edge processing; The high-resolution image after the edge processing is subjected to the first homotopy processing to obtain the first high-resolution image; the first high-resolution image is subjected to the second homotopy processing to obtain the enlarged high-resolution image Rate texture subimage.

Step 104, combining the enlarged cartoon sub-image and the enlarged texture sub-image to obtain a high-resolution image, specifically including: using formula (4) to synthesize the enlarged cartoon sub-image and the enlarged texture The sub-images are synthesized to obtain a high-resolution image;

f _H ＝f _c +f _t +λ ₁ *G(f _t )(4)

Among them, f _H is a high-resolution image, f _t is an enlarged texture image, f _c is an enlarged cartoon image, G( _ft ) is the modulus value of Robert's gradient for image f _t , and λ ₁ is a constant.

FIG. 4 is a structural block diagram of a system for improving image resolution provided by the present invention. Referring to Fig. 4, a system for improving image resolution, the system includes:

The image decomposition module 401 is used to decompose the original image by using the relative total variation method to obtain the cartoon sub-image and the texture sub-image; the image decomposition module specifically includes: an image segmentation sub-module, which is used to segment the original image to obtain The segmented image; the image decomposition sub-module, which is used to decompose the segmented image to obtain cartoon sub-images and texture sub-images by using the image total variation method of minimizing the window function.

The image segmentation sub-module specifically includes: a convolution unit for using templates respectively and The original image is convolved to obtain the gradient image I _x in the x direction and the gradient image I _y in the y direction; the gradient image modulus calculation unit is used to calculate the gradient image in the x direction and the gradient image in the y direction Gradient image modulus I _g ;

Wherein, (i, j) is pixel point coordinates; Pixel point selection unit is used for taking the pixel point whose gradient image modulus value is greater than the first threshold as an edge pixel point, and using the pixel point whose gradient image modulus value is greater than the second threshold value as a candidate Edge pixel points; a closed area acquisition unit, used to select a plurality of edge pixel points, and each edge pixel point is used as a starting point to connect edge pixel points or candidate points directly adjacent to the edge pixel point in eight directions Edge pixels, to obtain several closed areas, and delete the small closed areas located in the large closed area, to obtain several closed areas without containment relationship; the image segmentation unit is used to gray each closed area without containment relationship Intensity histogram, and the K-means method is used to segment the gray histogram to obtain a segmented image.

The cartoon sub-image enlargement module 402 is used to enlarge the cartoon sub-image by interpolation to obtain the enlarged cartoon sub-image;

A texture sub-image enlarging module 403, configured to amplify the texture sub-image using a homotopy method to obtain an enlarged texture sub-image;

The image synthesis module 404 is configured to synthesize the enlarged cartoon sub-image and the enlarged texture sub-image to obtain a high-resolution image.

The present invention also verifies the performance of the method and system proposed by the present invention through experiments, specifically including:

Windows 8 operating system, Matlab2013b software, 8GB memory, and 2.9GHz processor were selected as the experimental platform, and 91 images from the experimental sample library were selected as training samples. The average peak signal-to-noise ratio PSNR (peak signal to noise ratio) and RMSE (Root Mean Square Error). As shown in formulas (5) and (6):

where M×N is the size of the image, x(i,j) and are the image values at coordinates (i, j) of the original image and the reconstructed image, respectively.

In the experiment of the present invention, since the human eye is very sensitive to the brightness component of the image, the present invention first divides the color image into three channels of YCbCr, and then only takes the Y (brightness) channel component to carry out the following super-resolution processing. The low-resolution image is reconstructed and enlarged three times. For the texture part of the image, 50000 image blocks are selected for training, the number of iterations of K-SVD is set to 40, and the number of atoms in the dictionary is set to 1000. The overlap matrix size is 3x3 and the overlap pixel is 1. The experimental results are shown in Figure 5-6.

It can be seen from Figures 5 and 6 that the results processed by the bicubic interpolation method are too smooth, and at the same time, from the details of the characters and butterflies, there is a jagged effect. Yang's method has recovered better, but the freckles on the face have not been recovered, and many details have been lost. The method proposed by the present invention pays attention to the recovery of details, so it is superior in details.

The present invention uses an objective evaluation method to calculate PSNR and RMSE of three different reconstruction methods. Select the standard test image for comparison, as can be seen from Table 1 and Table 2, the method of the present invention has better performance compared with the bicubic interpolation method, the method of Yang, the Zeyde method, thus it can be seen that the method of the present invention A better reconstruction image quality can be obtained. The method used in the present invention has an improvement of about 0.2db over Zeyde's method.

Table 1 PSNR values of images processed by different methods

image bicubic interpolation Yang method Zeyde method The method of the invention zebra 26.6340 27.9531 28.4968 28.8147 ppt3 23.7066 24.9777 25.1789 25.5333 monarch 29.4259 30.7133 31.0921 31.3748 lenna 31.6778 32.6384 32.9951 33.2255 flowers 27.2301 28.2488 28.4236 28.6303 comic 23.1156 23.9023 23.9726 24.1433

Table 2 RMSE values of images processed by different methods

image bicubic interpolation Yang method Zeyde method The method of the invention zebra 11.8806 10.2067 9.5874 9.2428 ppt3 16.6422 14.3765 14.0473 13.4857 monarch 8.6148 7.4281 7.1111 6.8833 lenna 11.0926 9.8651 9.6685 9.4412 flowers 11.0926 9.8651 9.6685 9.4412 comic 17.8139 16.2714 16.1402 15.8261

According to the specific embodiments provided by the invention, the invention discloses the following technical effects:

1. Before the image is enlarged, the present invention first separates the image into texture and cartoon parts, wherein the texture part mainly includes the detail parts in the image, and the cartoon part includes the structure parts in the image. Therefore, in the process of image enlargement, finer enlargement of the texture part is required to overcome the problem of losing image details after image enlargement in traditional methods. At the same time, the present invention is different from the traditional TV (total variation) decomposition method, and has a better decomposition method than the TV method, and its effect has been proved in the experimental results, so the present invention first improves this method, and then The image is decomposed into an object-based structure part and a texture part.

2. For the cartoon part in the image, its high-frequency components are less, and using the traditional bi-cubic method to amplify this part can already achieve better performance and save the time required for image enlargement.

3. For the texture part in the image, it is necessary to maintain its details when zooming in. Therefore, effective learning-based methods are required. In the image training phase, in order to overcome the shortcomings of the Zeyde method that the high and low resolution dictionaries have the same sparse representation, the idea of joint training is added to ensure that the high and low image blocks have the same representation coefficients on the overcomplete dictionary. The traditional OMP (Orthogonal Matching Pursuit) method can be used for the online enlargement stage of the texture part, and its advantage is that the operation speed is fast and the enlargement effect is good. Moreover, the present invention uses the L1-homotopy (homotopy based on 1 norm) method, and compared with the OMP method, the reconstruction effect of this method is better.

Each embodiment in this specification is described in a progressive manner, each embodiment focuses on the difference from other embodiments, and the same and similar parts of each embodiment can be referred to each other. As for the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and for the related information, please refer to the description of the method part.

In this paper, specific examples are used to illustrate the principle and implementation of the invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. The described embodiments are only part of the embodiments of the present invention. , not all of the embodiments, based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work, all belong to the protection scope of the present invention.

Claims (10)

1. a kind of method improving image resolution ratio, which is characterized in that described method includes following steps：

Original image is decomposed using opposite total variance method, obtains cartoon subgraph and texture subgraph；

The cartoon subgraph is amplified using interpolation method, obtains amplified cartoon subgraph；

The texture subgraph is amplified using homotopy Method, obtains amplified texture subgraph；

The amplified cartoon subgraph and the amplified texture subgraph are synthesized, high resolution graphics is obtained Picture.

2. a kind of method improving image resolution ratio according to claim 1, which is characterized in that described to be become using relatively total Poor method decomposes image, obtains cartoon subgraph and texture subgraph；It specifically includes：

Original image is split, the image after being divided；

Using the full variational approach of image for minimizing window function, the image after segmentation is decomposed, obtains cartoon subgraph With texture subgraph.

3. a kind of method improving image resolution ratio according to claim 2, which is characterized in that described to be carried out to original image Segmentation, the image after being divided specifically include：

It is utilized respectively templateWithConvolution is carried out with original image, obtains the gradient map in the directions x As I_xWith the gradient image I in the directions y_y；

Calculate the gradient image modulus value I of the gradient image in the directions x and the gradient image in the directions y_g；

Wherein, (i, j) is pixel point coordinates；

Pixel using the gradient image modulus value more than first threshold is big by the gradient image modulus value as edge pixel point In second threshold pixel as candidate edge pixel；

Multiple edge pixel points are chosen, respectively using each edge pixel point as starting point, connection and the edge picture Vegetarian refreshments the edge pixel point of direct neighbor or candidate edge pixel on eight directions, obtain several enclosed areas Domain, and the small closed area in the big closed area is deleted, obtain the enclosed area of several not inclusion relations Domain；

Grey level histogram is done to the closed area of each not inclusion relation, and K- is used to the grey level histogram Means methods are split, image after being divided.

4. a kind of method improving image resolution ratio according to claim 2, which is characterized in that

The window function is：

Point (x_p,y_p) it is Current central pixel point, (x_q,y_q) it is (x_p,y_p) full variation pixel,σ² _p,qIt is (x_q,y_q) and (x_p,y_p) where pixel value in object variance, c_p,qFor multiplier, T_gFor third threshold value.

5. a kind of method improving image resolution ratio according to claim 1, which is characterized in that described to use interpolation method pair The cartoon subgraph is amplified, and is obtained amplified cartoon subgraph, is specifically included：

It uses interpolation stencil function into row interpolation the pixel of the cartoon subgraph, obtains the amplified cartoon subgraph Picture；The interpolation stencil function is：

L (x, y)=N₁*o(x,y)*d(x,y)*sinc(α*x/σ²)

*sinc(α*y/σ²)*exp(β*(x²+y²)/2σ²) (3)

Wherein, o (x, y) indicates whether the value using current pixel point to estimate the value of unknown pixel point, if current pixel point and Unknown pixel point is in the same object, then o (x, y)=1, that is, utilize current pixel point；Otherwise, o (x, y)=0, i.e., do not utilize Current pixel point, d (x, y) indicate the distance of current pixel point and interpolation template center point；Function sinc (x)=(sin (x)/ x)；σ²It is the mean square deviation participated in all pixels point of unknown pixel point interpolation；α and β is constant.

6. a kind of method improving image resolution ratio according to claim 1, which is characterized in that described to use homotopy Method The texture subgraph is amplified, amplified texture subgraph is obtained, specifically includes：

Obtain the image block dictionary of the texture subgraph, described image block dictionary packet offline using K-SVD dictionary learnings method Include high-definition picture block dictionary and low-resolution image block dictionary；

To the texture subgraph, amplified online using described image block dictionary and orthogonal matching tracking method, is obtained just Beginning high-definition picture；

The edged processing of arest neighbors is carried out to the initial high-resolution image；Obtain edged treated full resolution pricture；

Homotopy processing for the first time is carried out to the edged treated high-definition picture, obtains the first high-definition picture；

Second of homotopy processing is carried out to first high-definition picture, obtains amplified high-resolution texture subgraph.

7. it is according to claim 1 it is a kind of improve image resolution ratio method, which is characterized in that it is described to the amplification after Cartoon subgraph and the amplified texture subgraph synthesized, obtain high-definition picture, specifically include：

The amplified cartoon subgraph and the amplified texture subgraph are synthesized using formula (4), obtained High-definition picture；

f_H=f_c+f_t+λ₁*G(f_t) (4)

Wherein, f_HFor high-resolution image, f_tAmplified texture image, f_cFor amplified cartoon image, G (f_t) it is to figure As f_tAsk the modulus value of Robert gradients, λ₁For constant.

8. a kind of system improving image resolution ratio, which is characterized in that the system comprises：

Picture breakdown module obtains cartoon subgraph and texture for being decomposed to original image using opposite total variance method Image；

Cartoon subgraph amplification module obtains amplified card for being amplified to the cartoon subgraph using interpolation method Logical subgraph；

Texture subgraph amplification module is obtained amplified for being amplified to the texture subgraph using homotopy Method Texture subgraph；

Image synthesis unit, for being closed to the amplified cartoon subgraph and the amplified texture subgraph At acquisition high-definition picture.

9. a kind of system improving image resolution ratio according to claim 8, which is characterized in that described image decomposing module It specifically includes：

Image segmentation submodule, for being split to original image, the image after being divided；

Picture breakdown submodule, for using the full variational approach of image for minimizing window function, being carried out to the image after segmentation It decomposes, obtains cartoon subgraph and texture subgraph.

10. a kind of system improving image resolution ratio according to claim 9, which is characterized in that described image segmentation Module specifically includes：

Convolution unit, for being utilized respectively templateWithConvolution is carried out to original image, is obtained The gradient image I in the directions x_xWith the gradient image I in the directions y_y；

Gradient image magnitude calculation unit, the ladder of the gradient image of gradient image and the directions y for calculating the directions x Spend image modulus value I_g；

Wherein, (i, j) is pixel point coordinates；

Pixel selection unit, for the gradient image modulus value to be more than to the pixel of first threshold as edge pixel point, The gradient image modulus value is more than the pixel of second threshold as candidate edge pixel；

Closed area acquiring unit is respectively with each edge pixel point for choosing multiple edge pixel points Initial point, connection and the edge pixel o'clock the edge pixel point of direct neighbor or candidate edge picture on eight directions Vegetarian refreshments obtains several closed areas, and the small closed area in the big closed area is deleted, and acquisition is several not to be had There is the closed area of inclusion relation；

Image segmentation unit does grey level histogram for the closed area to each not inclusion relation, and to the ash Degree histogram is split using K-means methods, the image after being divided.