CN103198495B - The texture compression method that importance degree drives - Google Patents

Abstract

The importance-driven texture compression method consists of two parts: texture image compression and decompression. In the compression process, firstly calculate the control map and importance information map for the input texture, and then use the importance-driven texture compression algorithm to compress the original texture to obtain the corresponding compressed texture and control map; in the decompression stage, similar to the image In the analogy, the compressed control map, the compressed texture and the original control map are respectively used as input, and the decompressed texture output map is obtained through image analogy. Specifically, it includes: control graph calculation, importance graph calculation, importance-driven texture compression, and decompression based on image analogy.

Description

Importance-Driven Texture Compression Method

technical field

The present invention belongs to the field of information technology.

Background technique

Today's large-scale network-based scene rendering (such as Google Earth, street view technology, etc.) requires a large number of realistic texture images. There are two challenges in implementation: (1) the limitation of transmission bandwidth on the Internet; (2) the limitation of the storage space of the client to store these textures that need to be drawn. Corresponding techniques are required to reduce bandwidth and storage space.

Traditional texture compression algorithms usually use image coding techniques such as vector quantization to compress textures, and the compression rate is limited and GPU is not used for accelerated decompression drawing. The latest texture compression technology uses the idea of texture synthesis to compress texture images, and compresses texture images by extracting repeatedly used texel patterns in textures. This technology can not only compress the local redundant information of the image, but also compress the global repetitive information, and it is beneficial to use GPU to decompress and draw.

The reverse texture synthesis algorithm is based on the idea of texture synthesis, and compresses a large input texture into a small texture summary image through the optimization of the energy function. An important problem with inverse texture synthesis is that all regions are treated equally during the compression process. For some textures, not all regions are visually consistent and can be divided into foreground and background parts. From the visual saliency model, it can be seen that the foreground part of the texture is often the part that human eyes pay attention to, and it should be preserved as much as possible during compression. Therefore, it is necessary to perform adaptive compression on different parts of the texture when compressing. Finally, the control map plays a key role in the reverse synthesis framework, and the original texture can only be reconstructed by reverse synthesis based on a proper control map. In a network environment, texture decompression on the client side requires a suitable control map. Inverse synthesis does not propose a more general solution for the generation of control maps, which limits the application range of inverse synthesis in the field of texture compression.

In view of the above problems in reverse synthesis, no researchers have proposed an adaptive fast texture compression algorithm for real-time decompression rendering.

The invention modifies the energy function of reverse synthesis to make it suitable for adaptive texture compression, that is, to compress important regions and non-important regions of texture in different degrees. In the calculation stage, the texture importance map is first calculated according to the visual saliency model, and then the importance map, texture and its control map are used as input to iteratively optimize the new energy function.

Aiming at the control map problem of reverse synthesis, we propose to select the grayscale image reflecting the brightness change of image pixels as the control map.

Contents of the invention

The present invention overcomes the above-mentioned shortcomings of the prior art, and provides a method for compressing textures, effectively reducing texture memory occupied by large-scale texture images, realizing fast transmission on the network and real-time rendering on the GPU, and importance-based driving The adaptive texture compression method makes the image quality of the important area keep better after the texture image is decompressed.

The importance-driven texture compression method of the present invention consists of two parts: texture image compression and decompression. In the compression process, firstly calculate the control map and importance information map for the input texture, and then use the importance-driven texture compression algorithm to compress the original texture to obtain the corresponding compressed texture and control map; in the decompression stage, similar to the image For analogy, the compressed control map, the compressed texture and the original control map are respectively used as input, and the decompressed texture output map is obtained through image analogy; the specific steps are as follows:

Step 1. Control chart calculation

All the control charts of the present invention are grayscale images of the original image, and the grayscale image well preserves the brightness details of the original image; the YIQ calculation model is used to convert the original image (color image) into a grayscale image, and Y in the model Represents brightness, that is, the required grayscale information, I represents hue, and Q represents saturation. According to the corresponding model conversion matrix, the calculation formula for converting RGB into Y is as formula (1):

Y=0.299R+0.587G+0.114B (1)

Among them, R, G, and B are the three colors of red, green and blue respectively.

Step 2. Importance map calculation

The importance information of the texture directly affects the final result of the algorithm. In order to get better importance information, we first compute the saliency map of the image, which is consistent with the resolution of the original image. Calculate the saliency map based on the Saliency Filters algorithm: first abstract the image, that is, retain the structural features related to the image, and remove some unnecessary details of the image; then calculate the distribution of the underlying information such as unique elements and colors in the image ; Finally, combine the two information to get the saliency map saliency.

Assuming that the input image is X, the present invention uses formula (2) to calculate the importance value w(x,y) of each pixel in the importance map:

$\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; saliency</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span><span\; class="notranslate">\; |</span>\frac{<span\; class="notranslate">\; \&PartialD;</span>}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; x</span>}}\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; +</span><span\; class="notranslate">\; |</span>\frac{<span\; class="notranslate">\; \&PartialD;</span>}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}}\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

Among them, saliency(x, y) is the saliency value at the pixel (x, y), k is the weight of saliency information, usually k=1 can achieve the ideal effect; I(x, y) is the image pixel (x, y) y), the last two terms of the equation and is the absolute value of the x-direction component and y-direction component of the gradient value at the pixel (x, y). Further, we normalize the importance value calculated by formula (2), so that the value of each pixel in the importance map ranges between [0,1]. The larger the value of w(x,y), the higher the importance of the corresponding pixel.

Step 3. Importance-driven texture compression

After obtaining the importance map and the control map, compress the original map and the control map;

In this method, texture compression is defined as an energy function optimization problem. The energy function calculates the bidirectional similarity between all the pixel neighborhoods of the original image and the pixel neighborhood of the compressed image; for the input image X (it can be the original image or the control image ), the energy function between it and the compression target graph Z considering the importance information is shown in formula (4).

Where z/x is the sample value of Z/X, q/p are the pixel positions in the subset Z ⁺ /X ⁺ of Z/X respectively, and x _p /z _q represents the spatial neighborhood centered at point p/q , z _p /x _q is the most similar neighborhood to x _p /z _q in Z/X, α is a user-adjustable weight value, when α=0.01 is suitable for most textures; is the importance weight of the p-point neighborhood x _p , there are many ways to calculate the importance of an image area, such as taking the minimum/large value, median or mean value of all pixel weights in the neighborhood, etc., we take the average value of the pixel weight as the neighborhood Importance of x _p ; Use the same method to calculate the importance of the region x _q |·| ² Calculate the distance between two neighborhoods by summing the squares of the color differences of corresponding pixels in the neighborhood to calculate the distance between two neighborhoods.

The energy function of formula (3) is composed of the addition of two terms, which are similar in form, but have different calculation functions;

The first item is called the inverse item, which ensures that each neighborhood x _p in the input image X can find a similar z _p in Z; the second item is called the forward item, which ensures that there is no new one in Z z _q is not similar to any x _q in X;

Based on formula (3), the method for calculating the color value at each pixel q in the target texture Z is deduced below. For each pixel q∈Z in the target texture, its contribution to the overall energy value also includes the calculation of forward and inverse. The specific contribution value can be obtained by the following steps:

(1) The contribution of pixel q to the forward term

Represents all neighborhoods containing pixel q in the target graph Z (where q ₁ ,...,q _m are neighborhoods Note that here q ₁ ,...,q _m need to go through the corresponding offset to get q), and the value of m is the neighborhood The number of , which is related to the size of the neighborhood we choose, when the size of the neighborhood is 5x5, m=25 (in the present invention, each layer of the Gaussian pyramid has 2 groups of neighborhoods of different sizes, respectively 17x17, 9x9). Area The nearest neighbor in X is express neutralize The pixel position corresponding to q in (as shown in Figure 2). Thus, the contribution of q to the forward term is $\frac{\mathrm{\&alpha;}}{{\mathrm{\&Sigma;}}_{q\&Element;Z^{+}}{w}_{x_q}}\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}w\left({\stackrel{\&And;}{p}}_i\right){(I_x\left({\stackrel{\&And;}{p}}_i\right)-I_Z\left(q\right))}^2$ (where w( ) is the importance value of the pixel, I _X (), I _Z () respectively give the pixel color values of the input image X and the target image Z, the same below);

(2) Contribution of pixel q to the inverse term

are the neighborhoods in X and the nearest neighbors of these neighborhoods in Z are the neighborhoods containing q, where n is the neighborhood The number of is different from the m value above, and the size of the n value is not fixed, and varies with q. Pixels p ₁ ,...,p _n are neighbors The center point of , they also get the pixel points through the corresponding offset These pixels correspond to the pixel q in Z (see Figure 2). In this way, we can get the contribution value of q to inverse term as $\frac{1}{{\mathrm{\&Sigma;}}_{p\&Element;x^{+}}{w}_{x_p}}\underset{j=1}{\overset{\mathrm{no}}{\mathrm{\&Sigma;}}}w\left({\stackrel{\&OverBar;}{p}}_j\right){(I_x\left({\stackrel{\&OverBar;}{p}}_j\right)-I_Z\left(q\right))}^2;$

In summary, the energy of a single pixel q is the sum of the above forward and inverse items, and the specific calculation is as follows:

$\mathrm{<span\; class="notranslate">\; 能源</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; Z</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>$

$\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; +</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; Z</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\frac{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; Z</span>}}^{<span\; class="notranslate">\; +</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}}}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&And;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&And;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; Z</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

The color value at pixel q can be solved for by finding the minimum of Energy(I _Z (q)). Taking the derivative of the equation with respect to I _Z (q) and equaling it to 0, the solution of I _Z (q) can be obtained as:

${\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; Z</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; +</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\frac{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; Z</span>}}^{<span\; class="notranslate">\; +</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}}}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&And;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&And;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; +</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\frac{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; Z</span>}}^{<span\; class="notranslate">\; +</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}}}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&And;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

(3) Energy optimization method

The iterative energy optimization steps for solving I _Z (q) based on formula (5) are:

Step1: For each target pixel q∈Z, search for the neighborhood x _q most similar to its neighborhood z _q in the input image X. Multiply the color value of the pixel in x _q by the importance weight of the pixel and multiply by the weight Then vote the calculated value to the corresponding pixel in z _q ;

Step2: In contrast to step1, for each pixel p∈X, search for the neighborhood z _p that is most similar to its neighborhood x _p in Z. Multiply the pixel color value of x _p by the importance weight of the pixel and multiply , and then vote the calculated value to the corresponding pixel in z _p ;

step3: For each target pixel q∈Z, calculate the average of all the votes, that is, divide by $\frac{1}{\underset{p\&Element;x^{+}}{\mathrm{\&Sigma;}}{w}_{x_p}}\underset{j=1}{\overset{\mathrm{no}}{\mathrm{\&Sigma;}}}w\left({\stackrel{\&OverBar;}{p}}_j\right)+\frac{\mathrm{\&alpha;}}{\underset{q\&Element;Z^{+}}{\mathrm{\&Sigma;}}{w}_{x_q}}\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}w\left({\stackrel{\&And;}{p}}_i\right),$ Get the new I _z (q);

Repeat step1, step2, step3 until convergence.

Step 4. Decompression based on image analogy

Given a pair of pictures A and A' (respectively corresponding to the original picture and the picture generated after certain processing), input another unprocessed target picture B, and calculate a new picture B by image analogy method ', making:

A:A'::B:B'

That is, the relationship between B' and B is similar to the relationship between A' and A. The image analogy algorithm can learn the relationship between A and A', and apply the learned result to generate B'. Here we apply it to the decompression of compressed data.

We correspond the compressed control map and the compressed texture to A and A' in the image analogy, and use the original control map as B, and use the image analogy to synthesize a new B', which is the decompression result image.

The original image is recovered through the image analogy algorithm, so as to obtain the decompressed result.

The invention describes an importance-driven texture compression algorithm, which maps the importance-driven texture compression into an optimization problem of an energy function. Integrate the importance information into the energy function, and re-deduce the solver of the energy function, and use the grayscale image as the control map of the texture. Use formula (2) to calculate the importance map. The method of the invention can well preserve important areas in the texture when compressing the texture, and the compression efficiency is high.

The advantages of the present invention are as follows:

(1) Adaptive compression for important information preservation. The texture image is adaptively compressed by using the concept of importance, and the compression quality of important areas in the image is high during the compression process. Different from previous compression algorithms, the compression result of the algorithm is also a meaningful texture, which can be used to synthesize new textures.

(2) Novel idea, non-classical method: This method uses energy optimization to compress texture, and obtains the compressed texture result by iteratively solving the minimum value of the energy function. The idea is novel and innovative.

(3) The implementation is simple and the operation efficiency is high. The energy optimization calculation is simple, and the EM-like algorithm is used for iterative optimization to solve the energy minimum value, and only a simple design of step3, step1 and step2 is required. The search calculation in energy optimization has a high degree of parallelism, which is conducive to the use of GPU for parallel accelerated calculation, thereby greatly improving the operating efficiency of the algorithm.

(4) Convenient and easy. The user only needs to provide an image, and the algorithm will automatically calculate its control map and semi-automatically calculate the importance map through user interaction, and then perform importance-driven texture compression. In the decompression stage, only the control map of the original image is provided. The original image can be resynthesized.

(5) Good controllability: the user does not need complex input parameters to control the algorithm, just change the importance map, and the algorithm will perform high-fidelity compression on the important areas determined by the importance map during compression. Users can control the adaptive compression of images by editing the importance map.

Description of drawings

Fig. 1 is a technical framework diagram of the present technical invention.

Figure 2 is a labeled map of the pixel correspondence between the input map and the target map during the derivation process.

Detailed ways

As shown in the figure, the invention consists of two parts: texture image compression and decompression. In the compression process, the input texture first calculates its control map and importance information map, and then uses the importance-driven texture compression algorithm to compress the original texture to obtain the corresponding compressed texture and control map. In the decompression stage, similar to the image analogy, the compressed control map, the compressed texture and the original control map are respectively used as input, and the decompressed texture output map is obtained through image analogy. Below we specifically introduce the four calculation steps of the method of the present invention.

Step 1. Control chart calculation

All the control charts of the present invention are grayscale images of the original image, and the grayscale image well preserves the brightness details of the original image; the YIQ calculation model is used to convert the original image (color image) into a grayscale image, and Y in the model Represents brightness, that is, the required grayscale information, I represents hue, and Q represents saturation. According to the corresponding model conversion matrix, the calculation formula for converting RGB into Y is as formula (1):

Y=0.299R+0.587G+0.114B (1)

Among them, R, G, and B are the three colors of red, green and blue respectively.

Step 2. Importance map calculation

The importance information of the texture directly affects the final result of the algorithm. In order to get better importance information, we first compute the saliency map of the image, which is consistent with the resolution of the original image. Calculate the saliency map based on the Saliency Filters algorithm: first abstract the image, that is, retain the structural features related to the image, and remove some unnecessary details of the image; then calculate the distribution of the underlying information such as unique elements and colors in the image ; Finally, combine the two information to get the saliency map saliency.

Assuming that the input image is X, the present invention uses formula (2) to calculate the importance value w(x,y) of each pixel in the importance map:

$\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; saliency</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span><span\; class="notranslate">\; |</span>\frac{<span\; class="notranslate">\; \&PartialD;</span>}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; x</span>}}\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; +</span><span\; class="notranslate">\; |</span>\frac{<span\; class="notranslate">\; \&PartialD;</span>}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}}\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

Among them, saliency(x, y) is the saliency value at the pixel (x, y), k is the weight of saliency information, usually k=1 can achieve the ideal effect; I(x, y) is the image pixel (x, y) y), the last two terms of the equation and is the absolute value of the x-direction component and y-direction component of the gradient value at the pixel (x, y). Further, we normalize the importance value calculated by formula (2), so that the value of each pixel in the importance map ranges between [0,1]. The larger the value of w(x,y), the higher the importance of the corresponding pixel.

Step 3. Importance-driven texture compression

After obtaining the importance map and the control map, compress the original map and the control map;

In this method, texture compression is defined as an energy function optimization problem. The energy function calculates the bidirectional similarity between all the pixel neighborhoods of the original image and the pixel neighborhood of the compressed image; for the input image X (it can be the original image or the control image ), the energy function between it and the compression target graph Z considering the importance information is shown in formula (4).

Where z/x is the sample value of Z/X, q/p are the pixel positions in the subset Z ⁺ /X ⁺ of Z/X respectively, and x _p /z _q represents the spatial neighborhood centered at point p/q , z _p /x _q is the most similar neighborhood to x _p /z _q in Z/X, α is a user-adjustable weight value, when α=0.01 is suitable for most textures; is the importance weight of the p-point neighborhood x _p , there are many ways to calculate the importance of an image area, such as taking the minimum/large value, median or mean value of all pixel weights in the neighborhood, etc., we take the average value of the pixel weight as the neighborhood Importance of x _p ; Calculate the importance of the region x _q in the same way |·| ² Calculate the distance between two neighborhoods by summing the squares of the color differences of corresponding pixels in the neighborhood to calculate the distance between two neighborhoods.

The energy function of formula (3) is composed of the addition of two terms, which are similar in form, but have different calculation functions;

The first item is called the inverse item, which ensures that each neighborhood x _p in the input image X can find a similar z _p in Z; the second item is called the forward item, which ensures that there is no new one in Z z _q is not similar to any x _q in X;

Based on formula (3), the method for calculating the color value at each pixel q in the target texture Z is deduced below. For each pixel q∈Z in the target texture, its contribution to the overall energy value also includes the calculation of forward and inverse. The specific contribution value can be obtained by the following steps:

(1) The contribution of pixel q to the forward term

Indicates all neighborhoods containing pixel q in the target graph Z (where q ₁ ,...,q _m are neighborhoods Note that here q ₁ ,...,q _m need to go through the corresponding offset to get q), and the value of m is the neighborhood The number of , which is related to the size of the neighborhood we choose, when the size of the neighborhood is 5x5, m=25 (in the present invention, each layer of the Gaussian pyramid has 2 groups of neighborhoods of different sizes, respectively 17x17, 9x9). Area The nearest neighbor in X is express neutralize The pixel position corresponding to q in (as shown in Figure 2). Thus, the contribution of q to the forward term is $\frac{\mathrm{\&alpha;}}{{\mathrm{\&Sigma;}}_{q\&Element;Z^{+}}{w}_{x_q}}\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}w\left({\stackrel{\&And;}{p}}_i\right){(I_x\left({\stackrel{\&And;}{p}}_i\right)-I_Z\left(q\right))}^2$ (where w( ) is the importance value of the pixel, I _X (), I _Z () respectively give the pixel color values of the input image X and the target image Z, the same below);

(2) Contribution of pixel q to the inverse term

are the neighborhoods in X and the nearest neighbors of these neighborhoods in Z are the neighborhoods containing q, where n is the neighborhood The number of is different from the m value above, and the size of the n value is not fixed, and varies with q. Pixels p ₁ ,...,p _n are neighbors The center point of , they also get the pixel points through the corresponding offset These pixels correspond to the pixel q in Z (see Figure 2). In this way, we can get the contribution value of q to the inverse term as $\frac{1}{{\mathrm{\&Sigma;}}_{p\&Element;x^{+}}{w}_{x_p}}\underset{j=1}{\overset{\mathrm{no}}{\mathrm{\&Sigma;}}}w\left({\stackrel{\&OverBar;}{p}}_j\right){(I_x\left({\stackrel{\&OverBar;}{p}}_j\right)-I_Z\left(q\right))}^2;$

In summary, the energy of a single pixel q is the sum of the above forward and inverse items, and the specific calculation is as follows:

$\mathrm{<span\; class="notranslate">\; 能源</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; Z</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>$

$\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; +</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; Z</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\frac{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; Z</span>}}^{<span\; class="notranslate">\; +</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}}}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&And;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&And;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; Z</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

The color value at pixel q can be solved for by finding the minimum of Energy(I _Z (q)). Taking the derivative of the equation with respect to I _Z (q) and equaling it to 0, the solution of I _Z (q) can be obtained as:

${\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; Z</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; +</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\frac{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; Z</span>}}^{<span\; class="notranslate">\; +</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; w</span>}}}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&And;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&And;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; +</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\frac{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; Z</span>}}^{<span\; class="notranslate">\; +</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}}}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&And;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

(3) Energy optimization method

The iterative energy optimization steps for solving I _Z (q) based on formula (5) are:

Step1: For each target pixel q∈Z, search for the neighborhood x _q most similar to its neighborhood z _q in the input image X. Multiply the color value of the pixel in x _q by the importance weight of the pixel and multiply by the weight Then vote the calculated value to the corresponding pixel in z _q ;

Step2: In contrast to step1, for each pixel p∈X, search for the neighborhood z _p that is most similar to its neighborhood x _p in Z. Multiply the pixel color value of x _p by the importance weight of the pixel and multiply Then vote the calculated value to the corresponding pixel in _zp ;

step3: For each target pixel q∈Z, calculate the average of all the votes, that is, divide by $\frac{1}{\underset{p\&Element;x^{+}}{\mathrm{\&Sigma;}}{w}_{x_p}}\underset{j=1}{\overset{\mathrm{no}}{\mathrm{\&Sigma;}}}w\left({\stackrel{\&OverBar;}{p}}_j\right)+\frac{\mathrm{\&alpha;}}{\underset{q\&Element;Z^{+}}{\mathrm{\&Sigma;}}{w}_{x_q}}\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}w\left({\stackrel{\&And;}{p}}_i\right),$ Get the new I _z (q);

Repeat step1, step2, step3 until convergence.

Step 4. Decompression based on image analogy

Given a pair of pictures A and A' (respectively corresponding to the original picture and the picture generated after certain processing), input another unprocessed target picture B, and calculate a new picture B by image analogy method ', making:

A:A'::B:B'

That is, the relationship between B' and B is similar to the relationship between A' and A. The image analogy algorithm can learn the relationship between A and A', and apply the learned result to generate B'. Here we apply it to the decompression of compressed data.

We correspond the compressed control map and the compressed texture to A and A' in the image analogy, and use the original control map as B, and use the image analogy to synthesize a new B', which is the decompression result image.

The original image is recovered through the image analogy algorithm, so as to obtain the decompressed result.

Claims (1)

1. The importance-driven texture compression method consists of two parts: texture image compression and decompression; in the compression process, first calculate its control map and importance information map for the input texture, and then use the importance-driven texture compression algorithm Compress the original texture to obtain the corresponding compressed texture and control map; in the decompression stage, similar to the image analogy, the compressed control map, the compressed texture and the original control map are respectively used as input, and the decompressed texture is obtained by image analogy Output graph; the specific steps are as follows: Step 1. Control chart calculation All control charts are grayscale images of the original image, and the grayscale image well preserves the brightness details of the original image; the original image (color image) is converted into a grayscale image by using the YIQ calculation model, in which Y represents brightness; According to the corresponding model conversion matrix, the calculation formula for converting RGB to Y is as formula (1): Y＝0.299R+0.587G+0.114B (1) Among them, R, G, and B are three colors of red, green and blue respectively; Step 2. Importance map calculation The importance information of the texture directly affects the final result of the algorithm. In order to obtain better importance information, first calculate the saliency map of the image, which is consistent with the resolution of the original image; calculate the saliency map based on the Saliency Filters algorithm: Firstly, the image is abstracted, that is, the relevant structural features of the image are retained, and some unnecessary details of the image are removed; then the distribution of unique elements and color underlying information in the image is calculated; finally, the two information is combined to obtain a significant Sex figure saliency; Assuming that the input image is X, use formula (2) to calculate the importance value w(x,y) of each pixel in the importance map: Among them, saliency(x, y) is the saliency value at the pixel (x, y), k is the weight of saliency information, usually k=1 can achieve the ideal effect; I(x, y) is the image pixel (x, y) y), the last two terms of the equation and is the absolute value of the x-direction component and the y-direction component of the gradient value at the pixel (x, y); further, the importance value calculated by the formula (2) is normalized, so that the value of each pixel in the importance map The range is between [0,1]; the larger the value of w(x,y), the higher the importance of the corresponding pixel; Step 3. Importance-driven texture compression After obtaining the importance map and the control map, compress the original map and the control map; In this method, texture compression is defined as an energy function optimization problem. The energy function calculates the bidirectional similarity between all the pixel neighborhoods of the original image and the pixel neighborhood of the compressed image; for the input image X, it is the original image or the control image, The energy function considering the importance information between it and the compression target graph Z is shown in formula (3): Where z/x is the sample value of Z/X, q/p are the pixel positions in the subset Z ⁺ /X ⁺ of Z/X respectively, and x _p /z _q represents the spatial neighborhood centered at point p/q , z _p /x _q is the most similar neighborhood to x _p /z _q in Z/X, and α is a user-adjustable weight value; is the importance weight of the neighborhood x _p of point p, and takes the average value of the pixel weight as the importance of the neighborhood x _p ; Use the same method to calculate the importance of the region x _q ; |·| ² calculates the distance between two neighborhoods, and calculates the distance between two neighborhoods by summing the squares of the color differences of corresponding pixels in the neighborhood; The energy function of formula (3) is composed of the addition of two terms, which are similar in form, but have different calculation functions; The first item is called the inverse item, which ensures that each neighborhood x _p in the input image X can find a similar z _p in Z; the second item is called the forward item, which ensures that there is no new one in Z z _q is not similar to any x _q in X; Based on formula (3), the method for calculating the color value of each pixel q in the target texture Z is deduced below. For each pixel q∈Z in the target texture, its contribution to the overall energy value also includes the forward and inverse two The calculation of the item, the specific contribution value can be obtained by the following steps: (1) The contribution of pixel q to the forward term Represents all neighborhoods containing pixel q in the target image Z (where q ₁ ,...,q _m are neighborhoods Note that here q ₁ ,...,q _m need to go through the corresponding offset to get q), and the value of m is the neighborhood The number of , which is related to the size of the selected neighborhood, when the size of the neighborhood is 5x5, m=25; the neighborhood The nearest neighbor in X is express neutralize The pixel position corresponding to q in ; thus, the contribution value of q to the forward term is Where w( ) is the importance value of the pixel, I _X (), I _Z () respectively give the pixel color values of the input image X and the target image Z, the same below; (2) Contribution of pixel q to the inverse term are the neighborhoods in X and the nearest neighbors of these neighborhoods in Z are the neighborhoods containing q, where n is the neighborhood The number of; pixels p ₁ ,...,p _n are neighbors The center point of , they also get the pixel points through the corresponding offset These pixels correspond to the pixel q in Z, so the contribution value of q to the inverse term can be obtained as In summary, the energy of a single pixel q is the sum of the above forward and inverse items, and the specific calculation is as follows: ; Deriving the equation with respect to I _Z (q) and equal to 0, the solution of I _Z (q) can be obtained as: (3) Energy optimization method The iterative energy optimization steps for solving I _Z (q) based on formula (5) are: Step1: For each target pixel q∈Z, search for the neighborhood x _q most similar to its neighborhood z _q in the input image X, multiply the color value of the pixel in x _q by the importance weight of the pixel and multiply it by the weight Then vote the calculated value to the corresponding pixel in z _q ; step2: Contrary to step1, for each pixel p∈X, search for the neighborhood z _p that is most similar to its neighborhood x _p in Z, multiply the pixel color value of x _p by the importance weight of the pixel and multiply Then vote the calculated value to the corresponding pixel in _zp ; step3: For each target pixel q∈Z, calculate the average of all the votes, that is, divide by Get the new I _z (q); Repeat step1, step2, step3 until convergence; Step 4. Decompression based on image analogy Given a pair of pictures A and A', corresponding to the original picture and the picture generated after processing it, input another unprocessed target picture B, and calculate a new picture B' by image analogy method, Make the relationship between B' and B similar to the relationship between A' and A; the image analogy algorithm can learn the relationship between A and A', and apply the learning result to generate B'; here it is applied to the decompression of compressed data ; The compressed control map and the compressed texture correspond to A and A' in the image analogy, and the original control map is used as B, and a new B' is synthesized using the image analogy, and B' is the decompression result image; The original image is recovered through the image analogy algorithm, so as to obtain the decompressed result.