CN104217221A - Method for detecting calligraphy and paintings based on textural features - Google Patents

Abstract

The invention provides a method for detecting calligraphy and painting works based on texture features, comprising the following steps: obtaining a digital image of rice paper of authentic works of calligraphy and painting; preprocessing the digital image of rice paper to obtain a digital image of rice paper after preprocessing; using the SURF algorithm to extract Texture features: Match the texture features of the calligraphy and painting works to be detected. If they match, the calligraphy and painting works to be detected are genuine; if they do not match, the calligraphy and painting works to be detected are fakes. The method for detecting calligraphy and painting works based on texture features provided by the present invention is used to detect the authenticity of contemporary calligraphy and painting works, extract features from authentic calligraphy and painting works, and use the two-way FLANN algorithm combined with clustering of feature points to obtain the center of each category. Taking the center as the range to intercept the square area, judging the authenticity of the entire calligraphy and painting with the highest similarity of the square area, it can quickly and accurately identify the authenticity of the calligraphy and painting works, and can prevent criminals from forging calligraphy and painting works.

Description

Detection method of calligraphy and painting works based on texture features

technical field

The invention relates to a method for detecting calligraphy and painting works, in particular to a method for detecting calligraphy and painting works based on texture features, which is used to detect the authenticity of contemporary calligraphy and painting works.

Background technique

Calligraphy and painting works have always occupied a place in Chinese art works since ancient times. According to relevant data, calligraphy and painting works account for 70% to 80% of the Chinese art auction market.

In recent years, with the development of the social economy and the continuous improvement of the quality of people's living standards, more and more people and teachers are pursuing the art of calligraphy and painting. A variety of famous calligraphy and paintings began to emerge on the market, but people still know little about the identification of true and false calligraphy and paintings.

The development of science and technology has also made great progress in printing technology, which provides opportunities for criminals to make profits. They can easily use high-tech printing equipment to clone famous calligraphy and painting works, which is easier and more numerous than manual cloning. Also bigger. Pan Baoqing, executive director of Sichuan Tongqing Cultural Appraisal and Evaluation Co., Ltd., once said that the current Chinese art auction market is a bit like a runaway wild horse! Fakes and counterfeit works have flooded the market, and it is even more difficult to buy real genuine products.

According to relevant data, fakes of Chinese paintings and calligraphy are flooding the art auction market, and the more well-known writers, the higher the rate of fakes. And those small art auction houses, it is difficult to see an authentic work in individual special sales. Therefore, the lagging of authenticity identification technology for valuable collections such as calligraphy and paintings in the market is one of the main contradictions hindering and restricting their circulation and trading in the cultural and creative market.

Existing anti-counterfeiting technologies mainly include infrared anti-counterfeiting technology, DNA anti-counterfeiting technology and fingerprint recognition technology.

Infrared anti-counterfeiting technology is an earlier anti-counterfeiting technology. It uses infrared technology to make text, graphics, barcodes, etc. that are visible or invisible to the naked eye, and then uses special infrared toner, ink, and ink to print them on calligraphy and painting works. In this way, it can be detected by special instruments during inspection.

DNA anti-counterfeiting technology is related to biological knowledge. It is to transplant special DNA fragments on the works, so that the items also have the characteristics of DNA. Since an item corresponds to a set of genes, and each pair of genes is unique, the item will be With unique features, the difficulty of inspection is greatly reduced.

Fingerprint identification technology, when a calligrapher and painter completes his work, he dips his finger in the inkpad and stamps his own fingerprint impression under the seal. In this way, calligraphy and painting identification can be carried out through fingerprint identification.

But above-mentioned method all is to artificially add materials other than painting and calligraphy to assist the authenticity of judging painting and calligraphy, still there is the possibility that criminals utilize high technology to forge painting and calligraphy.

Those skilled in the art are committed to providing an anti-counterfeiting method and system that can quickly and accurately identify the authenticity of paintings and calligraphy and does not give criminals the opportunity to take advantage of loopholes, thereby effectively ensuring that paintings and calligraphy in auction houses are authentic.

Contents of the invention

The purpose of the present invention is to provide a method for detecting calligraphy and painting works based on texture features, which is used to detect the authenticity of contemporary calligraphy and painting works. In the detection process, using the features extracted from the authentic works of calligraphy and painting, using the two-way FLANN algorithm and combining the feature points to cluster to obtain the center of each type, intercepting the square area with the center as the range, and judging the entire square area with the highest similarity The authenticity of calligraphy and painting can quickly and accurately identify the authenticity of calligraphy and painting works.

The invention provides a method for detecting calligraphy and painting works based on texture features, comprising the following steps:

(1) Obtain digital images of Xuan paper of authentic works of calligraphy and painting;

(2) carry out preprocessing to the rice paper digital image, obtain the rice paper digital image after preprocessing;

(3) Utilize the SURF algorithm to extract texture features from the preprocessed rice paper digital image;

(4) Match the texture features of the painting and calligraphy works to be detected. If they match, the painting and calligraphy works to be detected are authentic works of calligraphy and painting; if they do not match, the painting and calligraphy works to be detected are fakes.

Further, in step (4), a two-way Flann matching algorithm is used to perform texture feature matching on the calligraphy and painting works to be detected.

Further, before performing texture feature matching in step (4), the following steps are also included:

(41) utilize improved K-Means algorithm to carry out clustering, obtain the center of each class;

(42) Through the center of the class obtained in step (41), extract the same area of the authentic painting and calligraphy work as that of the painting and calligraphy work to be detected.

Further, step (2) preprocesses the rice paper digital image using one or a combination of color image conversion to grayscale image, histogram equalization, morphological transformation, edge detection or Gabor filtering.

Further, step (3) utilizes the SURF algorithm to extract texture features from the preprocessed rice paper digital image, comprising the following steps:

(31) Utilize integral image to calculate Haar feature;

(32) Construct a Hessian matrix to obtain extreme points;

(33) Gaussian filtering is used to find out the basis for judging whether the point (x, y) is a local extremum point;

(34) feature points are extracted;

(35) Acquiring the characteristic direction of the characteristic point;

(36) Construct a SURF feature point description operator.

Further, step (35) obtaining feature direction comprises the following steps:

(351) Take the feature point as the center of the circle, and take six times the scale of the feature point as the radius to establish a circle area;

(352) scanning the circular area with a 60-degree fan-shaped area and a set interval;

(353) The maximum value of the harr wavelet response is four times the scale of the feature point, and the weight is given according to the distance from the center of the circle. The farther away from the center of the circle, the smaller the weight;

(354) adding the Harr wavelet response vectors of each point in the fan-shaped area;

(355) Find the Harr wavelet response vector and the largest fan-shaped area, and the vector direction of the fan-shaped area is the characteristic direction of the feature point.

Further, the method for extracting feature points in step (34) comprises the following steps:

(341) Compare the local extremum point with 8 adjacent points in the same scale and 18 points in the upper and lower two adjacent scales. If the local extremum point is an extremum point, then the local extremum point is preliminary Extreme point;

(342) After obtaining all the preliminary feature points, use the three-dimensional linear difference method to obtain sub-pixel-level feature points, and then remove the points smaller than the set threshold to obtain the final feature points.

Compared with the prior art, the method for detecting calligraphy and painting works based on texture features provided by the present invention has the following beneficial effects:

(1) After the author completes the calligraphy and painting works, feature extraction is performed on the authentic works of calligraphy and painting works. In the detection process of the calligraphy and painting works to be detected, the features extracted from the authentic works of calligraphy and painting works are used, and the feature points are clustered by using the two-way FLANN algorithm and combining them. Obtain each type of center, intercept the square area with the center as the range, judge the authenticity of the entire calligraphy and painting with the highest similarity of the square area, and can quickly and accurately identify the authenticity of contemporary calligraphy and painting works;

(2) There is no need to artificially add materials other than calligraphy and painting to assist in judging the authenticity of calligraphy and painting works. Since the rice paper texture characteristics of authentic calligraphy and painting works can hardly be copied, it can prevent criminals from using high technology to forge calligraphy and painting works.

Description of drawings

Fig. 1 is a flowchart of an embodiment of the present invention;

Figure 2 is an optical microscope-based image before histogram equalization;

Figure 3 is an optical microscope-based image after histogram equalization;

Figure 4 is an electron microscope-based image before histogram equalization;

Figure 5 is an electron microscope-based image after histogram equalization;

Figure 6 is the image before corrosion operation;

Figure 7 is the corrosion structure of the corrosion operation;

Fig. 8 is the image after corrosion operation;

Fig. 9 is the image after image thresholding based on the optical microscope image shown in Fig. 3;

Fig. 10 is the image after performing the opening operation based on the optical microscope image shown in Fig. 9;

Fig. 11 is the image after image thresholding based on the electron microscope image shown in Fig. 5;

Fig. 12 is the image after performing the opening operation based on the electron microscope image shown in Fig. 11;

Fig. 13 is the image after carrying out Laplacian edge detection based on the optical microscope image shown in Fig. 9;

Fig. 14 is the image after performing Laplacian edge detection based on the electron microscope image shown in Fig. 11;

Fig. 15 is the image after carrying out canny edge detection based on the optical microscope image shown in Fig. 9;

Fig. 16 is the image after canny edge detection based on the electron microscope image shown in Fig. 11;

Fig. 17 is the image after carrying out gabor filter based on optical microscope image shown in Fig. 9;

Fig. 18 is the image after carrying out gabor filtering based on the electron microscope image shown in Fig. 11;

Figure 19 is a Haar feature category diagram;

Fig. 20 is an integral explanatory diagram;

Figure 21 is the y-direction template of Gaussian filtering;

Figure 22 is the second-order mixed partial derivative template of Gaussian filtering;

Figure 23 is a scale pyramid;

Figure 24 is a scale pyramid;

Fig. 25 is a positional relationship diagram of local extremum points in adjacent points in the scale pyramid;

Figure 26a is a schematic diagram of characteristic directions;

Figure 26b is a schematic diagram of characteristic directions;

Figure 26c is a schematic diagram of characteristic directions;

Figure 27 is a schematic diagram of the operator direction;

Figure 28 is the matching result after preprocessing without rotation group;

Figure 29 is the matching result after preprocessing with rotation group;

Figure 30 is the effect of clustering texture feature points into four categories;

Figure 31 is the result of clustering texture feature points into four categories and extracting regions.

Detailed ways

The following are specific embodiments of the present invention and the technical solution of the present invention is further described in conjunction with the accompanying drawings, but the present invention is not limited to the following embodiments.

As shown in Figure 1, the detection method of the calligraphy and painting work based on texture feature of an embodiment of the present invention, comprises the following steps:

(1) Obtain the digital image of rice paper of authentic calligraphy and painting works;

(2) carry out preprocessing to the rice paper digital image, obtain the rice paper digital image after preprocessing;

(3) Utilize the SURF algorithm to extract texture features from the preprocessed rice paper digital image;

(4) Match the texture features of the calligraphy and painting works to be detected. If they match, the calligraphy and painting works to be detected are authentic; if they do not match, the calligraphy and painting works to be detected are counterfeit.

In step (1), the digital image of the rice paper of the authentic calligraphy and painting work can be obtained by using an optical microscope or an electron microscope.

An optical microscope uses the principle of optics to magnify tiny objects that cannot be discerned by the human eye into an image visible to the human eye, and then convert the image into a digital image through a series of processing. Because rice paper is made of sawdust, fiber, mudstone, etc., its optical image contains many impurities, and the contrast is low, and the sense of layering is poor, so the texture features are not obvious.

Electron microscopes are electron-optical devices that generate signals through the interaction of electrons with matter. Compared with optical microscopes, electromagnetic lenses are used instead of optical lenses to image electron beams invisible to our naked eyes onto fluorescent screens. Therefore, the image is a grayscale image with obvious texture features, high layering, and high depth of field.

In this embodiment, an electron microscope is used to obtain digital images of rice paper of genuine calligraphy and painting works.

Step (2) preprocesses the digital image of rice paper, the purpose is to eliminate the unimportant information in the image, strengthen the real and useful important information, and then enhance the detectability of the required information, so that the image data can be maximized. Simplify, and make feature extraction easy.

Preprocessing can use spatial domain processing or frequency domain processing.

Spatial domain processing includes methods such as color image conversion to grayscale image, histogram equalization, morphological transformation, and edge detection.

Frequency domain processing includes Gabor transform.

One or more of the above-mentioned methods can be used for preprocessing the digital image of rice paper.

In the method for detecting calligraphy and painting works based on texture features in this embodiment, histogram equalization and Gabor transformation are used for preprocessing.

The color image to grayscale image is used to convert the color rice paper digital image of the authentic calligraphy and painting works obtained through an optical microscope into a grayscale image.

The International Commission on Illumination (CIE) chooses red (wavelength 700.00nm), green (wavelength 546.1nm), blue (wavelength 438.8nm) as the three most basic colors, and any color can be represented by these three This is the RGB color representation system.

Since the human eye is most sensitive to the green channel, the green channel contains the most texture information, and compared with the original image, the contrast of the green channel is also improved, and the noise is less than the original image. Therefore, the present invention separates the color in the color image of the optical microscope into the green channel, and uses the value of the green channel to represent the gray value of the gray image.

Histogram equalization is a commonly used processing method in the field of image processing, which is used to improve the image quality in which both the background and the foreground are too bright or too dark. The principle is to use the histogram of the image to adjust the overall contrast of the image. It can be understood as changing the grayscale histogram of the original image from some relatively aggregated grayscale intervals to a uniform distribution in the entire grayscale range, so that the number of pixels in a certain grayscale range is about the same.

The probability of a pixel with grayscale i appearing in the image is:

${\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; 0,1</span>}<span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; L</span>}$

Among them, n _i represents the number of occurrences of gray level i in the grayscale image, L represents the highest pixel gray level in the image, and n represents the number of all pixels in the image.

Figure 2 and Figure 3 are images before and after histogram equalization based on optical microscope images, and Figure 4 and Figure 5 are images before and after histogram equalization based on electron microscope images, from which it can be seen that whether based on optical microscope images or based on Electron microscope images and histogram equalization can make texture features more obvious.

Mathematical morphology is a subject thought based on set theory, which plays an important role in image processing, so it is also a powerful algorithm for geometric morphology analysis and description. Mathematical morphology mainly studies the morphological characteristics of images, analyzes and processes the basic characteristics and structures of images, that is, analyzes and processes the relationship between elements and elements in images, and image parts and parts. Morphological image processing generally uses image neighborhood operations, uses neighborhood structural elements, and performs certain logical operations in the neighborhood structural element domains of each pixel, and the result is the final image result. Erosion and dilation are the most basic and common morphological operations, mainly used to reduce image noise and noise.

Let F(x) be the structural element, for each pixel point x in M in the space, the erosion transformation can be defined as:

X={x:F(X)∈M}

Figure 6 is the image M to be processed (generally a binary image, aimed at black dots), and Figure 7 is the structural element F(x), if the center point of F(x) is the black dot in the lower right corner, then the corrosion is It is to compare the center point of F(x) with the points on M in turn, if the black points on F(x) are all within the range of M, then the point will be kept, and if it is not within the range, it will be removed; 8 is the result after corrosion, from which it can be seen that the result after corrosion is still within the original range of M, but contains fewer points than M, as if M has been corroded.

Let F(x) be the structural element, for each point x in M in the space, the expansion transformation can be defined as:

The step of opening operation is to first perform erosion operation on an original image, and then perform expansion operation. After taking an appropriate threshold, the image is thresholded and binarized to obtain a binary image. Then perform the open operation to get the result.

Fig. 9 is the image after image thresholding based on the optical microscope image shown in Fig. 3; Fig. 10 is the image after the opening operation based on the optical microscope image shown in Fig. 9; Fig. 11 is the image based on the electron microscope shown in Fig. 5 The image after image thresholding; FIG. 12 is the image after opening operation based on the electron microscope image shown in FIG. 11 .

Because the image itself has a lot of noise, it has a great impact on the opening operation, making the result not ideal.

Edge detection is an important method of image processing. Since the edge is the area where the color of a certain part of the image changes the fastest, it is mostly between the target and the target, the target and the background, and the area and the area, and it is an important basis for image processing.

The two important attributes of the edge of the image are direction and magnitude. The edge direction is the most severe direction perpendicular to the edge pixels, and the edge magnitude is the degree of grayscale pixel transformation. Generally speaking, the pixels in the horizontal direction of the edge change gently, while the pixels in the direction perpendicular to the edge change sharply. This change on the edge can be detected using a differential operator, typically using first or second derivatives to detect edges.

Edge detection includes Laplacian edge detection and canny edge detection.

Laplace edge detection is a two-dimensional function f(x,y) whose Laplace transform is a second-order differential, defined as:

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{{<span\; class="notranslate">\; \&PartialD;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; f</span>}}{{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; +</span>\frac{{<span\; class="notranslate">\; \&PartialD;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; f</span>}}{{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

in:

$\frac{{<span\; class="notranslate">\; \&PartialD;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; f</span>}}{{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>$

$\frac{{<span\; class="notranslate">\; \&PartialD;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; f</span>}}{{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>$

In this way, for a 3*3 window, we can get:

f(x,y)=f(x+1,y)+f(x-1,y)+f(x,y+1)+f(x,y-1)-4f(x,y)

The convolution operation can be performed using the above formula.

Fig. 13 is the image after carrying out Laplacian edge detection based on optical microscope image shown in Fig. 9; Fig. 14 is the image after carrying out Laplacian edge detection based on electron microscope image shown in Fig. 11; From the figure It can be seen that the processing result of the image based on the electron microscope is better than that based on the optical microscope, but there are still many noises, and the effect is still not ideal.

Canny edge detection includes the following steps:

a) Denoising

Use two-dimensional Gaussian filter G(x,y) to smooth the image and reduce the influence of noise on the image. It can be equivalent to the product of filtering results of two one-dimensional Gaussian filters in the vertical direction G(y) and the horizontal direction G(x). The smoothed image is:

H(x,y)=G(x,y)*I(x,y)=G(y)*(G(x)*I(x,y)

Among them, I(x,y) represents the point (x,y) on the original image, and H(x,y) represents the point (x,y) on the filtered image.

b) Find the brightness gradient in the image

The vertical derivative and the horizontal derivative G(x) of G(y) are calculated by the finite difference of the first-order partial derivative, generally Sobel or Prewitt operator can be used. Then, use the directional derivative to calculate the gradient magnitude

$<span\; class="notranslate">\; \&dtri;</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; =</span>\sqrt{{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

c) Non-maximum suppression

The non-maximum value suppression operation is performed on the gradient value of each pixel to accurately position each edge point. For example, for the 3×3 neighborhood of the current pixel point, if the gradient magnitude of the point is greater than the gradient magnitude of two adjacent pixel points along the gradient direction, the point is considered to be an edge point to be selected, and its corresponding The point marked as 1. Otherwise, the point is considered as a non-edge point and marked as 0.

d) Thresholding

Double-thresholding processing is performed on the image after non-maximum suppression processing to remove false edges and connect broken edges. The final high threshold is calculated by the artificially given high threshold and the gray histogram of the image, and then the final low threshold is calculated by the given low threshold and the gray histogram of the image. Then compare each candidate edge point of the image after non-maximum suppression with the final high threshold and record the edge points smaller than the final high threshold. For all edge points, find a point larger than the final low threshold in the 8 neighborhood, and mark it as the final edge point.

Fig. 15 is the image after the canny edge detection based on the optical microscope image shown in Fig. 9; Fig. 16 is the image after the canny edge detection is carried out based on the electron microscope image shown in Fig. 11; it can be seen that the image based on the electron microscope The processing result is better than that based on the optical microscope, and the edges are continuous and clear. We can see the texture features intuitively with the naked eye. Secondly, there is no big difference between the processed image and the original image.

The Gabor transform in frequency domain processing is a windowed Fourier transform.

The Fourier transform is to decompose the signal into different frequency components, and any signal can be decomposed into the sum of complex sinusoidal signals. The Gabor function can extract the required relevant features in different scales and directions, and the expression of the human visual system in terms of frequency and direction is very close to the expression of the frequency and direction of the Gabor filter, so there are usually differences in texture processing and recognition. better effect.

Two-dimensional Gabor functions have two representations.

The first form is:

Original formula:

Real:

Imaginary part:

in:

x'=xcosθ+ysinθ

y'=-xsinθ+ycosθ

Indicates the phase shift, Indicates the wavelength of the sine function, θ indicates the direction of the kernel function, σ indicates the Gaussian standard deviation, and γ indicates the aspect ratio of the x and y directions.

The second form is:

${\mathrm{<span\; class="notranslate">\; g</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; v</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; e</span>}}^{<span\; class="notranslate">\; -</span>\frac{{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; 2</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}}<span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; e</span>}}^{\mathrm{<span\; class="notranslate">\; ik</span>}<span\; class="notranslate">\; \&CenterDot;</span>\left(\begin{array}{c}\mathrm{<span\; class="notranslate">\; x</span>}\\ \mathrm{<span\; class="notranslate">\; the\; y</span>}\end{array}\right)}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; e</span>}}^{<span\; class="notranslate">\; -</span>\frac{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; )</span>$

in:

${\mathrm{<span\; class="notranslate">\; k</span>}}_{\mathrm{<span\; class="notranslate">\; v</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; 2</span>}}^{<span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}\mathrm{<span\; class="notranslate">\; \&pi;</span>}$

The value of v indicates the wavelength of the Gabor filter, the value of u indicates the direction of the Gabor kernel function, K indicates the total number of directions, and the parameter σ/k indicates the size of the Gaussian window, which is taken here

The form used in this paper is the second one. The processed image can be obtained by performing convolution operation on the filter function and the image.

Since the Gabor function has a strong directionality, we select the Gabor transformation result graphs in six directions of 0°, 30°, 60°, 90°, 120° and 150°, as shown in Figure 17 and Figure 18, and Figure 17 9 is an image after performing Gabor filtering based on the optical microscope image shown in FIG. 9 , and FIG. 18 is an image after performing Gabor filtering based on the electron microscope image shown in FIG. 11 .

The result image of Gabor filtering is thresholded to obtain a binary image. In order to filter some image impurities, when merging image results in six directions, define the following filter conditions:

Let f _i (x, y) be the gray value of the Gabro transformed image at point (x, y) with the direction of the kernel function being 30*i, and this value can only be 0 or 255.

F(x, y) is the gray value of the merged image at point (x, y). but

if Then F(x,y)=0, that is, the point (x,y) is black.

if Then F(x,y)=255, that is, the point (x,y) is white.

It can be seen from Figure 17 and Figure 18 that the gabor filter processing has a good effect on both the electron microscope and the optical microscope, but it will produce some pseudo textures and noises, especially for images based on the optical microscope.

The implementation code of gabor filtering is as follows:

The earliest Haar feature was proposed by Papageorgiou C. et al. ("A general framework for object detection"), and later Paul Viola and Michal Jones proposed a method for quickly calculating Haar features using the integral image method ("Rapid object detection using a boosted cascade of simple features") . Later, Rainer Lienhart and Jochen Maydt extended the Haar feature library with diagonal features ("An extended set of Haar-like features for rapid object detection").

As shown in Figure 19, Haar features include edge features, linear features, center features, and diagonal features, which are combined into feature templates. There are two kinds of rectangles, white and black, in the feature template, and the feature value of the template is defined as the sum of the pixels of the white rectangle minus the sum of the pixels of the black rectangle.

In this embodiment, the first two edge features are selected as haarx and haary values for texture feature extraction.

Other Haar features or combinations of Haar features can also be selected for texture feature extraction.

Since the Haar feature is the sum of all pixel values in the black area in the rectangle minus the sum of all pixel values in the white area. The number of features in an image is far greater than the number of pixels. If the pixel sum of each feature is calculated, the amount of calculation will be very large, and many operations are repeated.

Step (3) utilizes the SURF algorithm to extract texture features from the preprocessed rice paper digital image, comprising the following steps:

a) Calculate the Haar feature using the integral image

Paul Viola proposed a method to quickly calculate Haar features using the integral image method, which is to construct an "integral image" (Integral image), also called Summed Area Table, and then any Haar rectangular feature can be obtained through the table lookup method ( Look Up Table) and a limited number of simple operations, greatly reducing the number of operations.

The integral image refers to the sum of all the gray levels in the rectangular area formed by the position of the current pixel point and the origin. The origin here refers to the pixel point in the upper left corner of the image.

The integral image of image I(x,y)(0≤x≤M, 0≤y≤N) can be expressed as:

${\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; total</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>$

Using the image method, it can be understood that the black part is the current pixel point, and the gray part is the integral area, as shown in Figure 20.

To calculate the sum S(x) of the gray levels of any rectangular area in the image, you only need to use the integral values of the four vertices of the rectangle to obtain:

S(x)＝S(X ₁ ,Y ₁ )+S(X ₂ ,Y ₂ )-S(X ₂ ,Y ₁ )-S(X ₁ ,Y ₂ )

Among them, S(X ₁ ,Y ₁ ), S(X ₂ ,Y ₂ ), S(X ₂ ,Y ₁ ) and S(X ₁ ,Y ₂ ) are the integral values of the four vertices of the rectangle respectively.

b) Construct a Hessian matrix to obtain extreme points

The Hessian matrix is the most important part of the entire Surf algorithm. Assuming that the function f(x, y) is a binary continuous differentiable function, the Hessian matrix is:

$\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left[\begin{array}{cc}\frac{{<span\; class="notranslate">\; \&PartialD;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; f</span>}}{{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}& \frac{{<span\; class="notranslate">\; \&PartialD;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; f</span>}}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}}\\ \frac{{<span\; class="notranslate">\; \&PartialD;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; f</span>}}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}}& \frac{{<span\; class="notranslate">\; \&PartialD;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; f</span>}}{{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{\mathrm{<span\; class="notranslate">\; 1</span>}}}\end{array}\right]$

Then each pixel will have a Hessian matrix, and the discriminant for the matrix is:

$\mathrm{<span\; class="notranslate">\; det</span>}\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; =</span>\frac{{<span\; class="notranslate">\; \&PartialD;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; f</span>}}{{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}\frac{{<span\; class="notranslate">\; \&PartialD;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; f</span>}}{{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span>\frac{{<span\; class="notranslate">\; \&PartialD;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; f</span>}}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}}\frac{{<span\; class="notranslate">\; \&PartialD;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; f</span>}}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}}$

If detH>0, the function f(x, y) obtains the extremum point at (x, y).

c) Gaussian filtering is used to find out whether the point (x, y) is the basis for judging whether the point is a local extremum point

Since the feature points need to be scale-independent, before constructing the Hessian matrix, it is necessary to perform Gaussian filtering on the image to form representations at different scales. Then the image function given Gaussian filter is:

L((x,y),t)=G(t)·I(x,y)

L((x,y), t) is the representation of the image at different scales, and the Gaussian kernel G(t) is convolved with the image function I(x,y) at the point (x,y), where Gaussian The kernel G(t) is:

$\mathrm{<span\; class="notranslate">\; G</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{{<span\; class="notranslate">\; \&PartialD;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; g</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&PartialD;</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

Where g(x) represents the Gaussian function and t represents the Gaussian variance. Using this method, it is possible to calculate the decision value of the H determinant of each pixel in the image, and then use the obtained value to distinguish the feature point. Herbert Bay later proposed that an approximate value can be used instead of L(x, t) to facilitate the application. In order to reduce the error of the correct value and the approximate value, the H matrix discriminant is expressed as:

$\mathrm{<span\; class="notranslate">\; det</span>}\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; =</span>\frac{{<span\; class="notranslate">\; \&PartialD;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; f</span>}}{{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}\frac{{<span\; class="notranslate">\; \&PartialD;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{<span\; class="notranslate">\; \&PartialD;</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 0.9</span>}<span\; class="notranslate">\; \&times;</span>{<span\; class="notranslate">\; (</span>\frac{{<span\; class="notranslate">\; \&PartialD;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; f</span>}}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

Figure 21 is the image obtained after Gaussian filtering. The second derivative template is calculated in the y-axis direction, and approximate processing is used to speed up the operation. The image after processing is shown in the right image, which is much simplified. And the right image can be calculated using the integral image, which speeds up the speed even more. Similarly, the second-order mixed partial guide template in the x and y directions is shown in Figure 22:

In this way, the basis for judging whether the point (x, y) is a local extremum point is found out. The convolution process of different dimensions can be represented by a pyramid model. Each layer of the pyramid represents a scale of the image, and the convolution of the Gaussian kernel and the original image can represent representations at different scales. The Surf algorithm keeps the original image unchanged for each layer and only changes the size of the filter.

d) Extract feature points

Figure 23 and Figure 24 show the scale pyramid. After obtaining the local extreme points of the image at different scales, as shown in Figure 25, it is adjacent to 26 points in the three-dimensional field of the pyramid, that is, in the same scale space Compare the 8 points with the 18 points of the upper and lower two adjacent scales, and if the point is an extreme point, it will be reserved as a preliminary feature point.

After obtaining all the preliminary feature points, use the three-dimensional linear difference method to obtain sub-pixel-level feature points, and then remove the points smaller than the set threshold to obtain the final feature points.

e) Get feature direction

In SURF matching, the feature vector is obtained by calculating the wavelet feature of the feature point field. If S is the scale of the feature point, generally the feature point is the center of the circle, 6S is the radius, and the circle area is established to calculate the harr wavelet characteristics of these points.

In the SURF algorithm, as shown in Figure 26a, Figure 26b, and Figure 26c, the entire circle area is scanned with a 60-degree fan-shaped area and a set interval. The maximum value of the harr wavelet response is 4s, and the weight is given according to the distance from the center of the circle. The distance The farther the center of the circle is, the smaller the weight is, and then add the Harr wavelet response vectors of each point in the fan-shaped area. In this way, each sector will get a vector value, and finally the sector area with the largest vector value is found, and the vector direction of the sector area is used as the feature direction of the feature point. Repeat the above method to get the feature directions of all feature points.

f) Construct SURF feature point description operator

Take a square area of 20s*20s around the feature point, and divide this area into 16 sub-areas of 5*5 according to the direction of the feature point, as shown in Figure 27, use the haar feature to calculate each The direction of the feature point.

The implementation code to calculate the eigenvector is as follows:

SiftFeatureDetector detector; //The constructor uses the internal default

std::vector<KeyPoint>keypoints_1, keypoints_2;//Construct 2 point vectors specially composed of points to store feature points

detector.detect(img_1,keypoints_1);//Store the feature points detected in the img_1 image and put them in keypoints_1

detector.detect(img_2,keypoints_2);//similar

// draw feature points in the image

Mat img_keypoints_1, img_keypoints_2;

drawKeypoints(img_1,keypoints_1,img_keypoints_1,Scalar::all(-1),DrawMatchesFlags::DEFAULT);//Draw feature points in memory

drawKeypoints(img_2, keypoints_2, img_keypoints_2, Scalar::all(-1), DrawMatchesFlags::DEFAULT);

imshow("sift_keypoints_1",img_keypoints_1);//display feature points

imshow("sift_keypoints_2", img_keypoints_2);

// Compute eigenvectors

SiftDescriptorExtractor extractor;//Define the description sub-object

Mat descriptors_1,descriptors_2;//Matrix to store eigenvectors

extractor.compute(img_1,keypoints_1,descriptors_1);//calculate feature vector

extractor.compute(img_2, keypoints_2, descriptors_2);

Step (4) performing texture feature matching on the calligraphy and painting works to be detected,

There are many methods for feature point matching, such as K Nearest Neighbors (KNN), Brute Force (BF) and Fast Approximate Nearest Neighbors (Flann), etc., because the Flann algorithm is suitable for High-latitude data is not more efficient than the two algorithms mentioned above, so this paper chooses the Flann algorithm.

However, the Flann matching algorithm has directionality between the target image and the candidate image, that is, matching the candidate image with the target image is not the same as matching the target image with the candidate image. Therefore, this paper uses two-way Flann matching to reduce this error.

Suppose a feature point a ₁ in the image M ₁ , find the initial pairing point (a ₁ , a ₂ ) with the minimum distance in the image M ₂ through the Flann algorithm, and then set the threshold according to the minimum distance dist _min among all matching pairs DIST = h*dist _min . If the minimum distance of the matching point is less than dist _min , a ₂ in the image M ₂ is used as a ₁ candidate matching point of the image M ₁ , otherwise, a ₁ is eliminated, and the next point of the image M ₁ is matched, and finally the image is obtained Matching point pairs of M ₁ for M ₂ .

Similarly, according to the above method, the matching point pairs of image M ₂ to M ₁ are obtained. Suppose a matching point pair of image M ₁ to M ₂ is (a ₁ , a ₂ ), and a matching point pair of point a ₂ in image M ₂ to M ₂ is (a ₃ , a ₂ ), if a ₃ =a ₁ , the match is considered successful, otherwise it is eliminated. In this way, the final matching point is obtained. Figure 28 and Figure 29 are feature point matching results. The straight line in Figure 28 represents the connection of two matching points, because the first group of rice paper images has only translation and no rotation, so the matches seen are all parallel straight lines. The second group is rotated, and the matching straight lines in Figure 29 will not be parallel.

Although two-way Flann matching greatly improves the matching accuracy of feature points, we cannot ensure that the feature points are matched correctly, and the target image itself may contain meaningless feature points, that is, noise points, which are not of interest to us. A valid description of the target.

Based on the above problems, it can be considered that the matching of feature points alone cannot ensure an accurate output result for the matching of calligraphy and painting images. Here we consider to use feature point clustering to extract the same area to solve this problem.

In the clustering algorithm, K-Means is a very representative and most commonly used algorithm. It divides n objects into k clusters, so that the sum of the similarities of objects in all clusters is the highest. Here, the similarity It is represented by the mean distance from the points in the cluster to the center. The algorithm starts to randomly select K objects, assuming that each object is the center of a cluster, judge the distance between all objects and k centers, add them to the nearest cluster, and then update the mean and center of each cluster, repeat the above Step until the sum of the mean square errors in each cluster is minimized.

However, the original K-Means algorithm can achieve a better clustering effect only when the differences between clusters are large, and it is sensitive to outliers and is easily disturbed. Therefore, this embodiment adopts the weighted K-means method (K-WMeans). The whole calculation process is as follows:

① Select any K points as the original cluster center. For a certain cluster S _j , if x _i is its center, define d( _xi , x _k ) as the distance from x _i to any other point x _k , then you can Define the weight value of a point as:

${\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{<span\; class="notranslate">\; \&prime;</span>}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{<span\; class="notranslate">\; \&prime;</span>}}$

in $w_k^{\&prime;}=\frac{1}{\mathrm{no}}{\mathrm{\&Sigma;}}_{i=1}^{\mathrm{no}}d(x_i,x_k).$

② For the points in the cluster, calculate its weighted average, and re-assign these points to the cluster with the closest weight. For a certain cluster S _j , the weight calculation formula is:

$\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; m</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}$

③ Calculate the weight of the heart cluster according to the appeal formula

④ Repeat steps 2 and 3 until each brittle sample is stable and the sum of mean square errors within each cluster is minimized.

In the method for detecting calligraphy and painting works based on texture features in this embodiment, the feature points are clustered into four categories, and the clustering effect is shown in FIG. 30 .

After clustering, the center point of each cluster is extracted, and a square area of fixed size is extracted with the point as the center. The area of the target image is smaller than the candidate image. This is because there is an offset error in the cluster center caused by the feature point matching error. The effect of clustering four categories and intercepting part of the area with the cluster center is shown in Figure 31.

At this point, we combine the target area extracted from the candidate image with the center point after clustering to perform affine transformation, and correct the errors of rotation and translation. Then the candidate image is expanded, which is also to reduce the error. Let the image size of the target area be M*M, the image size of the candidate area be N*N, N>M, take the point (i, j) as the center point field in the candidate area, and intercept the same square area as the target area as H _{i, j} , where 0≤i≤NM, 0≤j≤NM, then the similarity S(i,j) between H _i,j and the target area is:

Among them, L _M (n, m) and H _{i, j} (n, m) represent the gray value of the point (n, m) of the target area and the intercepted area, respectively. We define the similarity rate S of the region as:

$\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; Max</span>}<span\; class="notranslate">\; (</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; J</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>$

Finally, we select the highest similarity among all regions as the similarity of the two images.

The similarity of the first group of calligraphy and painting works is:

match rate 1

Match rate 0.995885

Matching rate 0.970803

match rate 1

The similarity of the second group of calligraphy and painting works is:

Matching rate 0.728111

Match rate 0.928571

Matching rate 0.871508

Match rate 0.769504

The first set of data is the similarity of the region extracted from each class in the first set of images, where the maximum matching rate is 1, then it can be considered that the two paintings and calligraphy are the same, then this painting and calligraphy is authentic. The second set of data is the similarity of the regions extracted from each class in the second set of images. Due to the existence of rotation and preprocessing errors, the matching rate of not every region is not particularly high, but there is still a region The matching rate reached 92.8%, so it can also be considered that this painting and calligraphy is authentic.

The implementation code for region matching is as follows:

IplImage*pRightImage11 = cvCloneImage(my_rio1);

cvSetImageROI(pRightImage11,Rect(maxx,maxy,rect_wide,rect_wide));

cvCopy(pRightImage11, matched, 0);

IplImage*matchmodel = andoperate(matched, my_rio);

cout<<"matching rate"<<maxvalue<<endl;

The method for detecting calligraphy and painting works based on texture features provided by the present invention, (1) only needs to perform feature extraction once on the authentic calligraphy and painting works, and in the detection process of the calligraphy and painting works to be detected, the features extracted from the authentic calligraphy and painting works are used. Combined with the FLANN algorithm, the feature points are clustered to obtain the center of each category, and the square area is intercepted with the center as the range, and the authenticity of the entire calligraphy and painting is judged by the highest similarity of the square area, which can quickly and accurately identify the authenticity of calligraphy and painting works False; there is no need to artificially add materials other than calligraphy and painting to assist in judging the authenticity of calligraphy and painting works. Since the rice paper texture characteristics of authentic calligraphy and painting works can hardly be copied, it can prevent criminals from using high technology to forge calligraphy and painting works.

The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make many modifications and changes according to the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art shall be within the scope of protection defined by the claims.

Claims (7)

1. A method for detecting calligraphy and painting works based on texture features, comprising the following steps: (1) Obtain digital images of Xuan paper of authentic works of calligraphy and painting; (2) carry out preprocessing to the rice paper digital image, obtain the rice paper digital image after preprocessing; (3) Utilize the SURF algorithm to extract texture features from the preprocessed rice paper digital image; (4) Match the texture features of the painting and calligraphy works to be detected. If they match, the painting and calligraphy works to be detected are authentic works of calligraphy and painting; if they do not match, the painting and calligraphy works to be detected are fakes. 2. the painting and calligraphy work detection method based on texture feature as claimed in claim 1, is characterized in that, adopts bidirectional Flann matching algorithm to carry out texture feature matching to the painting and calligraphy work to be detected in step (4). 3. the painting and calligraphy work detection method based on texture feature as claimed in claim 2, is characterized in that, also comprises the following steps before carrying out texture feature matching in step (4): (41) utilize improved K-Means algorithm to carry out clustering, obtain the center of each class; (42) By the center of the class obtained in the step (41), a square area is intercepted with the center of the class as the center, and the authenticity of the painting and calligraphy works to be detected is judged with the highest similarity of the square area. 4. the painting and calligraphy work detection method based on texture feature as claimed in claim 1, it is characterized in that, step (2) carries out preprocessing to rice paper digital image and adopts color picture to turn gray scale picture, histogram equalization, morphological transformation, edge detection Or one or a combination of Gabor filtering. 5. the painting and calligraphy work detection method based on texture feature as claimed in claim 1, is characterized in that, step (3) utilizes SURF algorithm, extracts texture feature from the rice paper digital image after preprocessing, comprises the following steps: (31) Utilize integral image to calculate Haar feature; (32) Construct a Hessian matrix to obtain extreme points; (33) Gaussian filtering is used to find out the basis for judging whether the point (x, y) is a local extremum point; (34) feature points are extracted; (35) Acquiring the characteristic direction of the characteristic point; (36) Construct a SURF feature point description operator. 6. the painting and calligraphy work detection method based on texture feature as claimed in claim 5, is characterized in that, step (35) obtains feature direction and comprises the following steps: (351) Take the feature point as the center of the circle, and take six times the scale of the feature point as the radius to establish a circle area; (352) scanning the circular area with a 60-degree fan-shaped area and a set interval; (353) The maximum value of the harr wavelet response is four times the scale of the feature point, and the weight is given according to the distance from the center of the circle. The farther away from the center of the circle, the smaller the weight; (354) adding the Harr wavelet response vectors of each point in the fan-shaped area; (355) Find the Harr wavelet response vector and the largest fan-shaped area, and the vector direction of the fan-shaped area is the characteristic direction of the feature point. 7. the painting and calligraphy work detection method based on texture feature as claimed in claim 5, is characterized in that, the method for step (34) extraction feature point comprises the following steps: (341) Compare the local extremum point with 8 adjacent points in the same scale and 18 points in the upper and lower two adjacent scales. If the local extremum point is an extremum point, then the local extremum point is preliminary Extreme point; (342) After obtaining all the preliminary feature points, use the three-dimensional linear difference method to obtain sub-pixel-level feature points, and then remove the points smaller than the set threshold to obtain the final feature points.