RU2538319C1 - Device of searching image duplicates - Google Patents

Abstract

FIELD: information technology.

SUBSTANCE: device comprises the pretreatment units of first and second images, the recording units of first and second images, the conversion units of first and second images into a color space YIQ, the enhancing units of the real component of first and second images, the image forming units as a result of rotation of the first and second image, the units of image forming in changing the angle of inclination of the first and second images, the units of storage of simulated images for the first and second images, the unit of application of the method SIFT, the calculation unit of quantity of equal descriptors, the unit of storage of the found pair of duplicates.

EFFECT: ensuring the ability to compare the descriptors applied to the task of searching image duplicates.

5 dwg

Description

The invention relates to methods for processing digital images and can be used in computer vision systems for identification and registration of objects in the image, multimedia applications that work with visual data.

- столбцы и строки изображения, $$k=\overline{1,3}$$

- цветовой канал. При использовании изображения в градациях серого, каждая точка представлена уровнем яркости в диапазоне от 0 (черный) до 255 (белый), с промежуточными значениями, представляющими различные уровни серого. Соответственно такое изображение обозначим как S_i,j.A simplified mathematical image model is a model in the RGB color space in the form of an array S _{i, j, k} , where $$\overline{i=\mathrm{one},I},j=\overline{\mathrm{one},J}$$

- columns and rows of the image, $$k=\overline{\mathrm{one},3}$$

- color channel. When using grayscale images, each dot is represented by a brightness level ranging from 0 (black) to 255 (white), with intermediate values representing different gray levels. Accordingly, we denote such an image as S _{i, j} .

The main task to be solved is matching the descriptors as applied to the problem of finding duplicate images.

The search for point correspondences between images of the same object is of great importance for scene identification, reconstruction of three-dimensional space objects, image search, and object recognition.

Image matching methods are based on the construction of descriptors. A descriptor is a feature vector that is calculated for each point feature of the image and describes the structure of its surroundings. As a rule, these vectors are formed on the basis of a set of values of the first and second derivatives of the image at a point.

Descriptors can be represented in the form of small invariant sections, image points, special areas. When choosing a method for representing regions, the following alternative arises: a region can be represented by its external characteristics (i.e., its boundary) or internal characteristics (a set of elements that make up this region). The external representation is usually chosen in those cases when the main attention is paid to the characteristics of the shape of the region. An internal representation is selected if the properties of the area itself, such as color or texture, are of interest. Sometimes both methods of presentation are used simultaneously. Data presented in the form of a plurality of pixels located along a boundary or within an area is used to obtain descriptors that take values in a feature space. If you set a measure on such a space, then you can compare the images with each other, calculating the distance between the corresponding feature vectors.

In order to satisfy the task of comparing images and their point features, the descriptors must be specific and selected so that with high probability the descriptors of different images (but belonging to the same point features) are correctly matched.

Descriptors must be:

- specific (distinguish different points);

- local (depend only on a small neighborhood);

- are invariant (rotation, tension, compression, monotonic change in brightness, affine and projective transformations);

- easy to calculate.

A common problem with descriptor matching methods is to recognize an object in photographs taken from different angles, at different viewing angles, at different scales, under different lighting conditions. This leads to the fact that the same object, depending on the shooting conditions, will be characterized by various feature vectors. Therefore, recognition methods must be invariant with respect to such changes.

Analysis of existing literary sources makes it possible to single out the most popular methods. The main problem of all methods of matching descriptors in images is finding affine invariant local features. Almost none of the methods considered satisfies this requirement. The Hessian-Affine and Harris-Affine methods are not resistant to changes in viewing angle and scale; therefore, they do not adequately implement affine invariance. The MSER method is not completely scale invariant: it cannot cope with sharp changes in the level of geometric blur. The number of different variants of the SIFT method, such as PCA-SIFT, GLOH, SURF, is constantly expanding. They are widely used in the field of recognition, detection, motion registration, but do not solve the problem of affine invariance.

A known method of recognizing objects in an image [Patent RU 2438174 C1, IPC G06K 9/68]. The invention relates to methods for recognizing objects in machine vision systems, television surveillance systems, information and control systems of robotic complexes.

The main technical task is to create a method that allows to increase the recognition accuracy by increasing the stability of the detectors of key areas in the image and increasing the number of invariant characteristics of these detectors. At the first stage, the input image is minimized with the given Green function. Next, the resulting convolutions are subtracted from each other to obtain a finite-difference approximation of the first derivative of the convolution of the input image with the filter; when searching for the local extremum of this convolution, the corresponding first derivatives are equal to zero. At the next stage, all local extrema are found and adaptive threshold filtering is performed to cut off minor features, while the selected points serve as the centers of the neighborhoods for which arbitrary descriptors are constructed.

The signs of the method-analogue, coinciding with the signs of the proposed technical solution, are as follows:

- construction of image descriptors.

The disadvantages of the known method and device that implements it are:

- low accuracy of determining local descriptors. The reasons that impede the achievement of the required technical result are as follows:

- the use of the convolution method with the Green function does not allow to detect various types of textures.

A known method and device for detecting an object in the image [US Patent No. 6711293, IPC G06K 9/68]. The invention relates to methods for recognizing objects in machine vision systems.

The main technical task is to recognize objects in the image using an invariant scale function. At the first stage, image differences are calculated: the convolution of the image with the Gaussian function is performed, then the convolution of the minimized image with the Gaussian function is again calculated to construct the difference image, and the difference image is subtracted from the input image. At the second stage, local extremes of pixel values are found. At the third stage, areas around the points of extremum are distinguished.

At the fourth stage, regions are divided into subdomains and at the fifth stage, many components are produced - descriptors of subdomains.

The signs of the method-analogue, coinciding with the signs of the proposed technical solution, are as follows:

- finding and matching descriptors;

- invariance of descriptors to zooming.

The disadvantages of the known method and device that implements it are:

- low accuracy of constructing descriptors.

The reasons that impede the achievement of the required technical result are as follows:

- The Gaussian function has a limited set of invariant characteristics characterizing the image features described by feature vectors, and, therefore, these vectors are less informative.

A known method of computer recognition of objects [Patent RU No. 2234127 C2, IPC G06K 9/68]. The invention relates to the field of automation and computer technology, namely to artificial intelligence systems.

The main technical problem is to increase the recognition rate of objects entered into the computer. At the first stage, the image of the object introduced into the computer is preliminarily brought to a normal, standard form for this method - zooming, rotation to the desired position, centering, inscribing the rectangle of the required size, converting the image of the object into an image made in gradations - of various degrees brightness - one color. Then, images sequentially stored in the computer memory are sequentially and alternately superimposed on it. The object recognition program step-by-step combines normalized images of recognizable objects centered and inscribed in the same cell size of the table and patterns centered and inscribed in the same cells of the template table with a step equal to the height of the row with cells or the column width of the table cells. At the same time, in each of the columns or in each row of the pattern table, the number of which is equal to the number of columns or rows in the table of recognized objects, there is a complete set of patterns

The signs of the method-analogue, coinciding with the signs of the proposed technical solution, are as follows:

- matching descriptors of two images.

The disadvantages of the known method and device that implements it are:

- lack of invariance to affine transformations;

- the need to ensure centering and normalization of images;

- the need for a complete enumeration of patterns in recognition.

The reasons that impede the achievement of the required technical result are as follows:

- when combining images step by step, even small differences in the structure of images lead to a sharp increase in recognition error.

A known method and device for recognizing images of objects (Patent RU No. 2361273 C2, IPC G06K 9/62). The invention relates to computer technology and can be used in vision systems to identify objects in the image.

The technical task is to increase the accuracy and quality of recognition through the use of a three-dimensional vector model of the standard of the object. The technical result is achieved as follows: the reference image is stored in the form of a vector three-dimensional model; for each such model, a set of parameters for affine transformations is fixed: rotation angles along the x, y, z axes and scale. This set of parameters is determined for each model, taking into account the complexity of its shape: the more complex the shape, the greater the number of angles necessary for the most complete description of possible options for the position of the object in space in order to most accurately identify.

The following sequence of actions is performed: a vector three-dimensional model of the reference object is obtained by geometric construction, then, changing its position in space (rotation, reflection, scaling), a number of the above parameters are obtained, which are saved and used in the future for recognition to recreate the corresponding angle of the object’s standard. Upon recognition, the three-dimensional image is rotated, each time a series of flat images are generated until a match can be found.

The signs of the method-analogue, coinciding with the signs of the proposed technical solution, are as follows:

- modeling of affine transformations of an object.

The disadvantages of the known method and device that implements it are:

- the presence of a preliminary processing stage;

- determination of a number of parameters of the object, the class to which the given object belongs, overall dimensions. This set of parameters is determined for each model, taking into account the complexity of its shape.

The reasons that impede the achievement of the required technical result are as follows:

- high computing costs.

Closest to the claimed solution, chosen by us for the prototype, is a method and device for affine-invariant pattern recognition [Patent No. US 2011/0069889 A1, IPC G06K 9/46]. The invention relates to the recognition of digital image objects.

The main technical requirement is the recognition of an object that is invariant to rotation, displacement, and scale. This requirement is achieved by adding an affine invariant extension to the SIFT (Scale-invariant feature transform) method. ASIFT (Affine-scale-invariant feature transform) allows you to reliably identify functions that have passed strong affine distortion. Instead of constructing affine-invariant descriptors, it models, with sufficient accuracy, all distortions caused by a change in the position of the optical cameras. Scale and change of camera position (includes two parameters of the camera axis position) - simulated parameters; rotation and movement (also includes 2 parameters of change of position) - normalized parameters. ASIFT first achieves the affinity invariance of image descriptors, then uses SIFT which stimulates scale and normalizes rotation and movement.

Consider a prototype device involves the following operations:

1. Each image is converted by simulating all possible affine distortions caused by a change in camera position from the front position. These distortions depend on two parameters: longitude φ and latitude θ. As a result, at the first step of this algorithm, we obtain a certain set of images that differ in all kinds of affine transformations;

2. Images obtained after affine transformations are processed by the SIFT algorithm;

3. SIFT includes the following operations:

- Finding singular points by constructing a pyramid of Gaussian (Gaussian) and differences of Gaussian (Difference of Gaussian, DoG).

- Refinement of singular points by approximating the DoG function (Gaussian difference) by a second-order Taylor polynomial taken at the point of the calculated extremum.

- Finding the orientation of the key point, which is calculated based on the directions of the gradients of adjacent points.

- Building descriptors. In the SIFT method, the descriptor is a vector. Like the direction of the cue point, the handle is calculated on the Gaussian closest in scale to the cue point, and based on the gradients in a certain cue point window.

The disadvantages of the known prototype device are:

- Different lighting conditions (e.g. day / night).

- The object has a reflective surface (usually cars, mirrors).

- The object has a strong 3-D structure.

- The object has similar descriptors or a periodic structure.

The reasons that impede the achievement of the required technical result are as follows:

- the ASIFT method examines the image array in grayscale S _{i, j} (the image array takes values in the range from 0 to 255), this method allows you to maintain resistance to color change. But with this approach, reflective surfaces retain their features, which significantly worsens the process of finding correspondence between descriptors.

The proposed device for searching for duplicate images allows you to solve one of the problems of the original device by using the YIQ color space. The device implements the following algorithm. At the first stage, we present an array of images S _{i, j, k} in the YIQ color space. This is the NTSC television standard format. This color space, as well as human vision, is more sensitive to light brightness (intensity), rather than color. The image consists of three components - brightness (Y) and two artificial color difference (I and Q) components. The signal I is called in-phase, Q - quadrature.

We decompose the YIQ channel into color components. The conversion from RGB to YIQ is carried out according to the following formulas:

, $$\text{Y}=\text{0.299}\cdot \text{R}+\text{0.587}\cdot \text{G}+\text{0.114}\cdot \text{B}$$

,

, $$\text{I}=\text{0.596}\cdot \text{R-0,274}\cdot \text{G-0,322}\cdot \text{Â}$$

,

. $$\text{Q}=\text{0.211}\cdot \text{R-0,522}\cdot \text{G}+\text{0.311}\cdot \text{Â}$$

.

An analysis of the images of the various components of the YIQ space showed that components I and Q do not contain the glare and reflection inherent in mirror surfaces. These components do not contain brightness information, but show only color, while the reflective surface structure is characterized to a greater extent by a change in brightness. We choose channel I to search for correspondence between descriptors on reflective surfaces.

Further, the obtained images are processed by the ASIFT algorithm.

Images are transformed by simulating all possible affine distortions caused by changing the position of the camera from the front position. These distortions are obtained by changing two parameters: φ longitude and latitude θ. As a result, at the first step of this method there is a certain set of images that differ in various affine transformations. Rotations and tilts are modeled for finite and small values of latitude and longitude of the angle, the choice of the step of these parameters is provided by a simulated image that takes any viewing opportunity at other values of latitude and longitude.

Then the step of finding the singular points is performed. The main point in the detection of singular points is the construction of a pyramid of Gaussians and differences of Gaussians. Gaussian (or, image blurred Gaussian filter) is the image:

, $$L(x,y,\sigma )=G(x,y,\sigma )*I(x,y)$$

,

where L is the Gaussian value at the point with coordinates (x, y), σ is the blur radius, G is the Gaussian core, I is the value of the original image, * is the convolution operator.

Gaussian difference is an image obtained by pixel-by-pixel subtraction of one Gaussian source image from an image with a different blur radius.

$$D(x,y,\sigma )=(G(x,y,k\sigma )-G(x,y,\sigma ))*I(x,y)=L(x,y,k\sigma )-L(x,y,\sigma ).$$

Scalable image space is a set of various versions of the original image smoothed by some filter. It is proved that the Gaussian space is linear, invariant with respect to shifts, rotations, scale, not displacing local extrema, and has the property of semigroups. A different degree of image blurring by a Gaussian filter can be taken as the original image taken at a certain scale.

Scale invariance is achieved by finding key points for the original image taken at different scales. To do this, a Gaussian pyramid is built: the entire scalable space is divided into some sections - octaves, and part of the scalable space occupied by the next octave is two times larger than the part occupied by the previous one. In addition, when moving from one octave to another, resampling of the image is done, its size is halved. Naturally, each octave spans an infinite number of Gaussian images, so only a certain number N of them is built, with a certain step along the blur radius. With the same step, two additional Gaussians (totaling N + 2), extending beyond the octave, are being completed. The scale of the first image of the next octave is equal to the scale of the image from the previous octave with the number N.

In parallel with the construction of the pyramid of Gaussians, a pyramid of differences of Gaussians is constructed, consisting of the differences of neighboring images in the pyramid of Gaussians (Fig. 1). Accordingly, the number of images in this pyramid will be N + 1.

Each difference is obtained from two neighboring Gaussians, the number of differences is one less than the number of Gaussians, when moving to the next octave, the image size is halved.

After building the pyramids we find the singular points. A point will be considered singular if it is a local extremum of the Gaussian difference. Figure 2 shows the step of determining the points of the extremum. It is considered such if the value of the difference of the Gaussians at the marked point is greater (less) than all other values at the points.

In each image from the pyramid of the difference of the Gaussians, local extremum points are determined. Each point of the current DoG image is compared with its eight neighbors and nine neighbors in the DoG, which are one level higher and lower in the pyramid. If this point is larger (less) than all neighbors, then it is taken as a point of local extremum.

The next step will be to refine the singular points - to check the suitability of the extremum point for the singular role.

The coordinates of the singular point are determined with subpixel accuracy. This is achieved by approximating the DoG function by the second-order Taylor polynomial taken at the point of the calculated extremum.

, $$D(x)=D+\frac{\partial D^T}{\partial x}x+\frac{\mathrm{one}}{2}{x}^T\frac{{\partial }^{2}D}{\partial x^2}x$$

,

where D is the DoG function, x = (x, y, σ) is the displacement vector relative to the decomposition point, the first DoG derivative is the gradient, the second DoG derivative is the Hessian matrix.

The extremum of the Taylor polynomial is found by calculating the derivative and equating it to zero. The result is a shift of the point of the calculated extremum, relatively accurate:

. $$\stackrel{\wedge }{x}=-\frac{{\partial }^2{D}^{-\mathrm{one}}}{\partial x^2}\frac{\partial D}{\partial x}$$

.

больше 0,5*шаг сетки в этом направлении, то это означает, что на самом деле точка экстремума была вычислена неверно и нужно сдвинуться к соседней точке в направлении указанных компонент. Для соседней точки все повторяется заново. Если таким образом мы вышли за пределы октавы, то следует исключить данную точку из рассмотрения.If one of the components of the vector $$\stackrel{\wedge }{x}$$

more than 0.5 * grid spacing in this direction, this means that in fact the extremum point was calculated incorrectly and you need to move to an adjacent point in the direction of the indicated components. For a neighboring point, everything repeats again. If in this way we went beyond the octave, then this point should be excluded from consideration.

When the position of the extremum point is calculated, the DoG value at this point is checked by the formula:

. $$D(\stackrel{\wedge }{x})=D+\frac{\mathrm{one}}{2}\frac{\partial D^T}{\partial x}x$$

.

If this test fails, then the point is excluded as a point with low contrast.

The last check includes checking if a particular point lies on the boundary of some object or is poorly lit, in which case such a point can be excluded from consideration. These points have a large bend (one of the components of the second derivative) along the boundary and small in the perpendicular direction. This large bend is determined by the Hessian matrix H. An H of size 2 × 2 is suitable for testing.

. $$H=\left[\begin{array}{cc}{D}_{xx}& D_{xy}\\ D_{xy}& D_{yy}\end{array}\right]$$

.

Let Tr (H) be the trace of the matrix, and Det (H) be its determinant.

, $$Tr(H)=D_{xx}+D_{yy}=\alpha +\beta$$

,

. $$Det(H)=D_{xx}{D}_{yy}-{(D_{xy})}^2=\alpha \beta$$

.

Let r be the ratio of the larger bend to the smaller,

, $$\alpha =r\beta$$

,

Then

, $$\frac{Tr{(H)}^2}{Det(H)}=\frac{{(\alpha +\beta )}^2}{\alpha \beta }=\frac{{(r\alpha +\beta )}^2}{\mathrm{<figure-callout\; id="2"\; label="r\; \beta "\; filenames="00000026.png"\; state="\{\{state\}\}">r\; </figure-callout>}{\beta }^{}{}^2}=\frac{{(r+\mathrm{one})}^2}{r}$$

,

and the point is considered further if

. $$\frac{Tr{(H)}^2}{Det(H)}<\frac{{(r+\mathrm{one})}^2}{r}$$

.

The orientation direction of the key point is calculated based on the direction of the gradients of the points adjacent to the singular. All calculations of the gradients are performed on the image in the pyramid of the Gaussians, with the scale closest to the scale of the key point. The magnitude and direction of the gradient at the point (x, y) are calculated by the formulas:

, $$m(x,y)=\sqrt{{(L(x+\mathrm{one},y)-L(x-\mathrm{one},y))}^2+{(L(x,y+\mathrm{one})-L(x,y-\mathrm{one}))}^2}$$

,

, $$\theta (x,y)={\tan }^{-\mathrm{one}}(\frac{L(x,y+\mathrm{one})-L(x,y-\mathrm{one})}{L(x+\mathrm{one},y)-L(x-\mathrm{one},y)}$$

,

- величина градиента, $$\theta$$

- его направление.Where $$m$$

is the magnitude of the gradient $$\theta$$

- his direction.

First, we define a window (neighborhood) of the key point at which the gradients will be considered. This will be the window required for convolution with the Gaussian core, it will be round and the blur radius for this core (σ) is 1.5 * the scale of the key point. For the Gaussian kernel, the so-called “three sigma” rule applies. It consists in the fact that the value of the Gaussian nucleus is very close to zero at a distance exceeding 3 * σ. Thus, the radius of the window is defined as [3 * σ].

The direction of the singular point is found from the histogram of directions (figure 3). The histogram consists of 36 components that evenly cover the gap of 360 degrees, and it is formed as follows: each point of the window (x, y) contributes m * G (x, y, σ) to that component of the histogram that covers the gap containing the direction of the gradient θ (x, y).

The direction of the key point lies in the gap covered by the maximum histogram component. The values of the maximum component (max) and two neighboring components are interpolated by a parabola, and the maximum point of this parabola is taken as the direction of the key point. If there are more components in the histogram with values of at least 0.8 * max, then they are likewise interpolated and additional directions are assigned to the key point.

The next step is the construction of descriptors. In the SIFT method, the descriptor is a vector. Like the direction of the cue point, the handle is calculated on the Gaussian closest in scale to the cue point, and based on the gradients in a certain cue point window. Before calculating the descriptor, this window is rotated by the angle of direction of the key point, thereby achieving invariance with respect to rotation.

Figure 3 schematically shows a part of the image and a descriptor derived from it. Here are the pixels indicated by small squares. These pixels are taken from the square window of the descriptor, which in turn is divided into four more equal parts (regions). The center of this window is between the pixels. It should be chosen as close as possible to the exact coordinates of the key point. A circle denotes a convolution window with a Gaussian core (similar to a window for calculating the direction of a key point). For this kernel, a is determined equal to half the width of the handle window. In the future, the value of each point in the window of the descriptor will be multiplied by the value of the Gaussian core at this point, as a weight coefficient.

On the right, the descriptor of a singular point, dimension 2 × 2 × 8, is schematically shown. The first two digits in the dimension value are the number of regions horizontally and vertically. Those squares that covered a certain region of pixels in the left image, on the right, cover histograms built on the pixels of these regions. Accordingly, the third digit in the dimension of the descriptor means the number of components of the histogram of these regions.

Three real coordinates (x, y, n) can be assigned to each gradient in the descriptor window, where x is the horizontal distance to the gradient, y is the vertical distance, n is the distance to the gradient direction in the histogram (meaning the corresponding histogram of the descriptor, to which this gradient contributes). The reference point is the lower left corner of the handle window and the initial value of the histogram. For single segments, we take the horizontal and vertical sizes of the regions for x and y, respectively, and the number of degrees in the histogram component for n. The coefficient of trilinear interpolation is determined for each coordinate (x, y, n) of the gradient as 1-d, where d is the distance from the coordinate of the gradient to the middle of the unit interval in which this coordinate fell. Each occurrence of the gradient in the histogram is multiplied by all three weighting coefficients of trilinear interpolation.

The cue point descriptor consists of all the histograms received. The resulting descriptor is normalized, after which all its components, the value of which is greater than 0.2, are trimmed to 0.2 and then the descriptor is normalized again. As such, the descriptors are ready for use.

In the next steps, the descriptors of the original images are compared with the images in the YIQ color space. At the final stage, the results of the comparison are combined.

The duplicate image search device (Fig. 4) contains an input that is connected to the input of the preprocessing unit of the first image 1 and to the input of the preprocessing unit of the second image 2, block 1 consists of a first image 1.1 registration unit, the output of which is connected to the input of the first image conversion unit YIQ 1.2 color space, the output of which is connected to the input of the in-phase component extraction unit of the first image 1.3, the output of which is connected to the input of the image forming unit as a result of rotation of the first 1.4 expressions, whose output is connected to the input of the imaging unit by changing the angle of inclination of the first picture 1.5, the output of which is connected to an input of the simulated image storage unit 1.6 for the first image, whose output is connected to the first input of the SIFT method of application; the output of the second image registration unit 2.1 is connected to the input of the second image conversion unit in the YIQ 2.2 color space, the output of which is connected to the input of the in-phase component extraction unit of the second image 2.3, the output of which is connected to the input of the image forming unit as a result of rotation of the second image 2.4, the output of which is connected to the input of the imaging unit when changing the angle of inclination of the second image 2.5, the output of which is connected to the input of the storage unit of simulated images for the second image 2.6, the output of which is connected to the second input of the SIFT method application block (the block diagram of the device is shown in FIG. connected to the input of the storage unit of the found pair of duplicates 5, the output of which is the information output of the device.

A device for searching for duplicate images is implemented and operates as follows. Two images enter the inputs of the registration blocks of the first 1.1 and second 2.1 images, on which it is necessary to find the corresponding descriptors and establish their similarity. For the purpose of processing images with mirror surfaces, images are converted into YIQ color space in blocks for converting images into YIQ 1.2 and 2.2 color space. Next, the in-phase component of the YIQ color space is selected in blocks 1.3 and 2.3. In the following blocks 1.4, 2.4, 1.5, 2.5, images are formed as a result of rotation and a change in the angle of inclination, that is, the use of affine transformations. In blocks 1.6 and 2.6, all received images are recorded as a result of these transformations. In the application block of the SIFT method, the SIFT descriptors of all images are constructed by the specified method. The resulting descriptors for a pair of images are sent to the input of block 4, in which the number of identical descriptors is calculated and the decision is made that the second image is a duplicate of the first, after which the found duplicates are written to the storage unit of the found pair of duplicates 5.

EFFECT: comparison of descriptors as applied to the task of searching for duplicate images.

Claims (1)

A duplicate image retrieval device comprising a preprocessing unit for a first image and a preprocessing unit for a second image, a preprocessing unit for a first image comprising a first image registration unit; an image forming unit as a result of rotation of the first image, the output of which is connected to the input of the image forming unit when the angle of inclination of the first image changes, the output of which is connected to the input of the simulated image storage unit for the first image, the output of which is connected to the first input of the SIFT method application block; the second image registration unit, the image forming unit as a result of rotation of the second image, the output of which is connected to the input of the image forming unit when the angle of the second image changes, the output of which is connected to the input of the simulated image storage unit for the second image, the output of which is connected to the second input of the application unit SIFT method, the storage unit for the found pair of duplicates, the output of which is the information output of the device, characterized in that the output of the register block radios of the first image are connected to the input of the unit for converting the first image into the YIQ color space, the output of which is connected to the input of the in-phase component extraction unit of the first image; the output of the second image registration unit is connected to the input of the second image conversion unit in the YIQ color space, the output of which is connected to the input of the in-phase component extraction unit of the second image; the output of the application block of the SIFT method is connected to the input of the unit for calculating the number of identical descriptors, the output of which is connected to the input of the storage unit of the found pair of duplicates.

Images