RU2514155C1 - Method for automatic identification of objects on images - Google Patents

Abstract

FIELD: physics, computer engineering.

SUBSTANCE: invention relates to computer science and can be used for automatic identification of objects on images. The method involves scanning the original high-resolution photographic image. The array of the obtained readings is converted to the scale of the reference array by normalising brightness pixels with a scaling coefficient. The obtained image is broken down into three two-dimensional arrays in the standard RGB colour palette. Outlines of objects are selected using techniques for spatial differentiation of the array signal function. The relief surfaces of the objects inside the selected outlines are approximated with a mosaic of triangles. The area of the mosaic in each channel is calculated using the Heron formula and the obtained surface relief areas of the objects are compared with values thereof for references based on a validation criterion: $$\frac{\left|S_{\text{ref}}-S_{\text{ob}}\right|}{S_{\text{ref}}}\cdot 100\%.$$

EFFECT: automating recognition with high accuracy.

7 dwg

Description

The invention relates to computer science and can find application in systems for the collection, conversion, processing of information in various fields of human activity: forensics, astronautics.

Remote sensing of the Earth from space in the interests of mineral exploration, forestry, ecology, monitoring of soil cover is carried out by obtaining digital images of the underlying surface. Selective features of objects in the images are: color, tone, texture, topology. There are multifunctional editing programs, image retouching, working with layers in the RGB, CMYK, Lab, etc. palette in interactive mode [see, for example, Minko P.A. Graphics processing in Photoshop CS2. M .: Publishing. Eksmo, 2007, pp. 71-89, pp. 145-151].

While there are no universal methods for automatic identification of objects in images. The largest amount of information for detecting objects in images and their identification is contained in their form. Psychologically, the perception of images by a human operator occurs at the outline level, i.e. outline shape of an object. The contour is the edge where the gradient of the signal function changes most rapidly.

There is a method of highlighting the contour of a drawing of an object [see, for example, Duda P.O., Hart P.E. Pattern recognition and scene analysis. Perev. from English, Moscow: Mir, 1976, §7-3, “Spatial differentiation”, pp. 287-288, Fig. 7.3] - analogue. The outline drawing is obtained by calculating the gradient of the scalar brightness function I (x, y) of the video image at each point as:

$$gradI\left(x,y\right)=\frac{dI}{dx}\cdot i+\frac{dI}{dy}\cdot j$$

To obtain a contour drawing, a regular operator with a window aperture of | 2 × 2 | item:

I j i, j + 1 i + l, j i + l, j + l

Window elements are connected along diagonals (two mutually orthogonal directions) by a subtraction operation. The Roberts operator is calculated at each point:

R (i, j) = | I (i, j) -I (i + 1, j + 1) | - | I (i + 1, j) -I (i, j + 1) |,

displays points for which R (i, j) ≥ threshold.

The disadvantages of the analogue should be considered:

- not all signs of the image signal are used to identify the object, in particular, the analog uses one signal parameter - the amplitude;

- lack of reliability in the visual analysis of the object by the operator.

The closest analogue to the claimed technical solution is the "Method for detecting anomalies of the underlying surface." RF patent No. 2160912, cl. G01V 8/00, 2000. The closest analogue method involves obtaining an image of the underlying surface in the form of a digital matrix of the brightness function I (x, y) from spatial coordinates, dividing the image into relatively homogeneous regions of tone based on a priori data, calculating the fractal dimension of each region , compiling a matrix of standards from the coefficients of the fractal dimension of each section, the difference between the current and reference values of the fractal dimension for the threshold level for the analyzed section.

The disadvantages of the closest analogue are:

- the uncertainty of dividing the image into a mosaic of sections, leading to an error in calculating the fractal dimension;

- not all independent features of the image are used for identification;

- a small interval of changes in the function of the fractal dimension of the image, in the range of 2.1 ... 2.7, which reduces the reliability of identification.

The problem solved by the claimed method consists in automatic, reliable identification of objects in the images by highlighting the contour drawing of the object and quantitative comparison of the surface area of the surface of the object inside the contour with the reference.

, разложение полученного изображения на три двумерные матрицы в стандартной палитре цветов RGB, выделение методами пространственного дифференцирования контурных рисунков объекта в каждом из каналов RGB, аппроксимацию поверхности рельефов объекта А внутри выделенных контуров мозаикой треугольников в окнах |2×2| элементов, расчет по формуле Герона площадей мозаики треугольников в каждом из каналов, сравнение полученных площадей с их значениями для эталонных объектов в каждом из каналов и их суммой, идентификацию образа объекта по совпадению площадей поверхности рельефов объекта (S_об) и эталона (S_этал) с установленной достоверностью:The technical result is achieved in that a method for automatically identifying objects in images includes scanning the original photo image with high linear and amplitude resolution, bringing the matrix of obtained samples I (x, y) to the scale of the reference matrix by normalizing the pixels with a scale factor $$M=\frac{I}{I_{\max }}\times I_{\max \text{scales}}$$

decomposition of the resulting image into three two-dimensional matrices in the standard RGB color palette, extraction by spatial differentiation of the contour drawings of the object in each of the RGB channels, approximation of the relief surface of object A inside the selected contours by a mosaic of triangles in the windows | 2 × 2 | elements, calculation according to the Heron formula of the areas of the mosaic of triangles in each channel, comparing the obtained areas with their values for the reference objects in each channel and their sum, identifying the image of the object by the coincidence of the surface area of the reliefs of the object (S _vol ) and the standard (S _etal ) with established reliability:

, $$\frac{|\; |\; |S_{\mathrm{uh}t\mathrm{but}l}-S_{\mathrm{about}b}|\; |\; |}{S_{\mathrm{uh}t\mathrm{but}l}}\cdot \mathrm{one\; hundred}\%$$

,

where: I _{max scale} - the maximum value of the parameter scale of the scanner used;

I _max - the maximum amplitude of the pixel signal in the matrix;

I is the current pixel value of the signal in the matrix.

The invention is illustrated by drawings, where:

figure 1 - source image;

figure 2 - contour drawing of an object in one of the layers of RGB;

figure 3 - sequence of dividing the image matrix into windows;

figure 4 - splitting the window into two pairs of adjacent triangles;

5 is a functional diagram of a device that implements the method.

The technical essence of the invention is as follows. Many scientific tasks in remote research from space are solved by obtaining an image of the underlying surface: the detection of earthquake foci, the identification of areas of environmental anomalies, conflagration, reconnaissance of military infrastructure, position areas of the troops, etc.

To date, the task of recognizing objects in images of the underlying surface is usually solved by an operator using a set of decryption features. As noted above, the largest amount of information contains the shape of the object, and on the image - its outline. Photos with a resolution of about 200 dots per 1 mm have the highest resolution. For digital processing, it is necessary to convert the image into a digital matrix, and any conversion of information leads to losses. To reduce losses use a high resolution scanner. The average brightness of the images depends on the shooting conditions: time of day, the height of the Sun, the state of the environment (atmosphere), and equipment parameters.

To exclude the influence of these factors on the reliability of identification, it is necessary to bring the compared images to a single scale, i.e. normalize the brightness pixels of the original matrix by multiplying their amplitude by the scale factor mentioned above. After scaling the source image, the matrix is decomposed into three orthogonal ones in the standard RGB color palette. The spatial differentiation of the signal function of the matrices of the method-analogue distinguish contour drawings of the object in all three channels. The selected contour pattern in one of the channels is illustrated in figure 2. The contouring program is shown below in an example implementation. Identification of objects by shape is carried out by calculating the surface area of their surface within the boundaries of the selected contours in the channels R, G, B and comparing them with the reference values.

Figure 3 presents an illustration of the sequence of partitioning the matrix into windows of 4 adjacent pixels. Each pixel is characterized by a resolution in the coordinates Δx, Δy, amplitude, depth ΔН, which are considered known from the technical characteristics of the means. The size of the outline is determined by the number of rows and columns of the matrix. The algorithm for calculating the elementary area of the window of a four-point pattern is based on triangulation, i.e. dividing it with diagonals 1-4 and 3-2 into two pairs of adjacent triangles. The splitting procedure is illustrated in FIG. The area of each triangle is calculated by the Heron formula. Previously, according to the Pythagorean theorem, the lengths of the sides of the triangles are calculated. In accordance with figure 4, the lengths of the sides of the triangle, for example, with vertices 1-3-4 are equal:

; $$l_{3-1}=\sqrt{\Delta y^2+\Delta {\left(H_1-H_3\right)}^2}$$

; $$l_{3-\mathrm{four}}=\sqrt{\Delta x^2+\Delta {\left(H_{\mathrm{four}}-H_3\right)}^2}$$

; $$l_{3-\mathrm{one}}=\sqrt{\mathrm{<figure-callout\; id="2"\; label="\Delta \; y"\; filenames="00000016.png"\; state="\{\{state\}\}">\Delta \; </figure-callout>}{y}^{}{}^2+\Delta {\left(H_{\mathrm{one}}-H_3\right)}^2}$$

;

diagonal $$l_{\mathrm{one}-\mathrm{four}}=\sqrt{{\left(2\Delta x\right)}^2+{\left(2\Delta y\right)}^2+\Delta {\left(H_{\mathrm{one}}-H_{\mathrm{four}}\right)}^2}$$

There are two ways to triangulate - on the main diagonal (left-top - right-down) and on the auxiliary diagonal (right-top - left-down). The area is calculated in both ways, and the geometric mean is selected as the result. If at least one vertex of the triangle is outside the boundary of the plot, the area of the triangle is considered equal to zero. If all the vertices belong to the site, the area of the triangle is calculated by the Heron formula.

An example implementation of the method.

The claimed method can be implemented according to the scheme of figure 5. Functional diagram of the device of Fig. 5 contains hardware-software input-output 1 of the Internet, a flash card 2 of client images in one of the commonly used formats, a digital scanner 3 that converts client pictures in a standard RGB color palette into a special format, a personal computer 4 of program image processing in the standard configuration of the elements: processor 5, random access memory 6, hard drive 7, display 8, printer 9, keyboard 10, 11 reference objects implemented on the microprocessor base of the video card, digital discrimination object comparison torus 12.

The procedure for automatically identifying images of objects is as follows. Current images of objects to be identified are pumped from the Internet or flash cards to a personal computer. The operator pre-sorts the images displayed on the display visually.

On figa presents the original image of the poet A.A. Blok in 1920. As a reference image used A.A. Blok - a student at St. Petersburg University in 1902, fig.1b. A specialized image processing program allocates an outline drawing of an object (outline outline of the face of A.A. Blok)

The result of the software processing is illustrated in FIG. Then calculate the surface area of the relief inside the selected path, for which they write to the hard drive a specialized program, the text of which is given above.

The treated area may have an arbitrary configuration. To process it, you should select the border of the “gulf” section of the neighborhood in white (maximum brightness in the printer quantization scale).

. Поскольку исходное и эталонное изображения приводятся к одному масштабу, то площадь рельефа измеряют в условных единицах: разрешение Δx=Δy=1, а единица глубины - шаг квантования шкалы яркости.To calculate the surface area of the site, the entire contour is viewed sequentially, using a template, from four neighboring points forming a square. The template scans the contour from left to right, from top to bottom, and for each image element, the elementary area is calculated, then all elementary areas are summed. To increase the accuracy of calculating the surface area of the relief, the calculations are carried out using two triangulation methods, and the total area of the relief is found as geometric mean $$S_{\Sigma }=\sqrt{S_{\mathrm{one}}\cdot {\mathrm{<figure-callout\; id="2"\; label="S"\; filenames="00000016.png"\; state="\{\{state\}\}">S</figure-callout>}}_2}$$

. Since the original and reference images are reduced to the same scale, the relief area is measured in arbitrary units: resolution Δx = Δy = 1, and the unit of depth is the quantization step of the brightness scale.

The calculated results of the image signal parameters were:

RGB Image

Relief area Sr = 434346.935

Projection Area Sp = 59700.000

Maximum Intensity Imax = 255

Minimum intensity Imin = 40

Average intensity Isr = 158.54935

Points processed count = 60200

Blue - Image

Relief area Sr = 459476.328

Projection Area Sp = 59700.000

Maximum Intensity Imax = 255

Minimum intensity Imin = 25

The average intensity Isr = 140.05271

Points processed count = 60200

Green - Image

Relief area Sr = 415642.552

Projection Area Sp = 59700.000

Maximum Intensity Imax = 255

Minimum intensity Imin = 48

Average intensity Isr = 161.33251

Points processed count = 60200

Red - image

Relief area Sr = 440343.389

Projection Area Sp = 59700.000

Maximum Intensity Imax = 255

Minimum intensity Imin = 34

Average intensity Isr = 148.34257

Points processed count = 60200

Similar calculations of the image signal parameters of the RGB image standard: Sp = 418218.659.

With age (due to wrinkles), the surface area of the face increases, so the standard area is less than the area of the analyzed object. The criterion for the discrepancy between the image of the object and the standard in the total RGB channel was:

$$\frac{|\; |\; |S_{\mathrm{uh}t\mathrm{but}P}-S_{\mathrm{about}b}|\; |\; |}{S_{\mathrm{uh}t\mathrm{but}P}}\cdot \mathrm{one\; hundred}\%=\frac{|\; |\; |418218-434346|\; |\; |}{418218}\cdot \mathrm{one\; hundred}\%=3.9\%$$

Since the signals are considered statistically independent in the orthogonal channels, the resulting reliability of the identification of the image of the object from the three orthogonal channels (in probability) will be:

P _Σ ≈ (1-0.961) ₃ ≈0.99996, i.e. discrepancy in the fifth sign.

The effectiveness of the method is determined by such indicators as efficiency, reliability, documentary, reproducibility, accuracy.

The method is implemented on an existing technical basis. PC type "Intel", a scanner type "Panasonic" with a resolution of 1024 dpi, video cards of the type "NVidia", which incorporate a graphics chip GTX 250 with 512 arithmetic-logical devices, as well as 2 GB of RAM.

Claims (1)

A method for automatically identifying objects in images includes scanning the original photo image with high linear and amplitude resolution, bringing the matrix of obtained samples I (x, y) to the scale of the reference matrix by normalizing the pixels with a scale factor

decomposition of the obtained image into three two-dimensional matrices in the standard RGB color palette, extraction by spatial differentiation of the contour drawings of the object in each of the RGB channels, approximation of the surface of the object reliefs inside the selected contours by a mosaic of triangles in the windows | 2 × 2 | elements, calculation according to the Heron formula of the areas of the mosaic of triangles in each channel, comparing the obtained areas with their values for the reference objects in each channel and their sum, identifying the image of the object by the coincidence of the surface area of the reliefs of the object (S _vol ) and the standard (S _etal ) with established reliability:

,
where: I _{max scale} - the maximum value of the parameter scale of the scanner used;
I _max - the maximum amplitude of the pixel signal in the matrix;
I is the current pixel value of the signal in the matrix.

Images