RU2522869C2 - Method for recognition and classification of object shape in labyrinth domain structures - Google Patents

Abstract

FIELD: physics.

SUBSTANCE: method includes determining the number of objects on an image of a structure; using roundness and filling coefficients as morphological characteristics for each object; creating a fuzzy knowledge base for separating objects into rounded, elliptical and dumbbell shaped using a triangular membership function, and for non-rounded strip-like and branched objects - using a trapezoidal membership function based on experimental data of values of said roundness and filling coefficients; performing recognition of domains; generating a fuzzy classifier for separating objects according to shape into rounded, elliptical, dumbbell shaped, strip-like and branched objects based on the ratio of the roundness coefficient to the filling coefficient of the object; performing classification of the shapes of the objects.

EFFECT: classifying objects based on geometric features in labyrinth structures.

2 cl, 6 dwg, 2 tbl, 1 ex

Description

The invention relates to computer technology, namely to the analysis of the structure of objects in digital images, and can be used in computer vision systems, when conducting scientific experiments of objects and structures of varying degrees of complexity.

When solving a number of problems, it becomes necessary to analyze labyrinth structures containing a large number of various objects that differ from each other in shape and size. It is known that the parameters of devices on magnetic films (memory, logical, magneto-optical devices, elements of magnetophotonics) are determined by the structure of magnetic films. The film structure contains a large number of domains - separate micro- and nano-objects of various shapes. Examples of such structures are labyrinthine domain structures in magneto-optical materials containing round, elliptical, dumbbell, strip and branching domains, for example, in ferrite-garnet films intended for the implementation of a wide class of devices. The need for recognition and analysis of the shape and classification of such micro-objects is necessary primarily for technological purposes.

Various methods of recognition, classification of micro-objects in digital images are described. So, in the invention (RU 2385494 C1, Nikitaev et al. March 27, 2010) a method for recognizing images of the texture of objects in digital photographs of biological objects is described. The method includes determining the texture features of an object, which consists in pre-processing the obtained images of the object through segmentation, calculating the values for the texture features and forming matrixes of numbers, and identifying and classifying the object. The most informative signs for classification are automatically determined on the basis of training samples from previously created image databases, based on which identification and classification are performed according to the following indicators: energy, moment of inertia, maximum probability of the neighbors that are most often found in this image, local uniformity, entropy, trace of the normalized matrix of spatial adjacency (NMPS), the average value of the brightness, the correlation of the brightness values of the image. In the patent (RU 118774 U1, Zhuravlev et al. 07/27/2012) a device is described that implements a method for analyzing objects against a heterogeneous background, including a morphological image filtering module that is capable of suppressing objects with the highest brightness values and a given shape, preserving the remaining boundaries of the desired objects unchanged; highlighting local minima of image brightness; determining the distance from each pixel of the image to the nearest background pixel and highlighting the local maximums of the image brightness. The boundaries of noise objects are filtered, the morphological gradient of the image is determined, the optical and geometric characteristics of the detected objects are calculated, and the database of coordinates, optical and geometric characteristics of the detected objects is calculated. However, the above methods do not include recognition of labyrinth and branching objects.

The invention (RU 2297039 C2, Menyaev et al. 04/10/2007) describes the identification of complex objects in an image by extracting a feature from the analyzed images that can effectively describe images of complex structure and distortion-invariant. Images of all reference objects are divided into intersecting domain blocks, and the image of the analyzed object is divided into disjoint rank blocks, the size of which is smaller than domain blocks. Then, a search is made for the best comparison of all ranking blocks of the analyzed image and domain blocks of all the reference images using compressive affine transformations. After that, the distance vectors between the geometric centers of the mapped domain for the reference object and the rank blocks for the analyzed object are formed. However, this method is based on comparing the analyzed object with reference objects and thus does not provide for the recognition of objects with a previously unknown shape.

Application (GB2397423 (A) KOKKO ERIC GERARD et al. 07/21/2004) describes a method for identifying and counting seeds and other small objects, including training a model of a single neural network with training sets of known objects; substantiation of the optimal neural network model; analysis of unknown objects with unknown parameters by visualizing them to obtain a digital image that includes pixels representing unknown objects, background and extraneous objects. However, this method does not provide for the simultaneous recognition of objects of several types.

The invention (RU 2132061C1, Honey et al. 06/20/1999) describes a method for segmentation and recognition of cells in images of cytological preparations in the field of view of a light microscope. First, the parameters of the models of the contours of cellular structures are set up on a small training set of images. In the process of segmenting the boundaries of cell structures, the digitized frame with the image is divided into facets and then the boundaries of the cell structures are formed. However, this method is designed to recognize objects of the same type - cells.

The closest analogue to the patented is a method for the recognition and classification of microobjects according to patent RU 2346331 C1, Kurkin et al. 02/10/2009 - prototype.

During testing, the image is transmitted using a web camera to a personal computer (PC). In a personal computer, a neural network trained for each type of LCD is created. Before starting the testing process, the LCD images are digitally filtered by isolating and analyzing the connected areas on the black-and-white image. By connected areas, a set of morphometric characteristics is determined to classify areas according to their shape, position and orientation. Comparison is carried out automatically, using a neural network stored in a PC and personally trained for each type of LCD to recognize those portions of the image of the LCD working surface that contain the previously determined morphometric characteristics of the connected areas.

However, this method does not provide for the classification of shape-changing objects that are components of two-dimensional labyrinth structures and cannot be used to recognize and classify objects in labyrinth structures.

The present invention is directed to solving the technical problem of object recognition and their classification according to geometric features in labyrinth structures, for example, magneto-optical domain media.

A method for recognizing and classifying the shape of objects in labyrinth domain structures involves obtaining a digital image of the structure, filtering it, binarizing and analyzing the morphological features of the objects.

The patented method is characterized in that the total number of objects in the image of the structure is determined, roundness and filling factors for each of the objects are used as morphological features.

A fuzzy knowledge base is formed for dividing objects into round, elliptical and dumbbell-shaped using a triangular membership function, and for non-circular strip and branching objects using a trapezoidal membership function based on experimental data of the values of these roundness and filling coefficients, and domain recognition is performed.

Next, a fuzzy classifier is formed to divide objects in shape into round, elliptical, dumbbell, strip and branchy objects based on the ratio of the roundness coefficient and the fill factor of the object and classify the shape of the objects. When classifying a non-circular object, the construction of its skeleton is carried out and the success / failure transformation is performed for it, then branch points are identified and, if present, the non-circular object is referred to as branchy, and in the absence, to strip domain structures.

The method can be characterized in that the fuzzy classifier is formed taking into account the numerical values of the coefficient k ₁ roundness and coefficient k ₂ filling, lying in the ranges:

for round objects - k ₁ = 0.77-1.0; k ₂ = 0.83-1.0;

for elliptical objects - k ₁ = 0.62-0.77, k ₂ = 0.71-0.83;

for dumbbell-shaped objects - k ₁ = 0.5-0.62, k ₂ = 0.62-0.71;

for non-circular strip, branched - k ₁ = 0.01-0.9, k ₂ = 0.1-0.62,

moreover, the roundness coefficients k ₁ and filling k ₂ are described by triangular membership functions for round, elliptical, dumbbell-shaped objects and trapezoidal membership functions for non-circular objects.

The technical result is the provision of recognition of objects and their classification by geometric features in labyrinth structures.

The essence of the patented method is illustrated in the figures, where:

figure 1 is a block diagram of a magneto-optical device for obtaining micrographs;

figure 2 - the initial micrograph of the domain structure in the film of ferrite garnet with the selection of the area for subsequent analysis;

figure 3 - membership functions µ (k) roundness coefficients k ₁ and filling k ₂ for classes d1-d4;

figure 4 - image selected in figure 2 of the area after pre-processing;

figure 5 - the results of the distribution of domains into classes;

6 is an image of the skeletons of non-circular domains with branch points found.

The method is implemented as follows. Images of structures are obtained using a digital camera using the magneto-optical method using the Faraday effect. On the block diagram of the installation (figure 1) shows: And - the light source; SL - light beam; L - lenses; P, A - polarizer and analyzer; EM is an electromagnet; MO — magneto-optical film under investigation; T - temperature controller; FC - digital camera, PC - personal computer. The origin and formation of the labyrinthine domain structure in magnetic materials occurs under the influence of external magnetic and / or temperature fields.

In a demagnetized state in a magnetic material, a ferrite garnet film several micrometers thick, a labyrinth domain structure containing domains of various shapes is usually realized. Figure 2 presents the original micrograph of the domain structure in a film of ferrite garnet. The analysis area is highlighted in the frame. The width of the labyrinth domains is comparable to the film thickness.

In order to determine the ranges of variation of the shape factors of objects of various classes, several hundred images of various labyrinth domain structures were analyzed. It has been established that for dividing objects into round (class d1), elliptic (class d2), dumbbell-shaped (class d3) and non-circular (strip, branchy: class d4), it is sufficient to use the roundness coefficient k ₁ and the fill factor k ₂ as informative signs. Membership functions µ (k) for objects of the indicated classes in domain structures are shown in FIG. 3. Membership functions are based on the distribution of experimental data of the values of the roundness and filling coefficients.

To describe the distribution of objects close to round (classes d ₁ , d ₂ , d ₃ ), we used the triangular membership function µ (k):

$$\mu \left(k\right)=\{\begin{array}{l}0k\le a\\ \frac{k-a}{b-a},a\le k\le b\\ \frac{c-k}{c-b},b<k<c\\ 0c\le k\end{array},(\mathrm{one})$$

where [a, c] is the range of the variable, b is the most possible value of the variable. For non-circular objects, the trapezoid function was used:

$$\mu \left(k\right)=\{\begin{array}{l}0k\le a\\ \frac{k-a}{b-a},a<k<b\\ \mathrm{one,}b\le k\le c\\ \frac{d-k}{d-c},c<k<d\\ 0d\le k\end{array},(2)$$

where [a, d] is the carrier of the fuzzy set, [b, c] is the core of the fuzzy set.

The classification task is to assign the object specified by the vector of informative features K = (k ₁ , k ₂ ) to one of the previously defined classes {d ₁ , d ₂ d ₃ , d ₃ } i.e. consists in performing the display of the form (see, Shtovba SD. Designing fuzzy systems using MATLAB tools. - M.: Hot line - Telecom, 2007):

$$K=\left(k_{\mathrm{one}},k_2\right)\to y\in \left\{d_{\mathrm{one}},d_2,d_3,d_{\mathrm{four}}\right\}.(3)$$

Here y is the output parameter of the fuzzy classifier. Classification based on fuzzy inference occurs on the basis of knowledge of the form:

$${\displaystyle \underset{p=\mathrm{one}}{\overset{\mathrm{four}}{\cup }}\left[{\displaystyle \underset{i=\mathrm{one}}{\overset{2}{\cap }}\left(k_i=a_i^{jp}\right)}\mathrm{from}\mathrm{at}e\mathrm{from}\mathrm{about}m{\omega }_{jp}\right]}\to y=d_j,j=\overline{\mathrm{one}...\mathrm{four}},(\mathrm{four})$$

- нечеткий терм, которым оценивается переменная k в строчке с номером jp (р=1,n);Where: $$a_i^{ip}$$

- the fuzzy term that evaluates the variable k in the line with the number jp (p = 1, n);

n _j is the number of lines - conjunctions in which the output y is estimated by the fuzzy term d _j , j = 1, m;

m is the number of terms used to linguistically evaluate the output parameter y.

Based on the source data (membership function of form factors, see Fig. 3), a fuzzy knowledge base is presented for the construction of a fuzzy classifier, presented in Table 1.

The weights of the rules ω _{jp are} tuned during the training of the fuzzy classifier. The training was conducted on the basis of about two hundred experiments.

, к классам d_j из базы знаний рассчитываются следующим образом:Degrees of belonging of the classification object, the informative features of which are given by the vector $$K^{*}=\left(k_{\mathrm{one}}^{*},k_2^{*}\right)$$

, to classes d _j from the knowledge base are calculated as follows:

$${\mu }_{dj}\left(K^{*}\right)=\underset{p=1.4}{\max }{\omega }_{jp}\underset{i=\mathrm{1,2}}{\min }\left[{\mu }_{dj}\left(k^{*}i\right)\right],j=\overline{\mathrm{one}...\mathrm{four}}(5)$$

As a solution, choose a class with a maximum degree of belonging

$$y^{*}=\underset{\left\{d_{\mathrm{one}},d_2,d_3,d_{\mathrm{four}}\right\}}{\arg }\max \left({\mu }_{d\mathrm{one}}\left(K^{*}\right),{\mu }_{d2}\left(K^{*}\right),{\mu }_{d3}\left(K^{*}\right),{\mu }_{d\mathrm{four}}\left(K^{*}\right)\right)(6)$$

As a criterion for dividing non-circular objects into strip and branching objects, the authors selected a criterion for the presence or absence of branching of the studied objects based on the success / failure transformation (Digital Image Processing using MATLAB / RC Gonzalez, RE Woods, SL Eddins. - Pearson Education. 2004) and includes the following steps:

1) Construction of the skeleton of a non-circular object.

2) Removal of parasitic components of the skeleton.

3) Performing a success / failure transformation for the object backbone.

4) Counting the number of branch points of an object.

In the presence of branch points, a non-circular object refers to branching, in their absence - to strip domains.

Example. As an illustration of the achievement of the technical result, an example of image processing containing round, elliptical, dumbbell, strip and branching domains is shown (in Fig. 2, for clarity, a small portion of the structure for processing is highlighted). Computer processing was performed in the MatLab environment.

Figure 4 shows the image for analysis after pre-processing, which includes filtering and binarization (see Analysis and recognition of nanoscale images: traditional approaches and new problem statements. / V.A.Soyfer, A.V. Kupriyanov // Computer Optics . - 2011. - T.35, No. 2. - S.136-144).

Figure 5 shows the results of the distribution of domains into classes. For each object in a binary image, a roundness coefficient k ₁ (vertical axis, FIG. 5) and a fill factor k ₂ (horizontal axis, FIG. 5) are calculated. The functions of belonging of µ objects to a certain type constructed using fuzzy clustering are given depending on the values of the coefficients k ₁ , k ₂ . On the axes, intervals corresponding to one of the predefined classes of objects are marked: round objects (class d1), elliptical (class d2), dumbbell-shaped (class d3) and non-circular (strip, branchy: class d4). As a result, objects belonging to the same class are localized in a strictly defined coordinate region in Fig. 5, which allows realizing domain recognition based on the proposed knowledge base.

Figure 6 shows the distribution of domains in areas that correspond to the previously mentioned classes of objects d1-d4.

With further analysis for non-circular objects, the construction of the cores is performed (Fig.6). For each object, the success / failure transformation is applied, which allows determining the presence of branch points at domains. For clarity, the image in Fig.6 is inverted, and the found branch points are highlighted.

Table 2 shows the results of the domain classification for the original test image shown in FIG. 4.

Thus, the patented method provides recognition of round, elliptical, dumbbell-shaped and non-circular objects in labyrinth domain structures and, in addition, the classification of non-circular objects into strip and branch ones. The invention can be used to analyze images of various origins, incorporating thousands of objects of various shapes.

TABLE 1 Fuzzy knowledge base for classifying objects k ₁ k ₂ ω _jp y Round Round one Round Elliptical 0.85 d1 Round Dumbbell 0.7 Round Non-circular 0.32 Elliptical Round 0.64 Elliptical Elliptical one d2 Elliptical Dumbbell 0.83 Elliptical Non-circular 0.37 Dumbbell Round 0.48 Dumbbell Elliptical 0.75 d3 Dumbbell Dumbbell one Dumbbell Non-circular 0.45 Non-circular Round 0.29 Non-circular Elliptical 0.45 d4 Non-circular Dumbbell 0.6 Non-circular Non-circular one

TABLE 2 Domain classification results No. Type of form Qty one Round domain 119 2 Elliptical domain 13 3 Dumbbell domain 12 four Strip domain 16 5 Branch domain eleven

Claims (2)

1. A method for recognizing and classifying the shape of objects in labyrinth domain structures, including obtaining a digital image of the structure, filtering it, binarizing and analyzing morphological features of objects,
characterized in that the total number of objects in the image of the structure is determined,
as morphological features using roundness and filling factors for each of the objects,
form a fuzzy knowledge base for dividing objects into round, elliptical and dumbbell-shaped using a triangular membership function, and for non-circular strip and branching objects using a trapezoidal membership function based on experimental data of the values of these roundness and filling coefficients, and domain recognition is performed,
further, a fuzzy classifier is formed for dividing objects in shape into round, elliptical, dumbbell, strip and branchy objects based on the ratio of the roundness coefficient and the fill factor of the object and classifies the shape of the objects,
when classifying a non-circular object, the construction of its skeleton is carried out and the success / failure transformation is performed for it, then branch points are identified and, if present, the non-circular object is referred to as branchy, and in the absence, to strip domain structures. 2. The method according to claim 1, characterized in that the fuzzy classifier is formed taking into account the numerical values of the coefficient k ₁ roundness and coefficient k ₂ filling, lying in the ranges:
for round objects - k ₁ = 0.77-1.0; k ₂ = 0.83-1.0;
for elliptical objects - k ₁ = 0.62-0.77, k ₂ = 0.71-0.83;
for dumbbell-shaped objects - k ₁ = 0.5-0.62, k ₂ = 0.62-0.71;
for non-circular strip, branched - k ₁ = 0.01-0.9, k ₂ = 0.1-0.62,
moreover, the roundness coefficients k ₁ and filling k ₂ are described by triangular membership functions for round, elliptical, dumbbell-shaped objects and trapezoidal membership functions for non-circular objects.

Images