KR20090082546A - Method for recognizing a target in images - Google Patents

Abstract

A target recognition method within an image is provided to pass through 2-stage matching processes, thereby obtaining a high recognition rate regardless of rotational and size changes of a target. A boundary line of a recognition object target within a random image is extracted(S101). A reference point is set inside a random-type figure based on boundary line information of the extracted target. A distance between the set reference point and all pixels in the boundary line is calculated. A DFV(Distance Feature Vector) is produced by using the calculated distance information(S104). The produced DFV is divided into plural sections. The first modeling is performed(S105). The second modeling is performed(S106). The first matching is executed to search optimal matching order for each target(S107).

Description

Method for recognizing a target in images}

The present invention relates to a method for recognizing a target in an image, and more particularly, to a method for recognizing a target in an image that can obtain a high recognition rate even when the rotation and size of the target change.

In order to recognize a target existing in a high resolution image, conventionally, region-based target information (eg, an airplane, an armored vehicle, a tank, etc.) is first extracted, and the brightness of each component pixel (pixel) is characterized. As a vector), a target model was identified by calculating a match value with a reference model stored in a previously constructed target database (Database). However, this method has a disadvantage in that the recognition rate of the target is significantly lowered when the size of the target changes due to the rotation of the target or the difference in resolution. Therefore, in general, a method of normalizing the size and calculating all matching values for arbitrary rotation angles is used. However, since the matching values need to be calculated for all cases, there is a problem in that calculation time is required. In addition, accuracy is also related to how finely rotated the match is. In particular, when the size is changed, the area of the target is small so that the recognition itself may be impossible when overlapping with another target model.

The present invention has been created in view of the above problems, and extracts the boundary of a target (graphic) image and sets the distance between the reference point in the relatively invariable target (graphic) and the pixels on the boundary of the figure as feature vectors. It is an object of the present invention to provide a method for recognizing a target in an image that can obtain a high recognition rate even by changing the rotation and size of the target by the matching process.

In order to achieve the above object, a target recognition method in an image according to the present invention includes:

Extracting a boundary of a target to be recognized in an image;

Setting a reference point in an arbitrary shape of a figure formed of a boundary line based on the extracted boundary line information of the target;

Calculating a distance between the set reference point and all pixels on the boundary line;

Generating a distance feature vector (DFV) using all the calculated distance information and dividing the generated DFV into a plurality of sections;

Performing first-order modeling on each of the divided sections to find an optimal matching order with a previously constructed reference model;

Performing secondary modeling to generate information necessary for calculating a final match value for the optimal match order;

Searching for an optimal matching order for each target by performing a first matching using the information extracted in the first modeling process;

Calculating a final match value by performing a second match using information extracted in the second modeling process according to the found optimal matching order; And

Distinguishing the target to be recognized based on the calculated final match value.

Here, in the first modeling step, information about the relative size of the peaks and the separation distance between the vertices is extracted for each of the divided sections.

In the second modeling step, the DFV is approximated to a series of polynomials for each divided section used in the first modeling, and coefficient information expressed as a series is extracted.

In the first matching step, the optimal matching sequence is searched by calculating the matching value while cyclically moving only the divided sections using the relative size between the vertices of the sections extracted during the first modeling and the distance information between the vertices.

In the second matching step, a matching error between the polynomial series coefficient extracted during the second modeling and the feature vector of the test target to be matched is calculated, and the final matching value is calculated based on the matching error.

According to the present invention as described above, the rotation of the target by extracting the boundary of the target image and performing a two-step matching process using the distance between all pixels on the boundary of the figure and the reference point in the relatively invariant target (graphic) as a feature vector And high recognition rate can be obtained regardless of the size change.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a flowchart illustrating a method of recognizing a target in an image according to the present invention.

Referring to FIG. 1, a boundary line 201 of a target to be recognized (figure) 200 (see FIG. 2) in an arbitrary image is first extracted according to a target recognition method in an image according to the present invention (step S101). In this case, a general image processing technique may be used to extract the boundary line 201 of the target (figure) 200.

When the extraction of the boundary line of the target is completed, the reference point 202 is set in the inside of the figure 200 having any form of the boundary line based on the extracted boundary line information of the target (step S102). Then, a projection distance d1 to d8 between the set reference point 202 and all the pixels 203a to 203h on the boundary line 201 is calculated (step S103). In this case, the distance can be easily calculated by obtaining coordinate values of the reference point and respective coordinate values of all pixels on the boundary line.

In this way, when the distances d1 to d8 between the reference point 202 and all the pixels on the boundary line are calculated, a distance feature vector (DFV) is generated as shown in FIG. 3 using all the calculated distance information. The generated DFV is divided into a plurality of sections (step S104). For each of the divided sections, first-order modeling is performed to find an optimal matching order with a previously constructed reference model (step S105).

In this regard, referring to the DFV characteristic of FIG. 3, it can be seen that the highest value and the lowest value exist in part according to the boundary feature of the target (tilt). First modeling is implemented as follows. That is, primary modeling is established by calculating the relative size between the vertices and the separation distance between each vertex for each section. FIG. 4 (A) shows the feature vectors of 16 sections obtained through the first-order modeling. In this way, only 16 feature section information is processed without performing matching while cyclically moving the entire pixel, thereby optimizing an optimal matching sequence. ), Which can greatly reduce the time required to find the best match.

As described above, when the primary modeling is completed, the secondary modeling is performed to generate information necessary for calculating the final matching value for the optimum matching sequence (step S106). In this secondary modeling step, the DFV is approximated to a series of polynomials for each divided section used in the first modeling, and coefficient information expressed as a series is extracted. 4B shows an example of the second order polynomial series obtained through the second-order modeling. Coefficients approximated in the form of the polynomial series for each section are used as matching information.

Subsequently, the first matching is performed using the information extracted in the first modeling process to search for an optimal matching order for each target (step S107). At this time, in the first matching step, the optimal matching sequence is searched by calculating a matching value while cyclically moving only the divided sections by using the relative size between the vertices of each section extracted during the first modeling and the distance between the vertices. .

FIG. 5 conceptually illustrates a cyclic shift matching process, in which one reference model has a first modeled section length of N and a section length of an input target to be matched has a cycle of M As we move, we calculate the matching values to find the optimal matching order.

In this way, when the search for the optimum matching sequence is completed, the second matching is performed by using the information extracted in the second modeling process according to the found optimum matching sequence to calculate the final matching value (step S108). In this second matching step, a matching error between the polynomial series coefficient extracted during the second modeling and the feature vector of the test target to be matched is calculated, and the final matching value is calculated based on the matching error.

When the final matching value is calculated by the above, the recognition target is identified based on the calculated final matching value (step S109).

FIG. 6 is a diagram illustrating an example of a figure pattern for a recognition experiment for a target (figure) according to the target recognition method in the image of the present invention as described above.

As in FIG. 6, six targets (figures) were prepared for recognition experiments on the targets (figures). Data set # 1 is the most basic target model, and the modeling was built and recognition experiments were conducted using the hexagonal object (number 1) in the upper left corner as the target target. Data set # 2 is a test target which rotates the basic target model of the data set # 1 at an arbitrary angle, and data set # 3 is for recognition experiments on targets whose rotation and size are changed.

7 shows the results for these recognition experiments. FIG. 7 (A) shows that the hexagonal target (No. 1), which is the reference model, and the second matching of the targets for the data set # 1 have been correctly recognized. Can be. In addition, it can be seen that the case of rotating at an arbitrary angle (Fig. 7 (B)) is also recognized as a difference in matching value of -15 dB or more. It can be seen that it is recognized as the difference of the matching values.

Therefore, as can be seen through such experiments, the method of the present invention can obtain a high recognition rate regardless of rotation and size change with respect to the figure (target) to be recognized.

As mentioned above, although the present invention has been described in detail with reference to the preferred embodiment, the present invention is not limited thereto, and it will be apparent to those skilled in the art that various changes and applications can be made without departing from the technical spirit of the present invention. Therefore, the true protection scope of the present invention should be interpreted by the following claims, and all technical ideas within the equivalent scope should be construed as being included in the scope of the present invention.

1 is a flow chart showing an execution process of a method for recognizing a target in an image according to the present invention.

FIG. 2 conceptually illustrates finding the distance between a reference point in a target (shape) and all pixels on the boundary of the figure in accordance with the method of the present invention.

3 shows generation of a DFV using calculated distance information between a reference point in a target (shape) and all pixels on the boundary line according to the method of the present invention.

4 is a view showing a 16-section feature vector obtained through first-order modeling and a second-order polynomial series obtained through second-order modeling according to the method of the present invention.

5 conceptually illustrates a cyclic movement matching process in the method of the present invention.

6 shows an example of a figure pattern for a recognition experiment on a target (figure) according to the method of the present invention.

FIG. 7 shows the results of a recognition experiment on the target (figure) of FIG. 6. FIG.

<Explanation of symbols for the main parts of the drawings>

200 ... Recognized target (figure) 201 ... Borderline

202 Reference points 203a to 203h ... pixels

d1 to d8 ... distance between reference point and pixel

Claims (5)

Extracting a boundary of a target to be recognized in an image; Setting a reference point in an arbitrary shape of a figure formed of a boundary line based on the extracted boundary line information of the target; Calculating a distance between the set reference point and all pixels on the boundary line; Generating a distance feature vector (DFV) using all the calculated distance information and dividing the generated DFV into a plurality of sections; Performing first-order modeling on each of the divided sections to find an optimal matching order with a previously constructed reference model; Performing secondary modeling to generate information necessary for calculating a final match value for the optimal match order; Searching for an optimal matching order for each target by performing a first matching using the information extracted in the first modeling process; Calculating a final match value by performing a second match using information extracted in the second modeling process according to the found optimal matching order; And And identifying the target to be recognized based on the calculated final matched value. The method of claim 1, In the first modeling step, the target recognition method in the image, characterized in that for extracting the information about the relative size of the peak (peak points) and the separation distance between the vertex for each divided section. The method of claim 1, And approximating the DFV to a series of polynomials for each division section used in the first modeling in the second modeling step, and extracting coefficient information expressed in a series. The method of claim 1, In the first matching step, the optimum matching sequence is searched by calculating a matching value while cyclically moving only the divided section by using the relative size between the vertices of the sections extracted during the first modeling and the distance information between the vertices. Target recognition method in the image. The method of claim 1, In the second matching step, a target error in the image is calculated by calculating a matching error between a polynomial series coefficient extracted during the second modeling and a feature vector of a test target to be matched, and calculating a final matching value based on the matching error. Way.

Images