KR101490027B1 - Apparatus and method for processing image - Google Patents

Abstract

The present invention relates to an image processing method and apparatus capable of efficiently configuring a feature region in an input image.
According to an embodiment of the present invention, there is provided an image processing method comprising: maintaining a database storing characteristic values of a target image; Receiving an input image; Extracting a partial image from the input image; Selecting a plurality of feature regions from the partial image; And verifying whether the selected plurality of feature regions are candidates that can be compared with the target image in consideration of a uniform distribution degree of the plurality of feature regions.

Description

[0001] APPARATUS AND METHOD FOR PROCESSING IMAGE [0002]

The present invention relates to an image processing method and apparatus, and more particularly, to an image processing method and apparatus capable of efficiently configuring a feature region in an input image.

Target recognition is one of the longest studied topics in image processing. Target recognition is based on a method such as SIFT or SURF, which computes the target similarity from the feature points of the target to unchanging feature information, the method of learning randomized trees, and the use of a very simplified number of decision trees Method (Fern algorithm) have been studied.

A conventional technique for detecting a target uses a method of selecting a plurality of regions having an arbitrary size and position to detect a target object from an input image and detecting the target using the arbitrary region. For example, in the target image, the feature extracted from each region is learned from the target image using the arbitrary region, the feature extracted from the same region is shifted by one pixel unit in the same size as the target image in the input image, Similar objects can be detected compared with learned features.

Since the prior art has selected the arbitrary region without any condition, it is possible to select a region shifted to one side, which may miss features that distinguish the target object from the non-target object.

Korean Patent Publication No. 10-2011-0062413 discloses a method of setting a first search area around a current center position of a target, calculating the center of each color of the first search area, Tracking the center position of the target using the weight of each color for the pre-computed target in the initial frame.

However, Korean Patent Publication No. 10-2011-0062413 does not disclose a method for verifying whether a selected region is comparable to a feature of a target image previously stored.

Therefore, it is necessary to study how to construct the feature region efficiently in the input image by verifying whether the selected region is a suitable candidate compared with the target image before comparing the selected region with the target image.

An object of the present invention is to provide an image processing method and apparatus capable of efficiently configuring a feature region in an input image in order to improve a recognition rate of a target in an input image.

According to an embodiment of the present invention, there is provided a method for processing a target image, the method comprising: maintaining a database storing characteristic information of a target image; Receiving an input image; Extracting a partial image from the input image; Selecting a plurality of feature regions from the partial image; And verifying whether or not the selected plurality of feature regions are candidates that can be compared with the target image in consideration of a uniform distribution degree of the plurality of feature regions.

According to an aspect of the present invention, there is provided an image processing apparatus including a database storing characteristic information of a target image; An image input unit for receiving an input image; An image extracting unit for extracting a partial image from the input image; A feature region selection unit for selecting a plurality of feature regions from the partial image; A verifying unit for verifying whether the selected plurality of feature regions are candidates that can be compared with the target image in consideration of a uniform distribution degree of the plurality of feature regions; And a control unit for controlling the database, the image input unit, the extraction unit, and the verification unit.

The image processing method according to an embodiment of the present invention can improve the recognition rate of the target in the input image by verifying the selection suitability of the plurality of feature regions in consideration of the uniform distribution degree of the plurality of feature regions selected from the input image .

1 is a block diagram of an image processing apparatus according to an embodiment of the present invention.
2 is a flowchart illustrating an image processing method according to an embodiment of the present invention.
FIGS. 3 to 4 are diagrams for explaining a method of acquiring a cumulative normalized image among image processing methods related to an embodiment of the present invention.
5 is a diagram for explaining a 2-bit binary form used for feature value comparison among image processing methods according to an embodiment of the present invention.
FIGS. 6 to 7 are diagrams showing the relationship between the correlation coefficient of the cumulative normalized image used for explaining the image processing method according to an embodiment of the present invention and the recognition rate. FIG.

Hereinafter, an image processing method and apparatus according to an embodiment of the present invention will be described with reference to the drawings.

As used herein, the singular forms "a", "an" and "the" include plural referents unless the context clearly dictates otherwise. In this specification, the terms "comprising ", or" comprising ", etc. should not be construed as necessarily including the various elements or steps described in the specification, Or may be further comprised of additional components or steps.

1 is a block diagram of an image processing apparatus according to an embodiment of the present invention.

The image processing apparatus 100 includes a database 110, an image input unit 120, an image extraction unit 130, a feature region selection unit 140, a verification unit 150, a feature value comparison unit 160 And a control unit 170. The control unit 170 includes a control unit 170,

The database 110 stores characteristic values of a target image on which a target object to be detected is captured. The feature value of the target image may be determined through learning.

The image input unit 120 may receive an input image. The input image may include an image to be compared with the feature value of the target image.

The image extracting unit 130 may extract a portion (hereinafter, referred to as 'partial image') from the input image.

The feature region selection unit 140 can select a plurality of feature regions having an arbitrary size and position, which can be overlapped in the partial image. That is, the feature region selection unit 140 can select a plurality of feature regions at random in the partial image.

The verification unit 150 can verify whether the plurality of feature regions selected from the partial image are uniformly distributed. When the selected plurality of feature regions are uniformly distributed, the verifying unit 150 may select the plurality of feature regions as candidates that can be compared with the target image. The verification may be performed using a correlation coefficient of the cumulative normalized image of the partial image. The degree of uniform distribution of the selected plurality of feature regions will be described later.

The feature value comparison unit 160 compares the feature values of the selected plurality of feature regions with the feature values of the target image stored in the database 110 when the selected plurality of feature regions are verified as candidates that can be compared with the target image Compare values and comparisons. A target in the input image can be detected through the feature value comparison. The feature value comparison may be performed using the fern algorithm. The fern algorithm will be described later.

The control unit 170 controls the overall operation of the database 110, the image input unit 120, the image extraction unit 130, the feature region selection unit 140, the verification unit 150, and the feature value comparison unit 160 .

2 is a flowchart illustrating an image processing method according to an embodiment of the present invention.

The database 110 may store characteristic information including characteristic values of the target image (S210). The feature value of the target image may be determined through learning.

The image input unit 120 may receive an input image to be subjected to target object detection (S220).

In this case, the image extracting unit 130 may extract the partial image of the input image for the convenience of detection of the target object (S230). For example, the image extracting unit 130 may extract a partial image while moving in units of one pixel with the same size as the target image.

The feature region selection unit 140 may select a plurality of feature regions from the partial image (S240). The plurality of feature regions may overlap each other and may have any size and position. That is, the feature region selection unit 140 can select a plurality of feature regions at random in the partial image.

The verifying unit 150 can verify whether the selected plurality of feature regions are uniformly configured (S250). This is to prevent a plurality of feature regions selected in the partial image from being shifted to one side to detect a target object with an error. When a plurality of selected feature regions are shifted to one side, there is a risk of missing a feature that distinguishes a target object from a non-target object.

The verification of whether the selected plurality of feature regions are uniformly constructed can be performed using correlation coefficients of the cumulative normalized image. The accumulated normalized image is an image obtained by accumulating a certain value in the same area as each area and normalizing the accumulated value.

3 is a view for explaining a method of acquiring a cumulative normalized image among image processing methods related to an embodiment. In particular, a process of generating a cumulative normalized image according to the number of selected feature regions is shown.

를 생성하는 과정을 단계별로 보여주고 있다.

은 n개의 특징 영역을 가지는 누적 정규화 영상을 나타낸다. 도 3(a)는 임의로 선택된 첫 번째 특징정보 영역에 대한

을 보여준다. 이하, 본 발명의 일실시예에서는 2비트 이진 형태를 특징 정보로 사용하기 때문에 특징 영역이 사각형 모양으로 형성된다. 도 3의 (b)는 임의로 선택된 네 번째까지의 영역에 대한

를 표현하고 있다. 도 3의 (c)는 임의로 선택된 스무 번째까지의 영역에 대한

을 보여준다. 도 3의 (d)는 300개의 특징정보 영역에 대한

을 나타내고 있다. 도 3의 (d)의 누적 정규화 영상을 통하여 특정 지역이 밝을수록 해당 부분의 누적양이 많고, 반대로 특정 지역이 어두울수록 해당 부분의 누적양이 적음을 시각적으로 확인할 수 있다. 만약 누적 정규화 영상에서 전체적인 아닌 지역적으로 밝고 어두운 부분이 다수 존재한다면, 특징 정보를 추출하는 영역 분포가 균일하지 않다는 것을 의미한다.3,

As shown in FIG.

Represents a cumulatively normalized image having n feature regions. FIG. 3 (a) is a graph showing the relationship between

Lt; / RTI > Hereinafter, in the embodiment of the present invention, since the 2-bit binary form is used as the feature information, the feature region is formed in a rectangular shape. FIG. 3 (b) is a graph showing the relationship between the

. FIG. 3 (c) is a graph showing the relationship between

Lt; / RTI > FIG. 3 (d) shows the relationship between

. The accumulative normalized image of FIG. 3 (d) can visually confirm that the cumulative amount of the portion is larger as the specific region is brighter, and conversely, as the specific region is darker, the cumulative amount of the portion is less. If there are many bright and dark parts in the cumulative normalized image as a whole and not in the region, it means that the area distribution for extracting feature information is not uniform.

4 is a view for explaining a method of acquiring a cumulative normalized image among image processing methods according to an embodiment of the present invention.

개의 특징정보 영역으로 생성된 표적의 누적 정규화 영상 t_N과 오 검출 영상의 누적 정규화 영상 e_N을 스케일링한 영상을 나타낸다.FIG.

The cumulative normalized image t _N of the target generated by the feature information areas and the cumulative normalized image e _N of the false detection image are scaled.

4 (a) shows a target image, and Fig. 4 (b) shows a cumulative normalized image t _N of the target image. FIG. 4 (c) shows an erroneously detected image, and FIG. 4 (d) shows a cumulatively normalized image e _N of an erroneously detected image.

The cumulative normalized image I _N is generated for two purposes. First, in order to calculate a correlation coefficient C of a plurality of feature areas included in a partial image to be described later, the second is to visually confirm how uniformly an arbitrary feature information area is constructed. The correlation coefficient C of the plurality of feature regions can be utilized as an index indicating how uniform the plurality of feature regions are. For example, the verification unit 150 can determine that the plurality of feature regions are uniformly distributed as the correlation coefficient C approaches zero.

In order to measure the degree of uniform distribution of the values accumulated in I _N , the correlation coefficient C can be calculated using Equation (1).

영상의 가로, 세로 길이를 의미하고.

와

는 각각 IN의

축 평균값의 위치,

축 평균값의 위치이다. T는 정규화 단위이고,

와

는 각각 IN의

축 표준편차,

축 표준편차를 의미한다. S는 IN의 전체 합이고, C는 상관계수를 의미한다. 검증부(150)는 상관계수 C가 0에 가까울수록 특징정보 영역이 균일하게 분포되었다고 판단할 수 있다. Here, IN denotes a cumulatively normalized image having N feature regions, X and Y are

It means the horizontal and vertical length of the image.

Wow

Respectively,

The position of the axial average value,

It is the position of the axis mean value. T is a normalization unit,

Wow

Respectively,

Axis standard deviation,

Axis standard deviation. S is the total sum of IN, and C is the correlation coefficient. The verification unit 150 can determine that the feature information region is uniformly distributed as the correlation coefficient C approaches zero.

The cumulative normalized image used in FIGS. 4 (b) and 4 (d) accumulates only the bright portion of the 2-bit binary form of the region arbitrarily selected in the partial image. Visually, Figures 4 (a) and 4 (c) are obviously different, but the respective accumulated normalized images (b) and (d) of Figure 4 are similar to each other. As a result, the Fern algorithm that does not perform verification of a plurality of selected feature regions classifies (c) in FIG. 4 into a pattern similar to (a). The reason why such an erroneous classification occurs is that the feature information of the arbitrarily selected region does not include feature information that classifies (a) and (c) in FIG. If the arbitrarily selected regions are not uniformly distributed as shown in FIG. 4, they may not include feature information for classifying the two images. Conversely, if the arbitrarily selected region is uniform, it is highly likely to include feature information for classifying the two images.

Therefore, the recognition rate of the target object can be increased through the verification of the selected plurality of feature regions of the verification unit 150. [ If the plurality of feature regions are uniformly distributed (i.e., the calculated correlation coefficient C is a value close to 0 within a specific range), the verification unit 150 may compare the plurality of feature regions with the feature values of the target image It can be selected as a comparable candidate.

If the plurality of feature regions are not selected as candidates, a process of selecting a plurality of feature regions (S240) and a process of verifying a plurality of feature regions (S250) may be performed again.

When the plurality of feature regions are selected as candidates, the feature value comparing unit 160 compares the feature values of the plurality of feature regions with the feature values of the target image stored in the database 110, The target object included in the partial image can be detected (S260, S270, S280). The feature value comparison may be performed using the Fern algorithm.

The Fern algorithm is an algorithm that divides a specific pattern by constructing a plurality of decision trees with arbitrarily selected simple feature information. The Fern algorithm can be briefly described as follows.

는 특징정보,

는 부류(Class),

는 Fern을 의미하며

개의

로 구성되어 있고, 각각 독립적이다.

는 이진값을 갖는 특징정보로써 수학식 3과 같이 표현될 수 있다.In Equation 2,

The feature information,

Class,

Means Fern

doggy

And are independent of each other.

Can be expressed as Equation (3) as feature information having a binary value.

In Equation (3), I (d _j , 1) and I (d _j , 2) _denote brightness values of a specific region in the partial image.

The Fern algorithm can use a 2-bit binary form as feature information.

5 is a diagram for explaining a 2-bit binary form used for feature value comparison among image processing methods according to an embodiment of the present invention.

The order of calculating the 2-bit binary form is as follows. First, a region having an arbitrary position and an arbitrary size is set in the partial image, and the size of the sum of the brightness values of the upper and lower regions is divided by dividing the half by half, and the brightness values of the left and right regions Is compared. Each result value has a value of 0 or 1. A value of 1 bit is generated in each of the horizontal and vertical directions, and a total of 2 bits is obtained. This result value is used as a 2-bit binary form value. In an embodiment of the present invention, a Fern algorithm using 2-bit binary form as feature information may be used for feature value comparison.

FIGS. 6 to 7 are diagrams showing the relationship between the correlation coefficient of the cumulative normalized image used for explaining the image processing method according to an embodiment of the present invention and the recognition rate. FIG.

FIG. 6 shows recognition rates for images of different sizes and targets of different sizes in order to evaluate the target recognition rate. The experimental environment was constructed as follows. Three sets of images were used, and each image set consisted of 10 consecutive images extracted from the video. The sizes of the image sets are 250 × 335, 180 × 90, and 85 × 70, respectively, and the targets are set to 50 × 67, 44 × 29, and 17 × 14, respectively.

Experiments were performed to detect target by randomly generating 100 feature information for each of three image sets. In the detection process, five candidates with the highest similarity to the target were selected. If the position of each of the five candidates matches the position of the target, it is treated as a correct recognition. If it does not match the position of the target, the recognition accuracy is calculated as a false recognition. The recognition rate is normalized to a value between 0 and 1, and the correlation coefficient is calculated as an absolute value.

FIG. 6 shows the relationship between the absolute value of the correlation coefficient and the recognition rate. In all three sets of images, the absolute value of the correlation coefficient and the recognition rate are inversely proportional to each other.

Fig. 7 (a) shows the detection result when the correlation coefficient is 0.107818, Fig. 7 (c) shows the detection result when the correlation coefficient is 0.032316, 0.00418. ≪ / RTI >

Fig. 7 (b) shows the result of detecting the target with respect to Fig. 7 (a) after configuring any feature information to have different correlation coefficients. Five candidate positions showing the highest degree of similarity in each image are indicated by rectangles. 7, (b), (c), and (d) of FIG. As the correlation coefficient decreases, the target is correctly classified.

As can be seen from the above experiments, it can be confirmed that the target recognition rate increases with the uniform distribution of the plurality of feature regions.

The image processing method described above may be implemented in the form of a program command that can be executed through various computer means and recorded in a computer-readable recording medium. At this time, the computer-readable recording medium may include program commands, data files, data structures, and the like, alone or in combination. On the other hand, the program instructions recorded on the recording medium may be those specially designed and configured for the present invention or may be available to those skilled in the art of computer software.

The computer-readable recording medium includes a magnetic recording medium such as a magnetic medium such as a hard disk, a floppy disk and a magnetic tape, an optical medium such as a CD-ROM and a DVD, a magnetic disk such as a floppy disk, A magneto-optical media, and a hardware device specifically configured to store and execute program instructions such as ROM, RAM, flash memory, and the like.

The recording medium may be a transmission medium, such as a light or metal line, a wave guide, or the like, including a carrier wave for transmitting a signal designating a program command, a data structure, and the like.

The program instructions also include machine language code, such as those generated by the compiler, as well as high-level language code that can be executed by a computer using an interpreter or the like. The hardware devices described above may be configured to operate as one or more software modules to perform the operations of the present invention, and vice versa.

The above-described image processing method and apparatus are not limited to the configuration and method of the embodiments described above, but the embodiments may be modified so that all or some of the embodiments are selectively combined .

100: image processing device
110: Database
120:
130:
140: Feature value extraction unit
150:
160: Feature value comparison unit
170:

Claims (12)

Maintaining a database in which characteristic values of a target image are stored;
Receiving an input image;
Extracting a partial image from the input image;
Selecting a plurality of feature regions from the partial image;
Generating a cumulative normalized image of the partial image using the plurality of feature regions;
Calculating a degree of correlation of the plurality of feature regions using the cumulative normalized image and determining a degree of uniform distribution of the plurality of feature regions based on the calculated degree of correlation; And
And verifying whether or not the selected plurality of feature regions are candidates that can be compared with the target image in consideration of a uniform distribution degree of the plurality of feature regions. delete 2. The method of claim 1,
Is defined as a correlation coefficient expressed by the following equation.
[Equation 1]

(Where I _N denotes a cumulatively normalized image having N feature regions, X and Y are respectively

It means the horizontal and vertical length of the image.

Wow

Lt; RTI ID = _{0.0 &gt};

The position of the axial average value,

It is the position of the axis mean value. T is a normalization unit,

Wow

Lt; RTI ID = _{0.0 &gt};

Axis standard deviation,

Axis standard deviation. S is the total sum of I _N , and C is the correlation coefficient.) 2. The image processing method according to claim 1,
Reselecting a plurality of feature regions in the partial image if the degree of uniform distribution of the plurality of feature regions is not more uniform than a predetermined reference; And
Further comprising the step of verifying whether or not the reselected plurality of feature regions are candidates that can be compared with the target image in consideration of the uniformity of distribution of the reselected plurality of feature regions. 2. The image processing method according to claim 1,
And comparing the feature values of the plurality of feature regions with feature values of the stored target images when the selected plurality of feature regions are verified as candidates. 6. The method of claim 5, wherein the feature value comparison step
Fern algorithm in accordance with an embodiment of the present invention. A database storing feature values of the target image;
An image input unit for receiving an input image;
An image extracting unit for extracting a partial image from the input image;
A feature region selection unit for selecting a plurality of feature regions from the partial image;
Generating a cumulative normalized image of the partial image using the plurality of feature regions, calculating a correlation degree of the plurality of feature regions using the cumulative normalized image, A verifying unit for determining a uniform distribution of regions and verifying whether the selected plurality of feature regions are candidates for comparison with the target image in consideration of a uniform distribution degree of the plurality of feature regions; And
And a control unit for controlling the database, the image input unit, the image extraction unit, and the verification unit. delete 8. The method of claim 7,
Is defined as a correlation coefficient expressed by the following equation.
[Equation 1]

(Where I _N denotes a cumulatively normalized image having N feature regions, X and Y are respectively

It means the horizontal and vertical length of the image.

Wow

Lt; RTI ID = _{0.0 &gt};

The position of the axial average value,

It is the position of the axis mean value. T is a normalization unit,

Wow

Lt; RTI ID = _{0.0 &gt};

Axis standard deviation,

Axis standard deviation. S is the total sum of I _N , and C is the correlation coefficient.) 8. The method of claim 7,
Wherein the feature region selection unit reselects the plurality of feature regions in the partial image when the degree of uniform distribution of the plurality of feature regions is not more uniform than the predetermined reference,
Wherein the verifying unit verifies whether the reselected plurality of feature regions are candidates that can be compared with the target image in consideration of the uniformity of distribution of the reselected plurality of feature regions. The image processing apparatus according to claim 7, wherein the image processing apparatus
Further comprising a feature value comparing unit for comparing a feature value of the plurality of feature regions with a feature value of the stored target image when the plurality of feature regions are verified as candidates. 12. The apparatus of claim 11, wherein the feature value comparison unit
And compares feature values of the plurality of feature regions with feature values of the stored target image using a Fern algorithm.

Images