KR101874471B1 - An apparatus, a method, and a computer-readable storage medium for refining object proposal - Google Patents

Abstract

The present invention relates to a method, a computer-readable storage medium, and an apparatus for improving an object search candidate region, and more particularly, to an object search candidate region and a minimum bounding box, To a technique for more accurately estimating an area estimated to be present.

Description

FIELD OF THE INVENTION [0001] The present invention relates to a method for improving an object search candidate region, a computer-readable storage medium, and a computer readable storage medium.

The present invention relates to a method, a computer-readable storage medium, and an apparatus for improving an object search candidate region, and more particularly, to an object search candidate region and a minimum bounding box, To a technique for more accurately estimating an area estimated to be present.

Generally, detecting an object (a person, a specific creature or an object) in an image (image) photographed by a camera generates an object search candidate region from a given image (image) And the process of finding out. Among them, Dense sliding window method and object proposal method are widely used as methods for creating an object search candidate region.

Dense sliding window method is a method to search the whole image given by a plurality of regions having different sizes and shapes by using a window according to a predetermined size and aspect ratio on the corresponding system, Object proposal technique is a method of selectively providing an object search candidate region, which is an area where an object is estimated to exist. Among these two methods, the dense sliding window technique requires a large number of windows to be searched. Therefore, since the data processing amount consumed for object detection can be long and it can take a long time, the latter object proposal technique estimates the object search candidate region, There is a technical need to improve accuracy.

One of the prior art documents with this technical objective is Y. Zhang et al. According to the disclosure in " Improving object detection with deep convolutional networks via Bayesian optimization and structured prediction ", Computer Vision and Pattern Recognition, 2015, pp. 249-258, Based neural network (CNN) -based classifier to a target object in the vicinity of a local optimal value, and searching for a local optimum value is repeated to start an algorithm for reducing an area for an object to be detected have.

However, according to this method, since it is difficult to generalize because it uses a specific object detector, and the size and the aspect ratio of a given object are maintained, there is a limit that there is an error according to the object position.

Y. Zhang et al. "Improving object detection with deep convolutional networks via Bayesian optimization and structured prediction," Computer Vision and Pattern Recognition, 2015, pp. 249-258

An embodiment of the present invention is to provide a technique for improving the ultimate object search candidate region by repeatedly searching and aligning the object search candidate region and the minimum bounding box to solve the limitations of the above- will be.

According to a first aspect of the present invention, a method for detecting an object from an image is disclosed. The method includes the steps of: (a) creating a primary object search candidate region by selecting a portion of a plurality of segmented images for an input image; (b) calculating an area of the primary object search candidate region; (c) selecting one or more inner regions associated with the primary object search candidate region; And (d) updating the primary object search candidate region to a secondary object search candidate region based on the width of the primary object search candidate region and the one or more inner regions. The step (d) (updating the object search candidate region) includes: selecting a first minimum bounding box defined by the one or more inner regions; And updating the primary object search candidate region to the secondary object search candidate region so that the difference between the primary object search candidate region and the primary minimum bounding box is minimized.

Meanwhile, the method according to the first aspect may repeat the steps (b) to (d) until the i-th object search candidate region coincides with the object search candidate region of the next rank Is an arbitrary positive integer that is equal to or smaller than N, and N is a maximum difference of the object search candidates.

In the method according to the first aspect, in the step (d) (the step of updating the object search candidate region), when all of the one or more inner regions are included in the i-th minimum boundary box, A case in which at least one inner region of the at least one inner region is located over the i-th object search candidate region, and a case where one inner region of the at least one inner region is to be excluded.

에 의한 LI 지표를 사용하여 객체 검색 후보 영역이 갱신될 수 있다. First, if all of the one or more inner regions are included in the i-th minimum bounding box,

The object search candidate region can be updated by using the LI indicator by the LI indicator.

에 의한 OMR 지표를 사용하여 객체 검색 후보 영역이 갱신될 수 있다.Secondly, if at least one inner region of the one or more inner regions spans the i-th object search candidate region,

The object search candidate region can be updated by using the OMR index by < RTI ID = 0.0 >

에 의한 OLR 지표를 사용하여 객체 검색 후보 영역이 갱신될 수 있다.Thirdly, if the inner area of one of the one or more inner areas should be excluded,

The object search candidate region can be updated using the OLR index by the OLR index.

According to a second aspect of the present invention, there is provided a computer-readable storage medium having recorded thereon a computer program for improving an object search candidate region from an image. The computer program comprising: instructions for generating a primary object search candidate region by selecting a portion of a plurality of segmented images for an input image; From the primary object search candidate region to the object search candidate region of the trailing order sequentially until the object search candidate region of the current order matches the object search candidate region of the next order.

In the computer program according to the second aspect, the instruction for updating includes: an instruction for calculating an area of the i-th object search candidate region; Selecting one or more inner regions associated with the i-th object search candidate region; And an instruction to update the i-th object search candidate region to an i + 1-th object search candidate region based on the width of the i-th object search candidate region and the one or more inner regions i is any positive integer equal to or less than N, and N is a maximum difference number of the object search candidate region)

According to a third aspect of the present invention, an apparatus for improving an object search candidate region is disclosed. The apparatus comprises: a primary object search candidate region generation module for selecting a primary object search candidate region by selecting a part of a plurality of segmented images for an input image; And an object search candidate region updating module for updating the object search candidate region from the primary object search candidate region to an object search candidate region of a trailing order sequentially until an object search candidate region of a current order matches an object search candidate region of a next order, .

The apparatus further includes an image selection module for selecting one or more interior regions associated with the i-th object search candidate region, where i is any positive integer less than or equal to N, and N is the maximum difference of the object search candidate regions ).

The object search candidate region update module includes an near field boundary calculation unit for calculating an i < th > order minimum bounding box defined by one or more inner regions associated with the i < th > And an area width calculator for calculating the i-th object search candidate area and the width of the i-th minimum boundary box.

According to the method for improving an object search candidate region according to an embodiment of the present invention, greedy search is performed on a nearest neighbor region boundary of a given object search candidate region, so that it is not limited by the size and the aspect ratio of the candidate region, Can be improved.

In addition, according to the object search candidate region improvement method according to an embodiment of the present invention, the object search candidate region is formed according to the near boundary search by defining the minimum bounding box, have.

It will be understood by those skilled in the art that the effects of the present invention may be obtained without departing from the scope of the present invention.

1 shows a functional block diagram of an apparatus for improving an object search candidate region according to the present invention.
2 shows a schematic flow diagram of a method for improving an object search candidate region according to the present invention.
Figure 3 shows a specific flowchart of a method that takes into account the boundary area alignment measure used in the technique according to an embodiment of the present invention.
4 is a diagram for explaining an aspect in which an object search candidate region is updated by a method according to an embodiment of the present invention.
5 is a diagram for explaining an exemplary aspect in which a technique according to the present invention is applied to an arbitrary image.

Specific structural and functional descriptions of embodiments of the invention disclosed herein are merely illustrative for purposes of illustrating embodiments of the invention and that the embodiments according to the invention may be embodied in various forms, And should not be construed as limited to the embodiments described in the application.

Since the embodiments according to the present invention can make various changes and have various forms, specific embodiments are illustrated in the drawings and described in detail herein. It should be understood, however, that the embodiments according to the concepts of the present invention are not intended to be limited to any particular mode of disclosure, but rather all variations, equivalents, and alternatives falling within the spirit and scope of the present invention.

The first and / or second term may be used to describe various components, but the components should not be limited by the terms. The terms are intended to distinguish one element from another, for example, without departing from the scope of the invention in accordance with the concepts of the present invention, the first element may be termed the second element, The second component may also be referred to as a first component.

The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular expressions include plural expressions unless the context clearly dictates otherwise. In this specification, terms such as "comprise," "include," "have," and the like, specify that there is a specified feature, number, step, operation, component, section, element, Steps, operations, elements, parts, or combinations thereof, whether or not explicitly described or implied by the accompanying claims.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. Terms such as those defined in commonly used dictionaries should be construed as meaning consistent with meaning in the context of the relevant art and are not to be construed as ideal or overly formal in meaning unless expressly defined herein .

Hereinafter, an object detection technique according to the present invention will be described in detail with reference to the accompanying drawings.

1 shows a functional block diagram of an apparatus for improving an object search candidate region according to the present invention. The object search candidate region enhancement apparatus 1 includes a primary object search candidate region generation module 10, an image selection module 20, and an object search candidate region update module 30.

The primary object search candidate region generation module 10 generates an initial object search candidate region including at least 50% of the object of interest, and performs a plurality of segmentation processing on the image input or captured by the image shooting / The segmented image can be generated without any limitation on size and aspect ratio. The image selection module 20 selects an image for updating an object search candidate region among a plurality of segmented images. Then, the object search candidate region update module 30 sequentially searches the primary (initial) object search candidate region for a next-order object search candidate region such as a second order, a third order, Until the object search candidate region of the next order matches the search region of the next order.

On the other hand, the image selection module 20 may be further capable of selecting one or more inner regions associated with the i-th object search candidate region, where i is any positive integer less than or equal to N and N is the maximum Car number). For example, one or more inner regions associated with the primary object search candidate region may be selected in the primary object search candidate region, and one or more inner regions associated with the object search candidate region may be selected in the posterior search region.

The object search candidate region update module 30 further includes a near region boundary calculation unit 31 and a region size calculation unit 32. [ The closest region boundary calculator 31 calculates a minimum bounding box defined by the inner regions selected by the image selecting module 20. [ For example, the first minimum bounding box is calculated by the inner regions selected in the primary object search candidate region, and the same applies to the inner regions selected in the object search candidate region of the posterior order. Meanwhile, the area size calculating unit 32 calculates the widths of the object search candidate area and the minimum bounding box. In the technique according to the present invention, since all areas are defined as the menopause, the area width can be calculated.

FIG. 2 shows a flowchart of a method for improving an object search candidate region according to the present invention, and a method considering the boundary region sorting measure for this method is specifically shown in FIG.

Referring to FIG. 2, which is a general flowchart, a primary object candidate region is created from an input image (Step S10), and an image for updating an object candidate region among a plurality of segment images is selected (Step S20). Then, the object search candidate region is re-initialized by the proximity region boundary search according to the present invention (Step S30), and the object boundary candidate region is sequentially updated by aligning the object boundary region (Step S40) An object search candidate region whose localization is improved is obtained (S50).

A method of improving the object search candidate region will be described in more detail with reference to FIG. In step S11, a primary object search candidate region (initial object search candidate region) is created by selecting a part of a plurality of segmented images with respect to the input image (S11). Wherein a plurality of segmented images may be generated without limitation on size and aspect ratio.

Next, in step S12, the area of the primary object search candidate area is calculated. In step S13, one or more internal areas associated with the primary object search candidate area are selected. In step S14, (Define) the primary minimum bounding box defined by Next, in step S15, the primary object search candidate region is updated to the secondary object search candidate region so that the difference between the primary object search candidate region and the primary minimum boundary box is minimized.

In step S13 to step S15, in order to update the object search candidate region according to the relation between the object search candidate region, the at least one inner region, and the minimum bounding box, that is, the difference between the object search candidate region and the minimum bounding box has the minimum value, the metric may be applied differently in some cases, and this will be described in detail with reference to FIG.

After the process is updated to the secondary object search candidate region according to the above process, the process returns to step S12, and until the i-th object search candidate region matches the object search candidate region of the next order (i.e., Is converged), steps S12 to S15 are repeatedly performed.

Although not shown in FIG. 3, in the process of updating from the i-th object search candidate region to the i + 1-th object search candidate region (i.e., from S12 to S15), the i- , One or more segmented images associated with the i-th object search candidate region may be selected.

An image of an exemplary aspect described above, in which a technique according to an embodiment of the present invention is applied to any image, is illustrated in Fig. Assume that an area indicated by initial and refined in an exemplary image is an object to be detected. An area indicated by 'initial' refers to a primary object search candidate area, and 'refined' Indicates a convergent object search candidate region searched according to the technique according to the present invention.

This will be described with reference to Fig. 4 (a). First, the minimum bounding box is obtained by using all the regions within the initial object search candidate. In this case, it is calculated as Equation (1) using (the area of the minimum bounding box defined by all inner regions) / (the area of the object search candidate).

는 i차 객체 검색후보

의 내부 영역의 세트를 표시하고,

는 임의의 영역 또는 임의의 박스 X의 면적을 표시하고,

는 영역들의 세트 χ에 의해 정의되는 최소 경계 상자의 면적을 표시하며, N은 객체 검색후보의 수를 표시한다.

Is an i-th object search candidate

A set of inner areas of the display area,

Represents an area of an arbitrary area or an arbitrary box X,

Denotes the area of the minimum bounding box defined by the set of regions x, and N denotes the number of object search candidates.

Next, in FIG. 4 (b), one of the straddling regions of the object search candidates and all the internal regions are obtained. (the area of the minimum bounding box defined by the inner regions having one spreading region) / (the area of the object searching candidate) similar to that described with reference to Fig. .

는 걸친 영역의 세트

에서 j번째 요소(영역)를 표시하고, M은 i번째 객체 검색후보 b_i의 걸친 영역의 수를 표시한다.

A set of spanned regions

(Area), and M denotes the number of regions of the i-th object search candidate b _i .

Lastly, in (c) of FIG. 4, the minimum bounding box is obtained using the inner regions except for one of all the inner regions in the object search candidate. That is, the calculation is performed using Equation (3) using the area of the minimum bounding box defined by the inner region excluding only one of all inner regions / (the area of the object search candidate).

은 내부 영역의 세트

에서 k번째 요소(영역)을 표시하고, P는 객체 검색후보 b_i의 내부 영역들의 수를 표시한다.

Lt; RTI ID = 0.0 >

(Region), and P denotes the number of inner regions of the object search candidate b _i .

Note that since the LI score in Equation (1) is estimated using all the inner regions of each object search candidate, a single LI score is obtained for each object search candidate. On the other hand, since there may be a large number of inner regions and overlapping regions in the object search candidate, a plurality of OMR or OLR scores (according to Equations 2 and 3) are obtained for each object search candidate. Since each measure expresses the difference between the initial object search candidate and the updated object search candidate, the nearest region boundary can be searched by selecting the minimum score among the LI, OMR, and OLR scores.

Object search candidates re-initialize

It may not match the region boundaries (in particular, the region boundaries generated by the objectivity-based method described above), even though the initial object search candidates are appropriate. Since objects are defined by the lung boundaries, the initial object search candidates are aligned to the nearest region boundaries.

The OLR score is always larger than the LI score, so that the OLR score matches the object search candidate to the nearest area boundary. It is not used for. Therefore, LI and OMR are used in this step.

에서 각각의 객체 검색후보에 대하여 영역 세트

로부터 내부 영역의 세트

및 걸친 영역의 세트

를 구한다. 그리고, 경계-정렬된 객체 검색후보에 대한 LI 스코어가 0이므로, 재-초기화에 필요한 객체 검색후보를 식별하기 위해 LI 측정을 사용한다. 그러므로 LI 스코어를 산출하고, 이에 대응하는 최소 경계 상자 b_LI를 초기 객체 검색후보로부터 구한다. 그리고 LI가 0이 아닌 객체 검색후보에 대한 걸친 영역의 세트

를 사용하여 OMR 스코어를 산출한다. 모든 스코어가 각각의 객체 검색후보에 대하여 산출되면, 산출된 스코어 중 최소값을 갖는 최소 경계 상자를 선택한다. 이러한 방식으로, 영역 경계-정렬 객체 검색 후보 세트

를 구한다.Since the LI and OMR scores are calculated using the inner region and the overlap region of the object search candidates,

For each object search candidate in the region set

Lt; RTI ID = 0.0 >

And a set of span areas

. Then, since the LI score for the boundary-aligned object search candidates is zero, the LI measurement is used to identify the object search candidates required for re-initialization. Therefore, the LI score is calculated and the corresponding minimum bounding box b _LI is obtained from the initial object search candidates. And a set of regions for object search candidates for which LI is non-zero

To calculate the OMR score. When all scores are calculated for each object search candidate, a minimum bounding box having a minimum value among the calculated scores is selected. In this manner, the region boundary-aligned object search candidate set

.

Greedy boundary - area alignment

로부터 가장 가까운 영역 경계를 그리디하게 검색한다. 재-초기화 단계에서와 같이, 내부 영역의 세트

및 걸친 영역의 세트

를 사용하여 가장 가까운 경계 영역들을 찾는다. 재초기화된 객체 검색후보는 영역 경계와 이미 정렬되었기 때문에, LI 스코어는 항상 0이다. 그러므로 OLR과 OMR 측정치를 사용하여 가장 가까운 경계 영역을 검색한다.To improve localization accuracy, the region boundary-aligned object search candidates

The region boundary closest to the target region is searched. As in the re-initialization step,

And a set of span areas

To find the nearest boundary regions. Since the re-initialized object search candidates are already aligned with the region boundaries, the LI score is always zero. Therefore, OLR and OMR measurements are used to search for the nearest boundary area.

에서 각각의 재초기화된 객체 검색후보에 대하여 영역 세트

로부터 내부 영역의 세트

및 걸친 영역의 세트

를 구한다. 그리고, 각각의 객체 검색후보에 대해 OLR 및 OMR 스코어를 추정하고, 추정된 OLR 및 OMR 스코어 중에 최소 스코어를 갖는 최소 경계 상자를 선택한다. 여기서, 검색 범위를 제어하기 위한 중첩 임계치 δ와 최대 반복

이 도입된다. 객체 검색후보를 갱신할 때마다 내부 영역의 세트

및 걸친 영역의 세트

가 변하기 때문에, 매 반복에서 차순위로 가장 가까운 영역 경계를 검색하기 위하여, 갱신된 객체 검색후보의 OLR 스코어 및 OMR 스코어를 추정할 것이다. 업데이트된 객체 검색후보가 수렴될 때까지 이러한 과정이 반복된다.first,

For each re-initialized object search candidate in the region set

Lt; RTI ID = 0.0 >

And a set of span areas

. Then, OLR and OMR scores are estimated for each object search candidate, and a minimum bounding box having a minimum score among the estimated OLR and OMR scores is selected. Here, the superposition threshold value < RTI ID = 0.0 >#< / RTI &

. Each time you update an object search candidate, a set of internal regions

And a set of span areas

The OLR score and the OMR score of the updated object search candidate will be estimated in order to search for the closest region boundary to the next in each iteration. This process is repeated until the updated object search candidates are converged.

값은 10으로 설정되었지만 통상적으로 7번의 반복 내에 수렴됨을 확인하였다.Experimentally, since the initial object search candidates are assumed to overlap by at least 50%, the overlay threshold [delta] is set to 0.5 and the maximum repeat

Value was set to 10, but it was confirmed that it converged normally within 7 repetitions.

In order to evaluate the quality of the object search candidate set according to the present invention, the Histogram of Best Overlap (HBO) metric was used. The HBO metric represents the distribution of object search candidates that are maximally overlapped with the ground-truth according to the overlap threshold. This is defined as (area based on intersection between actual and object search candidates) / (area based on the union of actual and object search candidates) and calculated as Equation (4).

는 실측을 표시하고,

는 객체 검색후보를 표시한다. IoU 값에 따른 객체 검색후보의 로컬화 품질이 그림 5와 같이 표시된다. 그리고 객체 검색후보의 전체적인 성능을 평가하기 위해 ABO(Average Best Overlap) 메트릭이 수학식 5과 같이 계산된다.In Equation 4,

Lt; / RTI >

Indicates an object search candidate. The localization quality of the object search candidates according to the IoU value is displayed as shown in Fig. And an ABO (Average Best Overlap) metric is calculated as shown in Equation (5) in order to evaluate the overall performance of the object search candidate.

는 i번째 실측을 표시하고, N은 실측의 수를 표시하고,

는 객체 검색후보 세트를 표시한다. 이러한 IoU 및 ABO 메트릭에 의한 시뮬레이션 결과, 관심객체 중 35%는 로컬화 정확도(IoU) 0.9 내지 1.0으로 포함하고 있음을 확인할 수 있었다.In Equation (5)

Denotes the i-th actual measurement, N denotes the actual measurement number,

Indicates a set of object search candidates. As a result of the simulation using the IoU and ABO metrics, it was confirmed that 35% of the objects of interest include the localization accuracy (IoU) of 0.9 to 1.0.

As described above, according to the object search candidate region improving method according to an embodiment of the present invention, greedy search is performed on the nearest region boundary of a given object search candidate region, so that it is not limited by the size and the aspect ratio of the candidate region The error due to localization can be improved. In addition, the localization performance can be improved by defining the minimum bounding box and performing the object search candidate region according to the nearest neighbor boundary search.

It should be understood that the example modules, logic blocks, means, steps or combinations thereof in connection with the embodiments described herein may be implemented as electronic hardware (digital designs designed by coding or the like), software (various types of applications including program instructions) Or a combination of these. Hardware, and / or software may vary depending on design constraints imposed on the user terminal.

In some implementations, one or more of the configurations described herein may be stored in memory as computer program instructions, which may execute the methods described herein, centering on a digital signal processor. It is to be understood that the connections between the components specified with reference to the drawings attached hereto are merely illustrative and at least some of them may be omitted or, conversely, may include additional components as well as these components.

It should be understood that the example modules, logic blocks, means, steps or combinations thereof in connection with the embodiments described herein may be implemented as electronic hardware (digital designs designed by coding or the like), software (various types of applications including program instructions) Or a combination of these. Hardware, and / or software may vary depending on design constraints imposed on the user terminal.

The foregoing description is merely illustrative of the technical idea of the present invention, and various changes and modifications may be made by those skilled in the art without departing from the essential characteristics of the present invention. Therefore, the embodiments disclosed in the present invention are intended to illustrate rather than limit the scope of the present invention, and the scope of the technical idea of the present invention is not limited by these embodiments. The scope of protection of the present invention should be construed according to the following claims, and all technical ideas within the scope of equivalents should be construed as falling within the scope of the present invention.

1: Improvement device of object search candidate region
10: Segment image generation module
20: Image selection module
30: object search candidate region update module
31: Near Area Boundary Calculator
32: area width calculation unit

Claims (11)

(a) generating a primary object search candidate region by selecting some of a plurality of segmented images for an input image;
(b) selecting one or more inner regions associated with the primary object search candidate region;
(c) selecting a primary minimum bounding box defined by the one or more inner regions; And
(d) updating the primary object search candidate region to a secondary object search candidate region so that the difference between the primary object search candidate region and the primary minimum bounding box is minimized;
Wherein the object search candidate region is an object search candidate region.
delete The method according to claim 1,
performing the steps (b) to (d) repeatedly until the i-th object search candidate region coincides with the object search candidate region of the next rank,
I is an arbitrary positive integer that is equal to or smaller than N, and N is a maximum degree of the object search candidate region.
A method for improving an object search candidate region.
The method of claim 3,
Wherein when the i-th object search candidate region includes all of the one or more inner regions, the updating to the i + 1 < st > -order object search candidate region uses the LI index according to Equation (1)
[Equation 1]

(

: i-th object search candidate region,

:

A set of inner regions included in <

: Area for arbitrary area X,

: The width of the minimum bounding box defined by the set of arbitrary regions x)
A method for improving an object search candidate region.
The method of claim 3,
The step of updating to the i + 1 < nd > -order object search candidate region when at least one inner region of the one or more inner regions spans the i-th object search candidate region includes using an OMR index according to Equation (2) ≪ / RTI >
&Quot; (2) "

(

: i-th object search candidate region,

:

A set of inner regions included in <

: Area for arbitrary area X,

: The width of the minimum bounding box defined by the set of arbitrary regions x,

J is the number of inner regions in the set of inner regions spanning the i-th object search candidate region, M is the number of inner regions spanning the i-th object search candidate region, j is the number of inner regions Any positive integer less than or equal to the number of regions)
A method for improving an object search candidate region.
The method of claim 3,
Wherein the step of updating to the i + 1 < st > -order object search candidate region uses an OLR index according to Equation (3) if the inner region of one of the one or more inner regions is to be excluded.
&Quot; (3) "

(

: i-th object search candidate region,

:

A set of inner regions included in <

: Area for arbitrary area X,

: The width of the minimum bounding box defined by the set of arbitrary regions x,

: remind

K is an arbitrary positive integer less than or equal to the number of inner regions of the i-th object search candidate region,
A method for improving an object search candidate region.
1. A computer-readable storage medium having recorded thereon a computer program for improving an object search candidate region from an image,
The computer program comprising:
(a) instructions for generating a primary object search candidate region by selecting some of a plurality of segmented images for an input image;
(b) selecting one or more inner regions associated with the primary object search candidate region;
(c) selecting a primary minimum bounding box defined by the one or more inner regions; And
and (d) updating the primary object search candidate region to a secondary object search candidate region so that a difference between the primary object search candidate region and the primary minimum bounding box is minimized.
Computer-readable storage medium.
8. The method of claim 7,
After the command to update,
(b) to (d) are repeated in order until the i-th object search candidate region coincides with the object search candidate region of the next order,
I is an arbitrary positive integer that is equal to or smaller than N, and N is a maximum degree of the object search candidate region.
Computer-readable storage medium.
A primary object search candidate region generation module for selecting a primary object search candidate region by selecting some of a plurality of segmented images with respect to an input image; And
An object search candidate region update module for updating an object search candidate region of the current order to an object search candidate region of a next order sequentially from the primary object search candidate region to an object search candidate region of a next order, Including,
The object search candidate region update module includes:
and a near-field boundary calculator for calculating an i-th ordered minimum boundary box defined by one or more inner regions associated with the i-th object search candidate region,
Where i is a positive integer equal to or less than N, and N is a maximum degree of the object search candidate region,
An apparatus for improving an object search candidate region.
10. The method of claim 9,
Further comprising an image selection module for selecting one or more interior regions associated with the i < th > object search candidate region,
An apparatus for improving an object search candidate region.
10. The method of claim 9,
The object search candidate region update module includes:
Order object search candidate region and an area width calculating section for calculating an area of the i < th > order object search candidate region and the i &
An apparatus for improving an object search candidate region.

Images