KR20200067072A - Facility defect inspection method and apparatus - Google Patents

Abstract

Disclosed are a facility defect inspection method and a device thereof. The facility defect inspection method comprises the steps of: (a) receiving two input video sets photographing a specific space including the same facility in the same section at different times; (b) matching the two input video sets; (c) inputting any one of the two matched input video sets into a classification model, and then applying a random forest to detect a detection target facility; and (d) detecting deformation of a region of the detected detection target facility with regard to two matched input videos to inspect a defect in the facility.

Description

Facility defect inspection method and apparatus

The present invention relates to a facility defect detection method and apparatus.

Image processing has been developed with the improvement of computer processing power and the integration technology of image sensors in cameras, and is used in various industries. In security and surveillance systems, image processing has been used for a long time by combining object detection and tracking technologies, and recently it has been used as a means to implement computer vision in the field of battlefields and robots.

The image processing technology for the maintenance of facilities in the railway sector aims to provide reliability and safety by detecting and analyzing defects that are difficult to find by humans. Unlike other means of transportation, the railroad sector relies on tracks and electricity supply facilities, and because the transportation section is very long, maintenance systems for various facilities are very important.

Recently, many studies have been conducted for efficient maintenance and inspection of facilities. Zhang et al. proposed a system for detecting cracks in a tunnel using morphological techniques and machine learning methods from tunnel images taken using a line scan camera. Rizvi et al. proposed a method to detect the presence or absence of track wear from images taken from two cameras. Similarly, Tastimur et al. studied a method for detecting the wear of a track using image processing techniques such as differential and morphological techniques. In Korea, many studies have been conducted for the safety of urban railroads and efficient maintenance of facilities, but the maintenance system of existing railway facilities still has a problem that largely depends on manpower.

The present invention is to provide an image processing-based facility defect detection method and apparatus for efficient railroad facility inspection.

In addition, the present invention is a facility defect detection method and apparatus that can reduce the risk of accidents due to efficient maintenance of technicians by analyzing the captured image using camera equipment installed outside the city railroad vehicle to detect facility deformation. It is to provide.

According to an aspect of the present invention, a method for detecting a facility defect based on image processing for efficient railroad facility inspection is provided.

According to an embodiment of the present invention, (a) receiving two sets of two input images of a specific space containing the same facility in the same section at different times; (b) matching the two input image sets; (c) inputting any one of the matched two input image sets into a classification model, and then applying a random forest to detect a detection target facility; And (d) detecting a defect of the facility by detecting a deformation of the detected area of the facility to be detected using the matched two input images.

The step (b) may include: calculating a phase correlation result of each pixel of the two input image sets; And deriving a maximum value of a phase correlation result between each pixel of the two input image sets as a displacement vector, and matching the two input images using the displacement vector.

In the step (c), the feature information is extracted by applying a histogram of oriented gradient (HOG) algorithm to one of the two input images, respectively; Vectorizing the extracted feature information; And inputting the vectorized feature information into the classification model and applying a random forest algorithm to detect and classify the detection target facility.

Before the step of detecting and classifying the facility to be detected, the method may further include learning the classification model by inputting the vectorized feature information into the classification model and applying a random forest algorithm.

The step (d) may include: (d-1) matching the shape of the detected detection target facility between the two input images by deriving and matching key point feature points of the two input images; And (d-2) determining whether there is a defect using the result of the regional difference of the detection target facilities having the same shape between the two input images.

The step (d-1) may include applying a SURF algorithm to the two input images to extract key point feature points, and matching key point feature points of the two extracted input images, respectively; And estimating a homography matrix using the matched pair of feature points, and then warping the input image captured at a previous time point from the two input images to determine the shape of the detected detection target facility between the two input images. And matching.

The step (d-2) may include deriving a sum of absolute difference (SAD) between facilities to be detected that have the same shape between the two input images; And if the minimum value of the derived SAD results is greater than or equal to a threshold, determining the defect of the facility to be detected.

According to another aspect of the present invention, a facility defect detection device based on image processing for efficient railroad facility inspection is provided.

According to an embodiment of the present invention, an input unit that receives two sets of input images photographing a specific space containing the same facility in the same section at different times; A matching unit that matches the two input image sets; A facility detection unit that detects a facility to be detected by applying a random forest after inputting one of the matched two input image sets into a classification model; And a defect detection unit configured to detect a defect in the facility by detecting a deformation of an area of the detected detection target facility, based on the matched two input images.

The matching unit calculates a phase correlation result of each pixel of the two input image sets, derives a maximum value of a phase correlation result between each pixel of the two input image sets as a displacement vector, and uses the displacement vector By doing so, the two input images can be matched.

The facility detection unit may include a feature extraction unit that extracts feature information by applying a histogram of oriented gradient (HOG) algorithm to either of the two input images; A feature rearranging unit that vectorizes the extracted feature information; And a classifier for detecting and classifying the facilities to be detected by applying a random forest algorithm after inputting the vectorized feature information into the classification model.

The facility detection unit may further include a learning unit for inputting the vectorized feature information into the classification model and learning the classification model by applying a random forest algorithm.

The defect detection unit includes: a matching unit that derives and matches key point feature points of the two input images to match the shape of the detected detection target facility between the two input images; And it may include a defect analysis unit for determining whether or not the defect by using the regional difference results of the detection target facility with the same shape between the two input images.

The matching unit extracts key point feature points by applying a SURF algorithm to the two input images, matches key point feature points of the two extracted input images, and matches the matched pair of feature points. After estimating the homography matrix by using, it is also possible to match the shape of the detected detection target facility between the two input images by warping an input image captured at a previous time point among the two input images.

The defect analysis unit, after deriving a sum of absolute difference (SAD) between the detection target facilities having the same shape between the two input images, if the minimum value of the derived SAD results is greater than or equal to a threshold, it can be determined as a defect of the detection target facility. have.

By providing a method and an apparatus for detecting a facility defect according to an embodiment of the present invention, it is possible to efficiently detect a railroad facility based on image processing.

In addition, the present invention also has the advantage of reducing the risk of accidents due to efficient maintenance of technicians by detecting the deformation of the facility by analyzing the captured image using the camera equipment installed outside the city rail car.

1 is a block diagram schematically showing the internal configuration of a facility defect detection device according to an embodiment of the present invention.
2 is a diagram illustrating a result of matching two input images according to an embodiment of the present invention.
Figure 3 is a block diagram showing the detailed configuration of another facility detection unit according to an embodiment of the present invention.
4 is a diagram showing a facility image used for learning according to an embodiment of the present invention.
5 is a diagram showing facility results detected by a learned classifier according to an embodiment of the present invention.
6 is a block diagram showing the detailed configuration of a defect detection unit according to an embodiment of the present invention.
7 is a view showing a defect detection result according to an embodiment of the present invention.
8 is a flow chart showing a method for detecting defects in facilities according to an embodiment of the present invention.

As used herein, a singular expression includes a plural expression unless the context clearly indicates otherwise. In this specification, the terms "consisting of" or "comprising" should not be construed as including all of the various components, or various steps described in the specification, among which some components or some steps It may not be included, or it should be construed to further include additional components or steps. In addition, terms such as "... unit" and "module" described in the specification mean a unit that processes at least one function or operation, which may be implemented in hardware or software, or a combination of hardware and software. .

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

1 is a block diagram schematically showing the internal configuration of a facility defect detection apparatus according to an embodiment of the present invention, and FIG. 2 is a view showing a result of matching two input images according to an embodiment of the present invention , FIG. 3 is a block diagram showing the detailed configuration of a facility detection unit according to an embodiment of the present invention, FIG. 4 shows a facility image used for learning according to an embodiment of the present invention, and FIG. 5 is a view of the present invention FIG. 6 is a block diagram showing the detailed configuration of a defect detection unit according to an embodiment of the present invention, and FIG. 7 is an embodiment of the present invention. It is a diagram showing the result of defect detection according to an example.

Referring to FIG. 1, the facility defect detection device 100 according to an embodiment of the present invention includes an input unit 110, a matching unit 120, a facility detection unit 130, a defect detection unit 140, and a memory 150 And a processor 160.

The input unit 110 is a means for receiving an input image set.

The input unit 110 receives two sets of input images captured in a specific space, such as a city railroad tunnel, at different times or dates by a camera. Here, the two image sets may be images in which the same facilities are sequentially photographed in the same section.

For example, the first input image of the two input image sets is an image of a specific space including a target facility in a first section at a first time point. In addition, the second input image of the two input images may be an image of a specific space including a target facility in a first section at a second time point.

In one embodiment of the present invention, the two input image sets may be images sequentially photographing a specific space including the same facility at different times. More specifically, one of the two input image sets is a reference image, and may be an image captured before a defect is generated in facilities included in the image.

Further, the other of the two input image sets may be an image captured after the reference image is taken to check whether a defect for facility maintenance occurs after a certain time has elapsed.

The image matching unit 120 is a means for matching two input image sets.

Even if a specific space including the same facility is photographed in the same section due to camera movement speed and shaking, mismatches may occur in the locations of the facilities photographed in each of the two input image sets.

Therefore, the image matching unit 120 may match the two input images using phase correlation.

,

라고 할 때, 두 입력 영상의 위상 상관 결과(R)는 수학식 1과 같이 나타낼 수 있다. For example, each of the two input images

,

In this case, the phase correlation result (R) of the two input images may be expressed as Equation (1).

는 푸리에 변환 함수이며,

는 요소간 곱셈을 나타내며,

는 켤레(conjugates) 신호를 나타낸다. here,

Is the Fourier transform function,

Denotes multiplication between elements,

Indicates a conjugates signal.

)가 계산될 수 있다. In Equation 1, R represents a response of correlation corresponding to each pixel, and a pixel having the largest response represents a moving relationship between two images. Therefore, in Equation 1, the displacement vector of the pixel with the largest response (

) Can be calculated.

This is expressed by Equation (2).

2 is a view showing a result of matching two input images captured by the image matching unit 120.

2(a) are input images photographed at different views, and FIG. 2(b) is a result of matching two input images. As shown in (a) of FIG. 2, mismatches may occur in input images photographed at different viewpoints at different locations of facilities. Therefore, the facility defect detection apparatus 100 according to an embodiment of the present invention may first perform a process of matching two input images for facility defect detection.

The facility detection unit 130 is a means for detecting facilities included in the input image.

The facility detection unit 130 may detect facilities included in the input image and grasp the type and location of the detected facilities.

The facility detection unit 130 described below may detect facilities by using a reference image among two input image sets. In addition, after the facility detection is completed using the reference image, it is natural that the process may be omitted.

3, the detailed configuration of the facility detection unit 130 is shown. Referring to FIG. 3, the facility detection unit 130 according to an embodiment of the present invention includes a feature extraction unit 310, a feature rearrangement unit 320, a classifier 330, and a learning unit 340.

The feature extraction unit 310 is a means for extracting feature information of an object (facilities) to be detected from an input image. Here, the feature extraction unit 310 may extract the feature information of the detection target facility from the input image by applying a histogram of oriented gradient (HOG) algorithm.

The HOG algorithm itself is obvious to those skilled in the art, so a separate description thereof will be omitted. In addition, in one embodiment of the present invention, it is assumed that the feature extraction unit 310 extracts feature information of a detection target facility using a HOG algorithm, but in addition, various known algorithms for extracting feature information of an object from an image It is natural that can be applied.

The feature rearrangement unit 320 is a means for vectorizing feature information of a detection target facility extracted by the feature extraction unit 310.

For example, the feature rearrangement unit 320 may obtain a histogram based on a gradient orientation or gradient magnitude by using feature information of a detection target facility, and then sum each histogram into a single vector. have.

The classifier 330 is a means for detecting and classifying facilities using vectorized feature information.

For example, the classifier 330 may detect and classify multiple detection target facilities by inputting vectorized feature information to the classification model and applying a random forest.

In addition, the classifier 330 may use either one of the two sets of input images in detecting multiple facilities based on a random forest. In the object detection, it is not necessary to use both input images, and when a specific object is detected/classified, based on the object region (location) detected in the subsequent process, two detection target object regions (ie, detection) It is possible to determine whether there is a defect by detecting different parts between target facility areas).

In one embodiment of the present invention, the reason that the classifier 330 detects a multi-detection facility based on a random forest algorithm is that multi-object detection is possible and classification accordingly.

In addition, random forest is a non-linear learning algorithm that improves the decision tree. It can learn multiple decision trees and detect and classify multiple objects using the best result among the learned decision trees when detecting an object. have. The random forest algorithm has an advantage of solving the learning error due to the structural design of the overfitting problem and the greedy algorithm.

Since the two input images are matched through an image matching process, a process of detecting a target facility can be performed by applying a random forest algorithm to only one input image of the two input images. Through the classifier 330, the detection target facility detected by the random forest and the corresponding area can be applied to the corresponding set of input images to perform defect analysis on the detection target facility area rather than the entire two input image areas. There is an advantage.

Since the random forest algorithm is obvious to those skilled in the art, a separate description thereof will be omitted.

In addition, the classifier 330 may remove the location and type of duplicated detection target facilities. For example, the classifier 330 may remove the location and type of duplicate detection target facilities based on non-maximum suppression.

The learning unit 340 is a means for learning the classifier 330. The learning unit 340 may learn the classifier 330 based on the random forest.

4 shows a facility image used by the learning unit, and FIG. 5 is a view showing facility results detected by the learned classifier 330 according to an embodiment of the present invention. 510 represents an insulator, 520 represents a branch current (A1), and 530 represents a branch current (A2). In addition, the solid line indicates the detection result, and the dotted line indicates the undetected facility.

Referring back to FIG. 1, the defect detection unit 140 may analyze deformation/combination based on a stereo matching technique for a detection target facility region among two input images.

6 is a detailed configuration of the defect detection unit 140 according to an embodiment of the present invention. Referring to FIG. 6, the defect detection unit 140 according to an embodiment of the present invention includes a matching unit 610 and a defect analysis unit 620.

The matching unit 610 derives the key point feature points of the two input images and matches them to match the shape of the detection target facility between the two input images.

For example, the matching unit 610 according to an embodiment of the present invention extracts different parts of two input images by applying a stereo matching technique to analyze defects in two input images. To this end, the matching unit 610 extracts key point feature points by applying a SURF algorithm to the two input images, and then matches key point feature points of the two extracted input images, respectively. Subsequently, the matching unit 610 estimates the homography matrix using the matched pair of feature points, and then warps the input image captured at a previous time point among the two input images and detects the detected between the two input images. You can match the shape of the target facility.

The defect analysis unit 620 determines whether there is a defect using the result of the regional difference of the detection target facilities having the same shape between the two input images.

For example, the defect analysis unit 620 differentially extracts two input images, extracts an occlusion candidate region, and calculates a sum of absolute difference (SAD) from the candidate region. Subsequently, the defect analysis unit 620 determines that the minimum value among the brightness values measured by the SAD of the candidate region is greater than or equal to a threshold.

That is, the defect analysis unit 620 may analyze the difference between the two detection target facility regions by performing a regional difference based on the SAD from the two detection target facility regions completely matched by the matching unit 610. The difference between the two detection target facility areas by the SAD can be expressed as Equation 3.

는 키 포인트 특징점 매칭을 통해 생성된 호모그래피 행렬에 의해 변환된 현재 시점의 입력 영상을 나타내며,

는 이전 시점의 입력 영상을 나타낸다. 결함 분석부(620)는 SAD 측정값이 일정한 임계치(T) 이상일 경우(

) 동일 시설물이 촬영된 두 시설물 영상내 결함으로 간주할 수 있다. here,

Denotes an input image of the current viewpoint converted by a homography matrix generated through key point feature point matching,

Indicates the input image from the previous viewpoint. When the SAD measurement value is equal to or greater than a certain threshold (T), the defect analysis unit 620 (

) It can be regarded as a defect in the image of two facilities where the same facility was photographed.

7 is a view showing a defect detection result according to an embodiment of the present invention. 7(a) and (c) show normal images, respectively, and (b) and (d) show defect detection results of the defect simulation image. As shown in FIG. 7, the facility defect detection device according to an embodiment of the present invention exhibits some defect detection errors due to differences in brightness, noise, and shape deformation of the entire facility, but simulated defects It can be confirmed that the part is accurately detected.

The memory 150 stores program codes (commands) required to perform a facility defect detection method using a random forest-based multi-object detection and stereo matching technique according to an embodiment of the present invention and various data derived from this process .

The processor 160 includes internal components (for example, the input unit 110, the matching unit 120, the facility detection unit 130, and the defect detection unit) of the facility defect detection device 100 according to an embodiment of the present invention (140, memory 150, etc.).

8 is a flowchart illustrating a method for detecting defects in facilities according to an embodiment of the present invention.

In step 810, the facility defect detection apparatus 100 receives two input image sets of a specific space including the same facility in the same section at different times.

In step 815, the facility defect detection apparatus 100 matches the two input image sets. For example, the facility defect detection apparatus 100 calculates a phase correlation result of each pixel of the two input image sets, and then derives a maximum value of a phase correlation result between each pixel of the two input image sets as a displacement vector Then, the two input images may be matched using a displacement vector.

In step 820, the facility defect detection device 100 inputs one of the two matched input image sets into the classification model, and then applies a random forest to detect the facility to be detected.

For example, the facility defect detection device 100 may apply a histogram of oriented gradient (HOG) algorithm to any one of the two input images, extract feature information, and vectorize it. Subsequently, the facility defect detection device 100 may input and vectorize feature information into the classification model, and then apply a random forest algorithm to detect and classify multiple detection target facilities.

In step 825, the facility defect detection device 100 analyzes whether deformation/combination is performed based on a stereo matching technique for the detected target facility area among the two input images.

For example, the facility defect detection device 100 may derive and match key point feature points of two input images to match the shape of the detected detection target facility between the two input images. Subsequently, the facility defect detection device 100 may determine whether the defect is based on a result of a regional difference of a detection target facility in which the shapes between two input images are matched.

The apparatus and method according to an embodiment of the present invention may be implemented in a form of program instructions that can be executed through various computer means, and may be recorded on a computer-readable medium. Computer-readable media may include program instructions, data files, data structures, or the like alone or in combination. The program instructions recorded on the computer-readable medium may be specially designed and configured for the present invention or may be known and usable by those skilled in the computer software field. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs, DVDs, and magnetic media such as floptical disks. Includes hardware devices specifically configured to store and execute program instructions such as magneto-optical media and ROM, RAM, flash memory, and the like. Examples of program instructions include high-level language code that can be executed by a computer using an interpreter, etc., as well as machine language codes produced by a compiler.

The hardware device described above may be configured to operate as one or more software modules to perform the operation of the present invention, and vice versa.

So far, the present invention has been focused on the embodiments. Those skilled in the art to which the present invention pertains will understand that the present invention can be implemented in a modified form without departing from the essential characteristics of the present invention. Therefore, the disclosed embodiments should be considered in terms of explanation, not limitation. The scope of the present invention is shown in the claims rather than the foregoing description, and all differences within the equivalent range should be interpreted as being included in the present invention.

100: facility defect detection device
110: input unit
120: registration
130: facility detection unit
140: defect detection unit
150: memory
160: processor

Claims (15)

(a) receiving two input image sets of a specific space containing the same facility in the same section at different times;
(b) matching the two input image sets;
(c) inputting any one of the matched two input image sets into a classification model, and then applying a random forest to detect a detection target facility; And
and (d) detecting a defect of the facility by detecting a deformation of the detected area of the facility to be detected using the matched two input images.
According to claim 1,
Step (b) is,
Calculating a phase correlation result of each pixel of the two input image sets; And
And deriving a maximum value of a phase correlation result between each pixel of the two input image sets as a displacement vector, and matching the two input images using the displacement vector.
According to claim 1,
Step (c) is,
Extracting feature information from each of the two input images by applying a histogram of oriented gradient (HOG) algorithm;
Vectorizing the extracted feature information; And
And inputting the vectorized feature information into the classification model and applying a random forest algorithm to detect and classify the detection target facility.
According to claim 3,
Prior to the step of detecting and classifying the facility to be detected,
And inputting the vectorized feature information into the classification model, and then learning the classification model by applying a random forest algorithm.
According to claim 1,
Step (d) is,
(d-1) deriving and matching key point feature points of the two input images to match the shape of the detected detection target facility between the two input images; And
(d-2) A facility defect detection method comprising determining whether there is a defect using a regional difference result of a detection target facility having a matched shape between the two input images.
The method of claim 5,
Step (d-1) is,
Extracting a key point feature point by applying a SURF algorithm to the two input images, and matching key point feature points of the two extracted input images, respectively; And
After estimating the homography matrix using the matched pair of feature points, the input image photographed at a previous time point among two input images is matched to match the shape of the detected detection target facility between the two input images. A method for detecting a defect in a facility, comprising the steps of making the defect.
The method of claim 5,
Step (d-2) is,
Deriving a sum of absolute difference (SAD) between the detection target facilities having a matched shape between the two input images; And
And if the minimum value of the derived SAD results is greater than or equal to a threshold, determining a defect of the facility to be detected.
A computer-readable recording medium product that records the program code necessary to carry out the method according to claim 1.
An input unit that receives two sets of input images photographing a specific space containing the same facilities in the same section at different times;
A matching unit that matches the two input image sets;
A facility detection unit that detects a facility to be detected by applying a random forest after inputting one of the matched two input image sets into a classification model; And
And a defect detection unit configured to detect a defect in the facility by detecting a deformation of an area of the detected detection facility using the matched two input images.
The method of claim 9,
The matching portion,
After calculating the phase correlation result of each pixel of the two input image sets, the maximum value of the phase correlation result between each pixel of the two input image sets is derived as a displacement vector, and the two vectors are used by using the displacement vector Facility defect detection device characterized by matching the image.
The method of claim 9,
The facility detection unit,
A feature extraction unit that extracts feature information by applying a histogram of oriented gradient (HOG) algorithm to any one of the two input images;
A feature rearranging unit that vectorizes the extracted feature information; And
And a classifier for detecting and classifying the facilities to be detected by applying a random forest algorithm after inputting the vectorized feature information into the classification model.
The method of claim 11,
The facility detection unit,
A facility defect detection device further comprising a learning unit for learning the classification model by applying a random forest algorithm after inputting the vectorized feature information into the classification model.
The method of claim 9,
The defect detection unit,
A matching unit that derives and matches key point feature points of the two input images to match the shape of the detected detection target facility between the two input images; And
And a defect analysis unit configured to determine whether a defect is defective by using a local difference result of a detection target facility having a shape matched between the two input images.
The method of claim 13,
The matching unit,
A key point feature point is extracted by applying a SURF algorithm to the two input images, the key point feature points of the two extracted input images are matched, and homography is performed using the matched pair of feature points. After estimating the matrix, a facility defect detection device characterized by matching the shape of the detected detection target facility between the two input images by warping an input image taken at a previous time point among the two input images.
The method of claim 13,
The defect analysis unit,
A facility defect characterized by determining a sum of absolute difference (SAD) between facilities to be detected in which the shapes between the two input images are matched, and determining that the minimum value of the derived SAD results is a defect of the facilities to be detected. Detection device.

Images