JP2019003575A - Image analysis device and maintenance management method of railroad facility component - Google Patents

Abstract

To provide an image analysis device capable of eliminating influence of luminance change and calculating a feature amount at which an image feature is appropriately reflected with a small load.SOLUTION: This image analysis device is provided with a feature amount calculation part for calculating a feature amount of a local area (K11) of an image. The feature amount calculation part includes a gradient direction calculation part for calculating a gradient direction obtained by eliminating size information from a luminance gradient between each of a plurality of pixels included in the local area (K11) and peripheral pixels, and a distribution calculation part for aggregating a plurality of gradient directions which are respectively calculated in correspondence with the plurality of pixels to calculate a frequency distribution of the gradient directions, and determines the frequency distribution of the gradient directions to be a feature amount of the local area.SELECTED DRAWING: Figure 8

Description

  The present invention relates to an image analysis apparatus for analyzing image feature amounts and a maintenance management method for railway equipment parts using the image analysis apparatus.

  The railroad track is provided with a track circuit that uses the rail as a part of the signal path in order to detect the presence or absence of the railway vehicle. The track circuit includes a detection circuit that inputs and outputs a signal current for detecting a railway vehicle to and from the rail, and also separates a signal current for vehicle detection from other currents such as a return current. A circuit called an impedance bond or a filter circuit is connected to the. And in order to electrically connect each of these circuits and the circuit of the rail part, the connection part called "feeding bond" is provided in the side surface of the rail. A connection part points out the site | part which fixed conducting wires, such as a strand wire, to the rail by copper thermite welding or brazing.

  If the connection part of the track circuit is damaged, disconnected or dropped, it cannot be correctly determined whether the railway vehicle exists on the track. For this reason, in the past, railway companies have regularly inspected a large number of connection portions provided on the rails by visual confirmation or the like for abnormalities.

As a technique related to the present invention, conventionally, there is a technique for performing various image analysis by calculating a feature amount called HOG (Histograms of Oriented Gradients) of an image. Patent Document 1 discloses a technique for determining the similarity between two images by dividing two images into a plurality of local regions and comparing the HOGs of the local regions.
HOG is an amount of a plurality of luminance gradient vectors in a local region of an image that is histogrammed with respect to the gradient direction. Specifically, the luminance gradient vector includes luminance gradient magnitude information and direction information, and HOG corresponds to an amount of accumulated luminance gradient vector magnitudes for each gradient direction. The HOG includes information on the magnitude of the luminance gradient, but is known as a feature quantity that is not easily affected by illumination fluctuations or shadows.

JP 2011-96081 A

  Here, a technique for realizing the inspection of the connection portion of the track circuit more accurately while using the image comparison processing of the computer while reducing the labor is considered. For example, images are periodically taken for a large number of connection parts of the track circuit, and a plurality of images taken for the same connection part over a number of days are compared with each other. And the connection part in which some change has arisen among many connection parts is narrowed down. Further, the inspector performs a detailed inspection on the connecting portion where the change occurs, and determines whether or not repair is necessary. With the support of such a computer, the number of connections that the inspector can inspect in detail can be greatly reduced, and a reduction in labor can be expected.

  Since the connection part of the track circuit is often outdoors, even when the same connection part is photographed in the same way at different dates and times, the brightness of the image varies due to the influence of the weather or the like. Therefore, it has been expected that by comparing images using HOG that is not easily affected by illumination fluctuations, it is possible to detect the presence or absence of changes in the connection portion by eliminating the influence of luminance variations.

  However, in the inspection of track circuit connections, it is necessary to detect minute changes such as small cracks, small cracks, partial breaks in stranded wires, and displacements of the conductors that induce them. . On the other hand, although HOG can reduce the influence of luminance variations, it cannot completely eliminate the influence. For this reason, when image comparison is performed using HOG, a small change in the connection portion is hidden by noise of luminance variation, and the presence or absence of the fine change in the connection portion may not be sufficiently detected.

  In order to avoid such a problem, it has been considered to perform image comparison using HOG after performing processing for correcting the luminance variation. However, simple processing such as simply equalizing the average value of luminance cannot significantly reduce the influence of luminance variation included in the HOG. For example, the calculation load of aligning the luminance frequency distribution in the entire comparison portion of the image is reduced. A large amount of processing was considered necessary. However, since there are a large number of track circuit connections and the number of image comparisons is enormous, there has been a demand to reduce the calculation load as much as possible.

  The present invention provides an image analysis apparatus that can largely eliminate the influence of luminance change as compared with HOG, and that can calculate a feature amount that appropriately reflects the feature of an image, such as HOG, with a small load. Objective. Another object of the present invention is to provide a maintenance management method for railway equipment parts that can significantly reduce the maintenance work of parts by using the image analysis apparatus.

In order to achieve the above object, an image analyzer according to the present invention provides:
A feature amount calculation unit that calculates the feature amount of the local region of the image,
The feature amount calculation unit includes:
A gradient direction calculation unit that calculates a gradient direction in which the magnitude information is removed from the luminance gradient between the plurality of pixels included in the local region and the surrounding pixels;
A distribution calculator that calculates a frequency distribution in the gradient direction by aggregating the plurality of gradient directions calculated respectively corresponding to the plurality of pixels;
Including
The frequency distribution in the gradient direction is used as a feature amount of the local region.

  According to this configuration, since the frequency distribution in the gradient direction in which the size information is removed from the luminance gradient is obtained as the feature amount of the local region, the feature amount in which the influence of the luminance change of the image is greatly eliminated is obtained. It is done. Furthermore, in order to calculate this feature amount, the feature amount can be obtained by calculating the frequency distribution by removing the size information from the luminance gradient including the size information and the direction information. Feature quantities can be calculated. For example, compared to the case where the calculation of HOG is performed after the correction calculation for aligning the luminance frequency distribution, the feature quantity that can eliminate the influence of the luminance change as compared with the HOG can be calculated with a smaller load than the calculation of the HOG.

Here, the image analysis apparatus according to the present invention is:
An image input unit for inputting a plurality of images taken of the inspection object;
An image comparison unit for comparing the plurality of images;
The image comparison unit
An image dividing unit that divides at least a portion including the inspection object into a plurality of local regions for each of the plurality of images;
The image processing apparatus may further include a change detection unit that detects the change in the inspection object by comparing the feature amounts of the plurality of local regions calculated by the feature amount calculation unit between the plurality of images.

  According to this configuration, even if there is a luminance variation between a plurality of images, it is possible to calculate the feature amount of the inspection target excluding the luminance variation, and to detect a change in the inspection target by comparing the feature amount. Accordingly, it is possible to detect the fine change of the inspection object precisely and accurately by eliminating the influence of the luminance variation between the plurality of images.

Furthermore, in the image analysis apparatus according to the present invention,
Three or more images obtained by photographing the inspection object are input to the image input unit,
The image comparison unit performs comparison processing of the feature amount between the comparison target and the plurality of comparison references, with a plurality of the three or more images as a comparison reference and at least one as a comparison target, The presence / absence of a change in the inspection object may be determined based on a result of a plurality of comparison processes each using a plurality of comparison criteria.

  According to this configuration, a plurality of comparison processes are performed using a plurality of comparison criteria, and the presence or absence of a change in the inspection object is determined based on these comparison results. Therefore, for example, even when an exceptional difference is included in one of the comparison criteria, it is possible to prevent an error from occurring in the determination of whether or not the inspection object has changed due to this exceptional difference. Therefore, the presence or absence of a change can be detected more accurately.

Furthermore, the image analysis apparatus according to the present invention is
A mask processing unit that masks at least a part of a background region of the inspection object may be further included at least before the change detection unit compares the feature amounts.
According to this configuration, it is possible to intensively detect changes occurring in the region of the inspection object by using the mask, and further reduce the calculation load by reducing the local region for calculating the feature amount.

In addition, the maintenance management method of railway equipment parts according to the present invention,
A maintenance management method for railroad equipment parts that manages and manages each of a plurality of parts connected to railroad rails using the image analysis apparatus described above,
With the imaging device mounted on the railway vehicle traveling on the rail, the plurality of parts are imaged as the inspection object several times over a period of days,
The image analyzer detects a change in each of the plurality of parts by comparing a plurality of images obtained by the plurality of times of photographing,
Of the plurality of parts, it is determined whether or not repair is necessary for one or a plurality of parts in which a change equal to or greater than a threshold is detected.

  According to this maintenance management method, for a plurality of parts connected to railroad rails, changes taken every day are accurately and accurately detected using images taken from railcars. The target to be judged can be narrowed down. Accordingly, it is possible to inspect a large number of parts existing on railroad rails more strictly while significantly reducing labor.

  According to the image analysis apparatus of the present invention, it is possible to greatly eliminate the influence of luminance change as compared with HOG, and it is possible to calculate a feature amount that can appropriately reflect the feature of an image like HOG with a small load. . Also, according to the maintenance management method for railway equipment parts of the present invention, the maintenance work of parts can be greatly reduced by detecting the change of the inspection object using the image analyzer.

1 is a configuration diagram illustrating an image analysis apparatus and an imaging apparatus according to an embodiment of the present invention. It is a perspective view which shows an example of the connection part of the track circuit which is a test object. It is a figure which shows an example of the picked-up image obtained by an imaging device. It is a flowchart which shows the procedure of a maintenance management process. It is a figure explaining the position detection process of a test object. It is a flowchart which shows the detailed procedure of the image comparison process of step S6 of FIG. It is a flowchart which shows the detailed procedure of the feature-value calculation process of step S12 of FIG. It is a figure explaining a feature-value calculation process. It is a figure explaining the comparison process of the feature-value. It is a figure explaining the effect of a mask process.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
FIG. 1 is a configuration diagram illustrating an image analysis apparatus and an imaging apparatus according to an embodiment of the present invention. FIG. 2 is a perspective view illustrating an example of a connection portion of a track circuit that is an inspection object.
The image analysis apparatus 1 according to the embodiment is an apparatus that compares the images of the inspection target T (see FIG. 2) taken a plurality of times at intervals of days and detects the presence or absence of the change in the inspection target T. The image analysis apparatus 1 is a computer, and includes a CPU (Central Processing Unit) 10, a RAM (Random Access Memory) 11, a display unit 12, an operation unit 13 such as a mouse or a keyboard, a data input interface 14 and a storage device 15. Prepare. The interface 14 corresponds to an example of an image input unit according to the present invention.

  The storage device 15 includes an imaging data storage unit 151 that stores imaging data of the inspection target T, and a feature amount storage unit 152 that stores feature amounts of images calculated in the past. In addition, the storage device 15 is provided with a template storage unit 153 in which image patterns of a large number of inspection objects T are stored in order to detect the inspection object T. Furthermore, the storage device 15 stores a plurality of program modules that realize processing for detecting whether or not the inspection object T has changed. These programs include a position detection program 154, a rough alignment program 155, a detailed alignment program 156, a feature amount calculation program 157, and an image comparison program 158. Among these, the CPU 10 that executes the feature amount calculation program 157 functions as a feature amount calculation unit according to the present invention, and the CPU 10 that executes the image comparison program 158 functions as an image comparison unit according to the present invention.

  As shown in FIG. 2, the inspection object T is a component (specifically, a connecting portion of the track circuit) provided on the rail R, and a fixed portion in which the leading end of the conducting wire 50 is fixed to the side portion of the rail R. 52 and the front end portion 51 of the conducting wire 50. A large number of inspection objects T exist on the route at an arbitrary interval in the longitudinal direction of the rail R. In general, the connection part of the track circuit is connected to the rail R in a plurality of forms such as a welding type using low-temperature brazing welding, copper thermite welding, and a hole drilling type. It is intended for connections connected by copper thermite welding.

FIG. 3 is a diagram illustrating an example of a photographed image obtained by the photographing apparatus.
The image analysis apparatus 1 inputs image data obtained by photographing the inspection target T from the outside via the interface 14. The image data of the inspection target T is photographed by a photographing device 31 (see FIG. 1) fixed to the lower part of the vehicle body of the railway vehicle 30. A lighting device 32 that illuminates a portion to be photographed is provided around the photographing device 31.

  The imaging device 31 is a line sensor in which a plurality of light receiving elements are linearly provided, although not particularly limited. The photographing device 31 is disposed at the lower part of the vehicle body of the railway vehicle 30 and is fixed toward the rail R from a direction oblique to the vertical. While the imaging device 31 repeatedly performs imaging at high speed, the imaging device 31 and the imaging location move relative to each other as the railway vehicle 30 travels, so that a predetermined width around the rail R is obtained as shown in FIG. A two-dimensional image continuous in the longitudinal direction of the rail R can be acquired. The photographed image of FIG. 3 is an image obtained by photographing one rail R from the photographing device 31 at the lower left end of the railway vehicle 30. For example, the imaging device 31 performs imaging in synchronization with the vehicle speed (for example, wheel rotation) of the railway vehicle 30, so that an image with a constant scale ratio in the longitudinal direction of the rail R can be obtained without depending on the speed of the railway vehicle 30. can get. The image data obtained by the imaging device 31 is divided into a plurality of pieces in the longitudinal direction of the rail R, for example, and sent to the image analysis device 1 via communication or a storage medium.

<Maintenance management method>
FIG. 4 is a flowchart showing a procedure of maintenance management processing.
The maintenance management method for railway equipment parts according to the present embodiment performs maintenance management of the connection part (inspection object T) of the track circuit by repeatedly performing the maintenance management process of FIG. 4 every predetermined period. When the maintenance management process is started on the inspection target date, the attendant first takes an image using the imaging device 31 from the traveling railway vehicle 30 (step S1). Photographing is performed in the route section subject to maintenance management, and a long image along the rail R is acquired by this photographing. A large number of inspection objects T are included in the image. For example, the image data is divided into predetermined pixels in the longitudinal direction and is handled as a plurality of images. Hereinafter, these plural images are referred to as “image sets”.

  Next, the clerk inputs image set data to the image analysis apparatus 1 via communication or a storage medium (step S2). Data of the image set is input via the interface 14 and stored in the shooting data storage unit 151. Note that the imaging data storage unit 151 stores data of image sets of a plurality of inspection target days obtained in the same manner in the past maintenance management process.

Subsequently, the clerk designates the image set on the inspection target date as a comparison target, and performs an operation for causing the image analysis apparatus 1 to execute the image analysis processing. Then, the CPU 10 of the image analysis device 1 executes the image analysis processing in steps S3 to S7.
When the image analysis process is started, first, the CPU 10 executes a position detection process for detecting the positions of a plurality of inspection objects T from the image set (step S3). This process is realized by the CPU 10 executing the position detection program 154 shown in FIG.

FIG. 5 is a diagram for explaining the position detection process of the inspection object.
In the position detection process, a template of the inspection target T stored in advance in the template storage unit 153 is used. The template is data indicating image feature amounts for many inspection objects T, and is created in advance by learning processing using the computer w1, as shown in FIG. In the learning process, many images HP of the inspection target T obtained by imaging similar to the imaging process in step S1 are input to the computer w1, and a frame indicating the position of the inspection target T in each image HP is input. f1 is designated. The designation of the frame f1 may be performed by a person at the beginning. The computer w1 extracts feature quantities on the image of the inspection object T from the frame images h1 of many inspection objects T having different shapes and the like, and a feature amount template for identifying the presence or absence of the inspection object T Are created in advance by statistical learning.

  In the position detection process in step S3, the CPU 10 recognizes the rail R from each image H0 on the inspection target date, and identifies a range where the inspection target T may exist based on the position of the rail R. Then, the CPU 10 sets a target frame f2 that identifies the presence or absence of the inspection target T in a range that may exist. Then, the CPU 10 extracts the feature amount of the image in the target frame f2, and performs pattern recognition by a support vector machine (SVM) or the like using a template prepared in advance. Thereby, the CPU 10 determines whether or not the inspection target T is included in the target frame f2. As indicated by an arrow in the image H0 in FIG. 5, the CPU 10 repeats the process of determining whether the inspection target T is included while shifting the target frame f2 within a range where the inspection target T may exist. . Then, the CPU 10 detects the position f3 of each inspection target T included in each image H0 on the inspection target date.

  In the position detection process in step S3, the CPU 10 detects the position of the inspection target T described above for all the image sets input in step S2. Thereby, the positions of all the inspection objects T included in the image set are detected.

  When the position detection process is completed, the CPU 10 executes a rough alignment process between the comparison reference image set and the image set on the inspection target day (step S4). This process is realized by the CPU 10 executing the general alignment program 155 of FIG. The rough alignment process associates the same inspection object T between the plurality of inspection objects T in the comparison reference image set and the plurality of inspection objects T in the image set of the inspection object day, This is a process of aligning images so that the same inspection object T is roughly at the same position.

  The image set for comparison and the image set for the inspection object date are images that are long along the rail, and the interval between each inspection object T and the subsequent inspection object T among the many inspection objects T depends on the shooting date. Almost unchanged. On the other hand, it is difficult to match the shooting start points of the comparison reference image set and the image set on the inspection target day. In addition, in the case where shooting is interrupted when the railway vehicle 30 that performs shooting stops at the station, it is difficult to match the shooting interruption point and the restart point between the comparison reference image set and the image set on the inspection target day. . Therefore, the inspection object T included in the image set on the inspection object day cannot be identified as it is among the many inspection objects T included in the comparison reference image set. Accordingly, rough alignment is performed by the processing in step S4 so that the same inspection object T can be associated between the comparison reference image set and the inspection day image set.

  In the rough alignment process, the CPU 10 calculates the interval length between each inspection object T included in the image set and the inspection object T subsequent thereto, and arranges the calculated interval length in the order in which the inspection objects T are arranged. To do. The CPU 10 performs such processing for both the comparison reference image set and the inspection day image set. Subsequently, the CPU 10 compares the array of interval lengths of the comparison reference image set with the array of interval lengths of the image set on the inspection target day, and sets the image sets so that the interval lengths are aligned within an error range. Correct the deviation of the start point, the interruption point, or the restart point. Thereby, the rough position of each inspection object T in both image sets is aligned, and the approximate alignment process is completed.

  When the rough alignment process is completed, the CPU 10 executes a detailed alignment process between the inspection target T in the comparison reference image set and the inspection target T in the image set on the inspection target date (step S5). This process is realized by the CPU 10 executing the detailed alignment program 156 of FIG. The detailed alignment process is a process of aligning the position in the comparison reference image and the position in the image on the inspection target day in units of pixels for the same inspection target T. When the imaging device 31 in FIG. 1 captures the same inspection object T several times over a number of days, due to shaking of the vehicle body of the railway vehicle 30 and the difference in relative position between the wheel and the rail R, etc. There is a slight difference in the shooting distance and shooting angle from the shooting device 31 to the shooting target. For this reason, the inspection target T has a slight shift in the position in the image and the direction in which the image is taken every time the image is taken. Such slight deviation is corrected by the detailed alignment process.

  In the detailed alignment process, the CPU 10 cuts out an image in which the same inspection target T is captured from the comparison reference image set and the inspection target day image set, and a plurality of images are displayed from both images. Extract feature points. The feature point is an inspection object T or an edge portion of an object in the background. Then, the CPU 10 associates a plurality of feature points between both images using a technique such as RANSAC (Random sample consensus), and inspects the feature points so that the deviation of the positions of the common feature points between both images is reduced. Projectively transform the image of the target day. As a result, the position of the inspection object T is aligned between both images in pixel units. When the above process is performed for all inspection objects T included in the image set, the detailed alignment process is completed.

  When the detailed alignment process is completed, the CPU 10 performs an image comparison process for comparing two images including the same inspection target T (step S6). This process is realized by the CPU 10 executing the image comparison program 158 of FIG. Details of the image comparison process will be described later. By this image comparison process, it is determined whether or not there is a change in each inspection target T included in the image set on the inspection target date.

  When the image comparison process is completed, the CPU 10 determines an inspection target T having a change equal to or greater than a preset threshold value from all the inspection targets T included in the image set, and determines the corresponding inspection target T. An image is extracted (step S7). The image is output to the display unit 12, printed out, or output as data. Further, position information indicating where the inspection target T is located on the route may be added to the image data. The position information may be acquired by performing positioning using GPS (Global Positioning System) when photographing the inspection target T.

  Once an image of the inspection object T that has been determined to have changed is extracted, the clerk checks this image in detail, and if necessary, visits the site to check the actual object to determine whether repair is necessary. (Step S8). Such a process completes one maintenance management process.

  According to such a maintenance management method, it is possible to greatly reduce the labor of inspection as compared with a method in which a staff member inspects all inspection objects T existing on a track. Furthermore, since the image analysis apparatus 1 detects changes in each inspection target T precisely and accurately, the strict inspection of the inspection target T is performed in the same manner as when an attendant inspects all inspection targets T visually. Inspection can be realized.

<Image comparison processing>
Next, the image comparison process in step S6 in FIG. 4 will be described in detail. FIG. 6 is a flowchart showing a detailed procedure of the image comparison process.
When proceeding to the image comparison process, the CPU 10 first cuts out a comparison target region K1 (see FIG. 8) including one inspection target T from the image set on the inspection target date (step S11). The comparison target area K1 is, for example, a rectangular area, and is set to a predetermined size so that the inspection target T is included with a margin and not other than the inspection target T is included. When the comparison target region K1 is cut out, the CPU 10 executes a feature amount calculation process for calculating the feature amount of the image for the comparison target region K1 (step S12). Hereinafter, each of a plurality of inspection objects included in one image set is described as “inspection object“ i ”” by adding an index number i (i = 1 to m). Assume that the total number of inspection objects included in one image set is m.

FIG. 7 is a flowchart showing a detailed procedure of the feature amount calculation process in step S12 of FIG. As will be described later, the order of each process is not limited to the example of FIG. FIG. 8 is an explanatory diagram showing the procedure of the feature amount calculation process.
When proceeding to the feature amount calculation process, first, the CPU 10 divides the comparison target region K1 into a plurality of local regions K11, for example, every 10 × 10 px (pixels) (step S31). The CPU 10 that executes the process of step S31 corresponds to an example of an image dividing unit according to the present invention. In FIG. 8, in order to avoid complexity, only some local regions K11 are denoted by reference numerals. The size of the local region K11 is not limited to 10 × 10 px and can be changed as appropriate. A pixel means a pixel.

When the CPU 10 is divided into a plurality of local regions K11, the CPU 10 repeatedly executes the processes of steps S32 to S33 for each local region K11. In the repetitive processing, the CPU 10 first calculates the luminance gradient direction for each pixel in the local region (step S32). CPU10 which performs the process of step S32 is equivalent to an example of the gradient direction calculation part which concerns on this invention. The luminance gradient direction indicates direction information obtained by omitting the size information from the luminance gradient vector including the size and direction information. The gradient direction of the luminance of an arbitrary pixel is expressed by the following equation (1), for example.
Here, x and y indicate the coordinates of the pixel, “x + 1” represents the right column for any x value, and “y + 1” represents the next row for any y value. Represents. Hereinafter, the pixel having the coordinates (x, y) is referred to as “pixel (x, y)”. Note that the value of “−180 ° to 0 °” may be corrected to a value of “180 ° to 360 °” so that the luminance gradient direction θ becomes a positive value.

The function in the right side of Expression (1) is defined by the following Expressions (2) to (4).
Here, L (x, y) represents the luminance of the pixel (x, y), and V _L (x, y) represents the luminance gradient vector of the pixel (x, y). Double-line characters _ex and _ey indicate vertical and horizontal basic vectors.

  In the examples of the expressions (2) to (4), as the luminance gradient vector of a certain pixel (x, y), the luminance difference between the upper pixel and the lower pixel of the pixel (x, y), The definition that the luminance difference between the left pixel and the right pixel is a component is shown. However, the definition of the brightness gradient vector is not limited to this. For example, as the luminance direction vector of a certain pixel (x, y), the luminance difference between the pixel (x, y) and the pixel above it, and the luminance difference between the pixel (x, y) and the pixel on the right thereof And may be defined as components. Moreover, it is good also as a definition in which the brightness | luminance difference with the diagonal pixel of the object pixel (x, y) is contained in a component. In addition to the definition that one luminance gradient vector is determined for each pixel, for example, if the number of pixels in the local region K11 is larger, one luminance gradient vector for each vertical and horizontal 2 × 2 or vertical and horizontal 3 × 3 pixel May be defined to be determined. Furthermore, in the example of Formula (1), the definition which makes the gradient direction of a brightness | luminance the direction component of a brightness | luminance gradient vector was shown. However, it is possible to adopt other definitions, for example, by defining the direction and direction of the top / bottom / left / right slant that has the largest brightness difference as the brightness gradient direction by comparing one pixel (x, y) with surrounding pixels. it can.

The luminance gradient vector V _L (x, y) in Equation (2) is an amount that includes information on the magnitude and direction of the luminance gradient, whereas the luminance gradient direction θ (x, y) in Equation (1). y) is an amount specialized in the direction of the luminance gradient without the information on the size of the luminance gradient.

  After the luminance gradient direction is calculated for all the pixels in the local region K11 by the above step S32, the CPU 10 then calculates the frequency distribution by totaling the luminance gradient directions of all the pixels in the local region K11 ( Step S33). The CPU 10 that executes the process of step S33 corresponds to an example of a distribution calculation unit according to the present invention. 8 shows a histogram of this frequency distribution. In the calculation of the frequency distribution, for example, the number of pixels having a luminance gradient direction included in each angle range is counted for each angle range obtained by equally dividing 360 ° into a plurality of angles, and this count value is used as the frequency of each angle range. That's fine. This frequency distribution is a feature amount of one local region K11.

  When the frequency distribution in the gradient direction of the luminance is calculated for one local region K11, the CPU 10 determines whether the calculation has been completed for all the local regions K11. If not yet, the process returns to step S32 to return to the next local region K11. Calculation for the region K11 is executed. On the other hand, if completed, the feature amount calculation process is terminated. By the feature amount calculation processing, a plurality of feature amounts corresponding to a plurality of local regions K11 of one comparison target region K1 are calculated.

  FIG. 7 shows an example in which the gradient direction of the luminance of each pixel is calculated for each local region after the comparison target region K1 is divided into a plurality of local regions. However, first, the luminance gradient direction may be calculated for all the pixels in the comparison target region K1, and then divided into a plurality of local regions to calculate the frequency distribution in the luminance gradient direction of each local region. . Even in this case, the frequency distribution in the gradient direction of a plurality of pixels included in each local region can be similarly calculated.

  Subsequently, the CPU 10 stores a plurality of feature amounts calculated for one comparison target region in the feature amount storage unit 152 of the storage device 15 (step S13). Here, for example, numbering is performed so that the CPU 10 can identify which comparison target region of which image set each of the plurality of feature amounts belongs to, and which local region within each comparison target region. Alternatively, a plurality of feature amounts are stored in association with correspondence information.

  Next, the CPU 10 designates n pieces of feature amounts of three or more image sets photographed at different dates as comparison criteria (step S14). For example, the CPU 10 specifies the feature amounts of the three image sets from the oldest among the feature amounts of the image sets of different dates and times stored in the feature amount storage unit 152 as the comparison reference. Hereinafter, each of the n comparison criteria is described as “comparison criteria“ j ”” with an index number j (j = 1 to n) added.

  Subsequently, the CPU 10 reads out the feature value for the inspection object “i” from the feature values of the image set of the comparison reference “j” (step S15). Then, the CPU 10 compares the feature amount calculated in step S12 with the feature amount read in step S15 (step S17).

FIG. 9 is an explanatory diagram of the feature amount comparison processing.
Here, as shown in FIG. 9, the CPU 10 features the feature quantity Q1 of each local region K11 of the comparison target region K1 and the feature of the local region K21 at the same position as the comparison target in the image region K2 of the comparison reference “j”. Compare the quantity Q2. By comparison, for example, the CPU 10 calculates the frequency difference of the same angle range of both frequency distributions, and calculates the difference amount such as the sum of absolute values or the square mean value of the frequency differences for all angle ranges. . The difference amount is standardized so that the maximum value is 1.0. The amount of difference in the frequency distribution is not limited to this example, and may be a value that appropriately represents the difference in the frequency distribution. Such a calculation is performed in all the local regions K11 of the comparison target region K1. The dissimilarity map M1 of FIG. 9 is obtained by imaging the calculated difference amount of each local region K11 in correspondence with the arrangement of the local regions K11. Usually, as in the dissimilarity map M1, a result in which the amount of difference varies depending on the location is obtained.

  FIG. 10 is a diagram for explaining the effect of the mask process. Prior to the comparison process in step S17, the CPU 10 may recognize the range of the rail portion from the comparison target region as an image of the background region, and add a mask to at least a part of this range (step S16). . The CPU 10 that executes the process of step S16 corresponds to an example of a mask processing unit according to the present invention. Adding a mask means not comparing feature values of the portion. For example, there may be a large difference in the background image depending on the weather, such as when the weather is sunny or rainy. The second column from the left in the upper two rows of FIG. 10 shows a photographed image in fine weather and a feature difference map, and a high luminance change appears at the base of the rail. Further, the third column from the left in the upper row of FIG. 10 shows a photographed image in the rain and a feature amount difference map, and a low luminance change appears on the back side of the inspection object. In this way, even if the background region is changed, by performing the mask setting N1, it is possible to compare only the feature values of the inspection object. In the example of FIG. 10, an abnormality has occurred in the inspection object in the image in the fourth column from the left, but as can be seen from the comparison of the feature amount difference maps with the lowermost mask in FIG. A change in the region is excluded, and a comparison result reflecting the change in the inspection object can be obtained.

  When the comparison process is completed, the CPU 10 stores the comparison result in the work area of the RAM 11 (step S18). Then, the CPU 10 determines whether or not the comparison with the plurality of comparison criteria “1 to n” has been completed (step S19). If not, the process returns to step S15 to use the next comparison criteria “j”. Steps S15 to S18 are repeated. On the other hand, if it is completed, the CPU 10 proceeds to a step that continues the process.

  When the comparison results with all the comparison references “1 to n” are obtained, the CPU 10 determines whether or not the inspection object “i” has changed based on these comparison results (step S20). As an example, the CPU 10 extracts a portion having a difference greater than or equal to a threshold value from the difference map M1 (see FIG. 9) as a comparison result, calculates the sum of the difference values of these portions, and compares this sum. The amount of change in the target area K1 is used. CPU10 performs the same calculation about the comparison result with a some comparison reference | standard, and acquires a some variation | change_quantity. Further, the CPU 10 calculates whether there is a change amount deviating from a predetermined deviation among a plurality of change amounts, and if there is a change amount deviating, excludes the change amount, and performs statistical processing such as averaging for the rest. To obtain the final amount of change. Then, the CPU 10 determines that there is a change in the inspection object “i” if the final change amount is larger than a predetermined threshold, and determines that there is no change if it is smaller. The CPU 10 that executes the processes of steps S17 to S20 corresponds to an example of a change detection unit according to the present invention.

  In this way, by determining the presence or absence of a change in the inspection object “i” based on the comparison results of a plurality of comparison criteria, for example, even when an exceptional difference is included in one of the comparison criteria, It is possible to suppress the occurrence of an error in determining whether or not the inspection object “i” has changed. Therefore, the presence or absence of a change can be detected more accurately. The method for determining whether or not the inspection object “i” has changed based on a plurality of comparison results is not limited to the above example.

  After determining whether or not there is a change in one inspection object “i”, the CPU 10 determines whether or not all the inspection objects “1 to m” have been determined (step S21). If not, the CPU 10 returns the process to step S11 and repeats the processes of steps S11 to S20 for the next inspection object “i”. On the other hand, if completed, the image comparison process is terminated.

  By such an image comparison process, a determination result of whether or not there is a change is obtained for each of the plurality of inspection objects “i” included in the image set.

  As described above, according to the image analysis apparatus 1 of the present embodiment, the frequency distribution in the luminance gradient direction is used as the image feature amount in order to determine whether or not the inspection target T has changed by image comparison. Thereby, it is possible to largely eliminate the influence of the luminance change of the image, determine the presence or absence of the change in the inspection target T, and realize a precise and accurate detection of the change in the inspection target T. In addition, the calculation load can be greatly reduced as compared with the case where the luminance correction for removing the influence of the luminance change and the calculation of the HOG are performed.

  Further, according to the maintenance management method of the present embodiment, the inspection target T having a change can be extracted from the large number of inspection target T using the image analysis apparatus 1. Therefore, the labor of the inspection can be greatly reduced as compared with the method in which all the inspection objects T are inspected by the staff. In addition, by detecting a precise and accurate change of the image analysis apparatus 1, it is possible to realize a strict inspection as in the case where an attendant inspects by visual inspection or the like.

  The embodiment of the present invention has been described above. However, the present invention is not limited to the above embodiment. For example, in the embodiment, an example in which the connection part of the track circuit is applied as the inspection object is shown, but other objects may be applied. In the above-described embodiment, an example is shown in which a feature amount that is a frequency distribution in the gradient direction of luminance is used when performing image comparison. For example, when detecting the position of an inspection object, detailed alignment processing is performed. A frequency distribution in the gradient direction of luminance may be used as the feature amount of the image used at the time. Moreover, in the maintenance management method of the said embodiment, it was set as the method by which a staff member judges the necessity of repairing the inspection target T with the change which the image analysis apparatus 1 detected by visual observation. However, the computer may also determine whether or not final repair is necessary using image analysis or the like. In addition, the details shown in the embodiments can be changed as appropriate without departing from the spirit of the invention.

1 Image analyzer 10 CPU
DESCRIPTION OF SYMBOLS 14 Interface 15 Memory | storage device 151 Image | photographing data storage part 152 Feature-value storage part 153 Template memory | storage part 154 Position detection program 155 Outline alignment program 156 Detailed alignment program 157 Feature-value calculation program 158 Image comparison program 30 Railway vehicle 31 Imaging device 32 Illumination Equipment R Rail T Inspection object

Claims (5)

A feature amount calculation unit that calculates the feature amount of the local region of the image,
The feature amount calculation unit includes:
A gradient direction calculation unit that calculates a gradient direction in which the magnitude information is removed from the luminance gradient between the plurality of pixels included in the local region and the surrounding pixels;
A distribution calculator that calculates a frequency distribution in the gradient direction by aggregating the plurality of gradient directions calculated respectively corresponding to the plurality of pixels;
Including
An image analysis apparatus characterized in that the frequency distribution in the gradient direction is used as a feature amount of the local region. An image input unit for inputting a plurality of images taken of the inspection object;
An image comparison unit for comparing the plurality of images;
The image comparison unit
An image dividing unit that divides at least a portion including the inspection object into a plurality of local regions for each of the plurality of images;
2. A change detection unit that detects a change in the inspection object by comparing the feature amounts of the plurality of local regions calculated by the feature amount calculation unit between the plurality of images. The image analysis apparatus described. Three or more images obtained by photographing the inspection object are input to the image input unit,
The image comparison unit performs comparison processing of the feature amount between the comparison target and the plurality of comparison references, with a plurality of the three or more images as a comparison reference and at least one as a comparison target, The image analysis apparatus according to claim 2, wherein presence or absence of a change in the inspection object is determined based on a result of the plurality of comparison processes each using a plurality of comparison criteria.   4. The mask processing unit according to claim 2, further comprising a mask processing unit configured to mask at least a part of a background region of the inspection object before at least the change detection unit compares the feature amounts. Image analysis equipment. A maintenance management method for railroad equipment parts for managing each of a plurality of parts connected to railroad rails using the image analysis device according to any one of claims 2 to 4,
With the imaging device mounted on the railway vehicle traveling on the rail, the plurality of parts are imaged as the inspection object several times over a period of days,
The image analyzer detects a change in each of the plurality of parts by comparing a plurality of images obtained by the plurality of times of photographing,
A maintenance management method for railway equipment parts, comprising: determining whether or not repair is necessary for one or more parts for which a change equal to or greater than a threshold is detected among the plurality of parts.