JP2020128877A - Linear object abnormality detection device and abnormality detection method - Google Patents

Abstract

To provide a linear object abnormality detection device and an abnormality detection method that can detect an abnormality more accurately by improving detection accuracy in the shape and position of the abnormality and that can present the determination basis in more detail when the abnormality is detected.SOLUTION: A linear object abnormality detection device 10 detects an abnormality of a linear object by analyzing an image obtained by imaging the linear object. The linear object abnormality detection device 10 includes: a learning model 43 that deeply learns as learning data, a learning image about the linear object and a correct value of whether or not a pixel corresponds to the abnormality in each pixel in the learning image; an abnormality inference part 42 that infers whether or not a pixel corresponds to the abnormality for each pixel in an input image that is the image input as the input image to the learning model 43; and an abnormality determination part 51 that determines whether or not the abnormality exists in the input image based on the inference result of the abnormality inference unit 42.SELECTED DRAWING: Figure 2

Description

The present invention relates to a linear body abnormality detection device and an abnormality detection method.

BACKGROUND ART Conventionally, linear bodies such as wires, ropes, and electric wires have been widely used as important parts constituting equipment and the like. The linear body is used for various purposes such as supporting an object by binding and suspending the object and conducting electricity.
The linear body may be deteriorated with age or an external force may be applied to cause abnormalities such as disconnection and laceration. Abnormalities in the linear body may be detected by an inspection worker visually confirming the linear body itself or by viewing an image of the linear body. Need time and effort. Therefore, reduction of time and labor required for the inspection work of the linear body has been studied.

Patent Document 1 discloses an abnormal area detection device that detects the presence or absence of an abnormality in an inspection target and the position of the abnormality using high-order local autocorrelation characteristics. More specifically, a subspace of the normal area feature is generated in the area feature space based on the higher-order local autocorrelation feature, and the abnormal area is detected using the distance or angle from the subspace as an index. For example, a principal component analysis method is used to generate the normal region subspace, and the principal component subspace is formed by, for example, a principal component vector with a cumulative contribution ratio of 0.99.
In the abnormal area detection device as in Patent Document 1, it is necessary to obtain a normal data distribution from a normal data set as a preliminary work. Further, it is necessary to set a threshold value for judging abnormality from the distribution of normal data. These tasks require specialized knowledge and are not easy to introduce.

On the other hand, Patent Document 2 discloses a patrol inspection system that detects an abnormal portion occurring in a power transmission facility by image processing. In the patrol inspection system of Patent Document 2, since there are few samples of image data of arc traces generated by lightning strikes on overhead ground lines or power transmission lines, not only image data of abnormal points actually generated, Generates a pseudo image based on the image data of the abnormal/degraded state that is artificially added to the object to be inspected, and learns a large amount of learning models using the deep learning technology that utilizes a neural network. To do.

JP, 2011-14152, A JP, 2008-74757, A

In the system or apparatus as in the above-mentioned Patent Document 2, for example, a general convolutional neural network (Convolutional Neural Network, hereinafter referred to as CNN) is applied, and a line image is input to the line to input the image. It is possible to detect abnormalities in the state. The CNN basically includes a convolution layer that extracts features in an image by various filters, and a pooling layer that reduces the feature image so that an effective value remains in the feature image output by the convolution layer. ing. That is, in CNN, in the process of extracting and reducing the feature in the image, the position information of the feature in the image becomes ambiguous, and as a result, pattern recognition that is strong against the influence of the displacement of the object becomes possible. There is.
However, when the CNN is applied to the abnormality detection of the linear body, the positional information and the shape in the image are ambiguous even if the abnormality as described above can be detected although the characteristic as described above can be received. In addition, since the CNN performs the processing by the filter as described above, it is possible to determine in detail the position and the shape of the detected abnormality in the rectangular area to which the filter is applied. Difficult to know in.
Therefore, when the CNN is applied to the linear object detection device, when the abnormality is detected, the shape of the object at which position in the image the CNN has determined to be abnormal, and the basis for the determination are described in detail. I can't figure it out. Therefore, even if the abnormality detection is automated, in some cases, the operator must eventually visually check the image to determine the size, type, and position of the abnormality.
There is a need for an anomaly detection device and an anomaly detection method for linear objects that can detect anomalies more accurately by presenting more accurate anomaly shape and position detection, and that can provide a more detailed basis for making judgments when anomalies are detected. ing.

The problem to be solved by the present invention is to improve the detection accuracy of the shape and position of the abnormality, thereby detecting the abnormality more accurately, and presenting the determination basis at the time of the abnormality detection in more detail. An anomaly detection device and an anomaly detection method are provided.

The present invention adopts the following means in order to solve the above problems. That is, the present invention is a linear body abnormality detection device that analyzes an image of a linear body to detect an abnormality of the linear body, the learning image relating to the linear body, and the learning image. For each pixel in, the correct value of whether or not the pixel is abnormal, and a learning model deep learning has been provided as learning data, the image is input as an input image to the learning model, For each pixel in the input image, it is inferred whether or not the pixel corresponds to an abnormality, and the presence or absence of abnormality in the input image is determined based on the abnormality inference unit and the inference result of the abnormality inference unit. An abnormality detection device for a linear body, comprising:

Further, the present invention is a method for detecting an abnormality of a linear body, which detects an abnormality of the linear body by analyzing an image obtained by capturing an image of the linear body, the learning image relating to the linear body, and the learning image. For each pixel in, the correct value of whether or not the pixel is abnormal, and the learning model deep learning as learning data, the image is input as an input image, each pixel in the input image On the other hand, there is provided a method for detecting an abnormality in a linear body, which infers whether or not the pixel corresponds to an abnormality, and determines the presence or absence of abnormality in the input image based on the inference result.

According to the present invention, by increasing the detection accuracy of the shape and position of the abnormality, it is possible to detect the abnormality more accurately, and to present the determination basis at the time of abnormality detection in more detail, and An abnormality detection method can be provided.

FIG. 4 is an explanatory diagram of an overhead ground wire which is a target of abnormality detection in the abnormality detection device and the abnormality detection method for a linear body according to the embodiment of the present invention. It is a block diagram of the above-mentioned linear body abnormality detection device. It is explanatory drawing of the learning image input into the learning model in the abnormality detection device of the said linear body. It is explanatory drawing of the teacher image in the said learning model. It is a schematic block diagram of the said learning model. It is a flowchart of the abnormality detection method of the said linear body. It is a block diagram of an abnormality detection device of a linear object concerning the 1st modification of the above-mentioned embodiment. Explanatory diagrams of an inference result image which is an output image of the abnormality inference unit and an inspection image which is an output image of the inference result inspection unit, which are input to the inference result inspection unit of the abnormality detecting device for the linear body according to the first modification. Is. It is a flowchart of the abnormality detection method of the linear body regarding the said 1st modification. An inference result image which is an output image of the abnormality inference unit, which is input to the inference result inspection unit of the abnormality detecting device for a linear body, and an image which is being processed by the inference result inspection unit, regarding the second modified example of the embodiment. 9A and 9B are explanatory diagrams of a detailed examination image which is an output image of the inference result detailed examination unit.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
The linear body abnormality detection device in the present embodiment detects an abnormality of the linear body by analyzing an image of the linear body, and includes a learning image regarding the linear body and each of the learning images in the learning image. A corrective value of whether or not the pixel is abnormal is provided for a pixel, and a learning model that is deeply learned as learning data is provided, and an image is input as an input image to the learning model, and each pixel in the input image is input. On the other hand, the presence/absence of an abnormality in the input image is determined based on the abnormality inference unit that infers whether or not the pixel corresponds to the abnormality and the inference result of the abnormality inference unit.

FIG. 1 is an explanatory view of an overhead ground wire.
As described above, the abnormality detection device 10 is a device that detects an abnormality in the linear body. The linear body 2 in the present embodiment is, for example, an overhead ground wire 2 that is installed between the steel towers 1. The overhead ground wire 2 mainly protects the overhead wire 3 similarly installed between the steel towers 1 for the purpose of power transmission, distribution, etc. from lightning.
The abnormality detection device 10 includes an inspection machine 11, a control device 12, and a display device 13.

FIG. 2 is a block diagram of the abnormality detection device 10 for the linear body 2.
The inspection machine 11 is configured to travel on the linear body 2 autonomously or by being operated. The inspection machine 11 includes an imaging device 20, a storage unit 21, and a GPS reception unit 22.
The imaging device 20 captures an image of the surface of the linear body 2 as a moving image when the inspection machine 11 travels on the linear body 2. The imaging device 20 stores the moving image in the storage unit 21.
The GPS receiving unit 22 receives a GPS signal from a GPS satellite (not shown) at a predetermined time interval and identifies the coordinate position of the inspection machine 11. The GPS receiving unit 22 associates the specified coordinate position with the time when the coordinate position is acquired, and stores the combined coordinate position in the storage unit 21 in combination.
The display device 13 is, for example, a display or the like, and displays a detection result of an abnormality by the control device 12 according to an instruction from the control device 12 described below.

The control device 12 is an information processing device such as a personal computer or a server. The control device 12 includes a data storage unit 30, an image conversion unit 31, a learning unit 40, and an abnormality detection unit 50. The learning unit 40 includes a learning data storage unit 41 and an abnormality inference unit 42. The abnormality detection unit 50 includes an abnormality determination unit 51, a determination result storage unit 52, and a display control unit 53.
Among the components of the control device 12, the image conversion unit 31, the abnormality inference unit 42, the abnormality determination unit 51, and the display control unit 53 are, for example, software and programs executed by the CPU in each of the above information processing devices. You can Further, the data storage unit 30, the learning data storage unit 41, and the determination result storage unit 52 may be realized by a storage device such as a semiconductor memory or a magnetic disk provided inside or outside each of the information processing devices.

As described later, the abnormality inference unit 42 infers, with respect to each pixel in the input image, whether or not the pixel corresponds to an abnormality. In order to effectively perform this inference, the abnormality inference unit 42 includes a machine learning device as described later, and deep learning is performed on the machine learning device to generate the learning model 43. When actually performing the process of detecting the abnormality of the linear body 2, the abnormality inference unit 42 uses the learning model 43 for which the learning has been completed, with respect to each pixel in the input image. It is inferred whether or not the pixel corresponds to an abnormality.
In other words, the control device 12 roughly performs two operations, that is, learning during deep learning and inference as to whether or not each pixel corresponds to an abnormality when actually detecting an abnormality of the linear body 2. In order to simplify the explanation, in the following, first, a description will be given of each component of the control device 12 at the time of deep learning, and then it is inferred whether or not each pixel corresponds to an abnormality when actually detecting an abnormality. The behavior of each component of will be described.

First, the behavior of the constituent elements of the control device 12 during deep learning will be described.
Each of the moving image and the plurality of combinations of coordinate positions and times stored in the storage unit 21 of the inspection machine 11 are transferred to and stored in the data storage unit 30.
The image conversion unit 31 acquires a moving image from the data storage unit 30 and converts this into a plurality of still images. The image conversion unit 31 transmits the still image to the learning unit 40.

The learning unit 40 receives a still image from the image conversion unit 31, and saves the received still image as a learning image in the learning data storage unit 41.
FIG. 3 is an explanatory diagram of a learning image. The linear body 2 may be provided with a linear facility 4 such as a snow-hardening ring, which is a facility provided on the linear body 2. The learning image 45 shown in FIG. 3 is an image of a portion of the linear body 2 where the on-line equipment 4 is provided.
Abnormalities such as lacerations and arc marks may occur on the linear body 2. In the learning image 45 shown in FIG. 3, two abnormal portions 5 are shown on the linear body 2 and include an abnormality. However, the learning image 45 may naturally include an image of the linear body 2 in a normal state, that is, that does not include any abnormality.
In the present embodiment, the moving image, which is the generation source of the learning image 45, is imaged from the vicinity of the linear body 2 by the inspection machine 11 traveling on the linear body 2. Therefore, the portion of the linear body 2 including the on-line equipment 4 and the abnormal portion 5 occupies an area of a predetermined ratio or more in the learning image 45. The background 6 is provided except for the portion of the linear body 2.

In the learning data storage unit 41, the still image received from the inspection machine 11 via the image conversion unit 31 as described above is stored as a learning image, and also prepared by a learning operator by other means. An image of the linear body 2 including normal or abnormal is stored as a learning image.
In such a learning image prepared by other means as well, the portion of the linear body 2 including the on-line equipment 4 and the abnormal portion 5 is imaged so as to occupy an area of a predetermined ratio or more in the learning image 45. , Or processed.

The abnormality inference unit 42 includes a machine learning device. The learning model 43 is generated by performing deep learning on the machine learning device. The learning model 43 infers whether or not each pixel corresponds to an abnormality when actually detecting the abnormality. In other words, the machine learning device is used as a program module that is a part of artificial intelligence software to generate a learned model 43 in which appropriate learning parameters have been learned.
Hereinafter, for simplification of description, the machine learning device included in the abnormality inference unit 42 and the learning model generated by learning the same are allotted with the same reference numeral, and the machine learning device 43 or the learning model 43 is used. I call it.
Each learning image 45 stored in the learning data storage unit 41 is used as an input image in the deep learning of the machine learning device 43. In order for the machine learning device 43 to perform deep learning on the input learning image 45, teacher data corresponding to the learning image 45, which is a target of deep learning, is necessary.

The teacher data corresponding to each learning image 45 is, for each pixel in the learning image 45, a correct value of whether or not the pixel corresponds to an abnormality. More specifically, the correct answer values are classified into the classification categories in which the objects displayed corresponding to each pixel in the learning image 45 are roughly classified into the background 6, the linear body 2, the linear equipment 4, and the abnormal portion 5. It indicates which one is applicable.
In particular, in the present embodiment, for each pixel, a background value that is a value indicating whether or not the pixel corresponds to the background 6 shown in FIG. 3, and whether or not the pixel corresponds to the linear body 2. , A linear body value that is a value that indicates the pixel, a linear facility value that is a value that indicates whether the pixel corresponds to the on-line equipment 4, and an abnormality that is a value that indicates whether the pixel corresponds to the abnormal portion 5. The combination of four partial values is the correct value. The background value, the linear object value, the on-line equipment value, and the abnormal portion value are set to 1, for example, when the pixel corresponds to each element in the classification section, and when not corresponding, for example, 0. Therefore, for each pixel, a combination of the background value, the linear body value, the line facility value, and the abnormal portion value in which each element is 0 or 1 (background value, linear body value, linear facility value, abnormal portion value) Value) is the correct value.
For example, the correct answer value of the pixel corresponding to the background 6 in FIG. 3 may be (1, 0, 0, 0). The correct value of the pixel corresponding to the linear body 2 can be (0, 1, 0, 0). The correct value of the pixel corresponding to the on-line facility 4 can be (0, 0, 1, 0). Further, the correct value of the pixel corresponding to the abnormal portion 5 can be (0, 0, 0, 1).
In this way, the correct answer value is provided corresponding to each pixel in the learning image 45, and indicates which of the classification category including the background 6, the linear body 2, and the abnormal portion 5 the corresponding pixel corresponds to. It is a thing.

An image in which each pixel is colored with a color of a classification category to which a corresponding correct value corresponds when a unique color is associated with each classification category is hereinafter referred to as a teacher image. That is, the teacher image has a resolution equivalent to that of the learning image 45, and each pixel has a color corresponding to the classification category to which the object displayed in the pixel corresponding to the pixel in the learning image 45 corresponds. There is.
FIG. 4 is an explanatory diagram of the teacher image 46. Each pixel in the background area 6R of the teacher image 46, which is an area corresponding to the background 6 in FIG. 3, is colored with a dot pattern that looks light (corresponding to light gray, for example). Each pixel in the linear body region 2R of the teacher image 46, which is a region corresponding to the linear body 2 in FIG. 3, is colored with a dot pattern that appears dark (corresponding to, for example, black). Each pixel in the linear equipment region 4R of the teacher image 46, which is a region corresponding to the linear facility 4 in FIG. 3, has a dot pattern with a different density between the linear body region 2R and the background region 6R (corresponding to dark gray, for example). It is colored with. Each pixel in the abnormal portion area (abnormal area) 5R of the teacher image 46, which is an area corresponding to the abnormal portion 5 in FIG. 3, is colored in white, for example.
The teacher data may be, for example, the teacher image of such a configuration to which the correct value corresponding to each pixel is added.
The teacher data is stored in the learning data storage unit 41 in association with the learning image 45.

The learning model 43 of the abnormality inference unit 42 includes a learning image 45 regarding the linear body 2 as described above, and a teacher that includes, for each pixel in the learning image 45, a correct value of whether or not the pixel is abnormal. Deep learning is performed using the data and as learning data.
FIG. 5 is a schematic block diagram of the machine learning device (learning model) 43. In the present embodiment, the learning model 43 is realized by a full-layer convolutional network (Fully Convolutional Network, hereinafter referred to as FCN), and whether or not each pixel in the input image corresponds to abnormality is realized by the semantic segmentation. Learned to reason. As will be described below, the FCN does not have a fully connected layer and directly inputs the feature map processed and generated in the convolutional layer to the deconvolutional layer. The FCN allows an image of an arbitrary size to be used as an input, and allows learning to select and estimate an appropriate feature.

The machine learning device 43 includes a convolution processing unit 47 and a deconvolution processing unit 48. The convolution processing unit 47 includes a plurality of convolution layers 47a, 47b, 47c connected in series. The deconvolution processing unit 48 includes a plurality of deconvolution layers 48c, 48b and 48a which are similarly connected in series. In FIG. 5, the number of convolutional layers and the number of deconvolutional layers are illustrated as three, respectively, but the number of convolutional layers and deconvolutional layers is not limited to three.

The abnormality inference unit 42 acquires the learning image 45 from the learning data storage unit 41 and inputs it to the convolutional layer 47a in the first stage.
The convolutional layer 47a includes a predetermined number of filters (not shown). The machine learning device 43 positions each filter on the learning image 45 and assigns the weight set in the filter corresponding to the pixel position to the pixel value of each pixel of the learning image 45 in the filter. The convolution filter processing is executed by adding the sum and calculating the sum. Thereby, the pixel value of one pixel in the convolutional layer 47a is calculated. The machine learning device 43 calculates a plurality of pixel values by executing such convolution filter processing while moving the filter on the learning image 45 at a predetermined resolution step, and arranges the pixel values so that 1 corresponding to the filter is obtained. Generates one image.
The machine learning device 43 executes this process for all the filters, and generates a feature map according to the number of filters.

The filter is actually expressed by weighting the enhancement or smoothing of the pixel value of the learning image 45. In the feature map generated by executing the convolution filter process using such a filter, the grayscale pattern of the image such as the edge feature is effectively extracted. Further, since the feature is extracted from the local area of the learning image 45 through the filter, the object is robust against the displacement of the position of the object existing in the learning image 45.

In the convolutional layer 47a, pooling processing is executed after the convolutional filter processing. More specifically, each feature map is divided into a plurality of small regions each having a size of 2×2, the maximum pixel value in each small region is calculated, and the pixel of one pixel is calculated. By setting the value, each of the 2×2 small areas of each feature map is converted into a 1×1 pixel, and the information is contracted. As described above, in the pooling process, the maximum pixel value is selected from the local area of each feature map, so it is possible to leave only the appropriate features that are specialized for the image, and the subsequent processes can be performed efficiently. It can be carried out.

The pooled feature map generated in the convolutional layer 47a becomes the input image of the convolutional layer 47b in the next stage.

Similar to the convolutional layer 47a, the convolutional layer 47b sequentially executes the convolutional filter process and the pooling process. The convolutional layer 47b has a predetermined number of filters smaller in size than the filters of the convolutional layer 47a. By using these filters, the convolutional filtering process is executed, and further the pooling process is executed. A predetermined number of pooled feature maps according to the above are generated.
Also in the convolution layer 47c, the convolution filter process and the pooling process are sequentially executed. The convolutional layer 47c is provided with a predetermined number of filters smaller in size than the filters of the convolutional layer 47b. By using these, the convolutional filtering process is executed, and further the pooling process is executed. A predetermined number of pooled feature maps according to the above are generated.
The filter weight in each convolutional layer 47a, 47b, 47c is adjusted by machine learning.
The feature map subjected to the pooling processing in the convolutional layer 47c is input to the deconvolutional layer 48c of the deconvolution processing unit 48.

The deconvolution processing unit 48 has a symmetrical structure with the convolution processing unit 47. That is, the learning image 45 is compressed into a low dimension by the convolution processing unit 47, but the deconvolution processing unit 48 operates so as to be expanded and restored from the low-dimensional compression state. More specifically, the deconvolution layer 48c that executes the deconvolution process corresponding to the convolution layer 47c, the deconvolution layer 48b that executes the deconvolution process corresponding to the convolution layer 47b, and the deconvolution process corresponding to the convolution layer 47a are executed. The output data 49 is generated by sequentially passing through the deconvolution layer 48a to be executed.

The machine learning device 43 outputs, as the output data 49, the probability that each pixel in the learning image 45 can correspond to each of the classification categories, in the same manner as the teacher data. For example, with respect to a certain pixel, the probability that the pixel corresponds to the background 6 is 0.6, the probability that the pixel corresponds to the linear body 2 is 0.2, and the probability that the pixel corresponds to each of the linear equipment 4 and the abnormal portion 5 is 0. When the value is 1, when the machine learning unit 43 infers, a combination of values (0.6, 0.2, 0.1, 0.1) is output for the pixel.
An image expressed based on the output data 49 in the same manner as the teacher image 46 shown in FIG. 4 is hereinafter referred to as an inference result image 56. That is, the inference result image 56 is an image in which information regarding each of the inference results is reflected and arranged at the position of the pixel corresponding thereto as a color. As shown in FIG. Among the plurality of probability values included in the output value corresponding to, the color corresponding to the classification section corresponding to the highest value probability is provided.

The feature map generated in each of the convolutional layers 47a, 47b, and 47c as described above is less susceptible to the positional shift and the size difference of the object in the image. In other words, much of the information regarding the position of the object is lost in the feature map generated by the convolution processing unit 47.
Here, in each of the convolutional layers 48c, 48b, 48a, the processing corresponding to the corresponding convolutional layers 47c, 47b, 47a is executed as described above, and the processed data in each of the convolutional layers 47c, 47b, 47a is used. By doing so, the positional information lost in each convolutional layer 47a, 47b, 47c is complemented. As a result, the output data 49 output from the deconvolution processing unit 48 and the inference result image 56 corresponding to the output data 49 have position information restored in pixel units.
The position information is complemented, for example, by storing the position of the pixel selected as the maximum pixel value in the pooling process of each convolution layer 47a, 47b, 47c, and enlarging the image in the deconvolution layers 48c, 48b, 48a. At the time of restoration, after copying the pixel values of the feature map to the stored location, the pixel values of the pixels between the copied pixels are set based on the pixel values of the pixels copied from these feature maps, for example, bilinear. Can be done by complementing with.

In the machine learning device 43, machine learning is performed so that the inference result regarding each pixel in the output data 49 becomes a value close to the correct value of the teacher data corresponding to the input learning image 45. For this reason, the abnormality inference unit 42 acquires the teacher data corresponding to the input learning image 45 from the learning data storage unit 41, and compares the correct answer value in the teacher data and the value of the output data 49 in pixel units. Then, for example, the squared error of these differences between each pixel is calculated as a cost function.
Then, the machine learning device 43 is machine-learned by adjusting the weight value of each filter by the error back propagation method or the like so as to reduce the cost function.
As a result, for example, the correct value of the pixel corresponding to the background 6 in FIG. 3 on the teacher data is (1, 0, 0, 0), and therefore the output value corresponding to the pixel of the output data 49 is It is possible that the probability value corresponding to the first background 6 is closer to 1 than the other, and the other probability values are closer to 0 than the first probability value.
As described above, when learning is completed, the learning model 43 outputs the probability that, when an image is input as an input image, for each pixel in the input image, the pixel may correspond to each of the classification categories.

Next, the behavior of each component of the control device 12 in the case of inferring whether or not each pixel corresponds to an abnormality in actual abnormality detection, that is, after the deep learning of the learning model 43 is finished will be described. To do.

In the same manner as described above, the inspection machine 11 is caused to travel on the linear body 2 that is the target of abnormality detection, and a moving image and a plurality of combinations of coordinate position and time are stored in the storage unit 21.
Each of the moving image and the plurality of combinations of coordinate positions and times stored in the storage unit 21 of the inspection machine 11 are transferred to and stored in the data storage unit 30.
The image conversion unit 31 acquires a moving image from the data storage unit 30 and converts this into a plurality of still images. The image conversion unit 31 acquires a plurality of combinations of coordinate position and time from the data storage unit 30, and for each of the converted still images, based on the plurality of combinations and the time when the still image was captured. The coordinate position where the still image is captured is estimated.
The image conversion unit 31 transmits the still image as an input image to the learning model 43 to the learning unit 40.
The image conversion unit 31 also transmits the input image and the imaged coordinate position corresponding to the input image to the learning unit 40 and at the same time to the determination result storage unit 52, which will be described later, to store the same.

The input image is, for example, an image 55 similar to the image shown as the learning image 45 in FIG. In the present embodiment, the moving image, which is the generation source of the input image 55, is captured by the inspection machine 11 traveling on the linear body 2 similarly to the learning image 45, and thus includes the linear equipment 4 and the abnormal portion 5. The part of the linear body 2 occupies an area of a predetermined ratio or more in the input image 55.

The abnormality inference unit 42 of the learning unit 40 receives the input image 55 from the image conversion unit 31.
The anomaly inference unit 42 inputs the input image 55 as an input image to the learning model 43, and infers, for each pixel in the input image 55, whether or not the pixel corresponds to an anomaly, by the learning model 43. ..
The learning model 43 is a learned model that has undergone deep learning so as to infer whether or not each pixel of the input image 55 corresponds to an abnormality.

The anomaly inference unit 42 performs a deep learning in advance by the learning unit 40 and adjusts and determines the value of each parameter constituting the FCN such as the weight of the filter, and the learned learning model 43 is programmed on the CPU, for example. Is executed to infer whether or not each pixel of the input image 55 corresponds to an abnormality.
More specifically, when the anomaly inference unit 42 inputs the input image 55 to the learning model 43, the learning model 43 sequentially traces the convolution layers 47a, 47b, 47c and the deconvolution layers 48c, 48b, 48a, Executes convolution processing, pooling processing, deconvolution processing, and the like. Finally, the output data 49 corresponding to the input image 55 is output from the deconvolution layer 48a.

When the input image 55 is input, the learning model 43 that has finished learning outputs the probability that each pixel in the input image 55 can correspond to each of the classification categories, as output data 49.
The anomaly inference unit 42 infers that for each pixel, the classification category having the highest probability of being applicable on the output data 49 is the classification category corresponding to the pixel.
The abnormality inference unit 42 also generates an inference result image 56 as shown in FIG. 4 for the output data 49. That is, the inference result image 56 is generated such that each pixel has a color corresponding to the classification section inferred to be applicable as described above.
The abnormality inference unit 42 transmits the input image 55, the output data 49, and the inference result image 56 generated based on the output data 49 to the abnormality determination unit 51 of the abnormality detection unit 50.

The abnormality determination unit 51 receives the input image 55, the output data 49, and the inference result image 56 from the abnormality inference unit 42.
The abnormality determination unit 51 determines the presence or absence of abnormality in the input image 55 based on the inference result of the abnormality inference unit 42, in this case, the inference result image 56.

When the classification section is inferred on a pixel-by-pixel basis by the semantic segmentation using the FCN, the classification section is inferred on a pixel-by-pixel basis, so that a misjudgment may occur, for example, when similar image patterns exist. .. Further, it is difficult to infer a boundary portion between the background 6 and an object such as the linear body 2, and an erroneous judgment may occur due to such a reason. In such a case, the misjudgment often appears as noise in a minute area in the inference result image 56.
In order to reduce such noise, the abnormality determining unit 51 causes the abnormality inferring unit 42 to infer that the pixels inferred by the abnormality inferring unit 42 are continuous on the inference result image 56, for example. When the area of the abnormal region) 5R is equal to or larger than a predetermined threshold value, it is determined that the input image 55 has an abnormality.
Further, when there is an abnormal portion region 5R having an area equal to or smaller than the above-mentioned predetermined threshold value, the abnormality determination unit 51 corresponds the corresponding portion in the inference result image 56 to the surrounding classification of, for example, the linear body 2. (For example, black corresponding to a dot pattern that appears dark in FIG. 4) is reset.
As a result, the possibility that the input image 55 is determined to be abnormal is reduced due to the abnormal partial region 5R that is regarded as noise.

The abnormality determination unit 51 resets the determination result, the output data 49, and the coloring in association with the input image 55 that is already stored in the determination result storage unit 52 by the image conversion unit 31, thus reducing noise. The inference result image 56 is stored in the determination result storage unit 52.

The determination result storage unit 52 stores the input image 55 together with the coordinate position where the input image 55 was captured.
The determination result storage unit 52 also stores the determination result corresponding to each input image 55, the output data 49, and the inference result image 56 in which coloring is reset and noise is reduced, by the abnormality determination unit 51. ing.

The display control unit 53 extracts the determination result that is determined to be abnormal from the determination result storage unit 52, and inputs the input image 55 corresponding to the determination result, the coordinate position where the input image 55 is captured, and the coloring is restored. The inference result image 56 in which the noise is reduced by being set is acquired.
The display control unit 53 displays the input image 55 and the inference result image 56 in contrast with each other on the display device 13, and also displays that the input image 55 is determined to be abnormal. The display control unit 53 also displays the coordinate position at which the input image 55 is captured by, for example, displaying a map in the vicinity and then marking the corresponding coordinate position on the map.

Next, an abnormality detection method using the abnormality detection device 10 for the linear body 2 will be described.
The abnormality detecting method of the linear body 2 according to the present embodiment detects an abnormality of the linear body 2 by analyzing an image obtained by capturing the linear body 2, and includes a learning image 45 regarding the linear body 2, For each pixel in the learning image 45, the correct value as to whether or not the pixel corresponds to an abnormality is input as an input image 55 to the learning model 43 deeply learned as learning data, and the input image 55 is input. For each pixel inside, whether or not the pixel corresponds to an abnormality is inferred, and whether or not there is an abnormality in the input image 55 is determined based on the inference result.
First, the operation of the learning unit 40 during deep learning will be mainly described with reference to FIGS. 1 to 5.

In the learning data storage unit 41, a still image received from the inspection machine 11 via the image conversion unit 31 is stored as a learning image 45. Further, the learning data storage unit 41 stores, as a learning image 45, an image of the linear body 2 prepared by the learning operator by other means and including a normal or abnormal state.

The abnormality inference unit 42 acquires the learning image 45 from the learning data storage unit 41 and inputs it to the convolution processing unit 47 of the learning model 43.
The convolution processing unit 47 sequentially performs the convolution processing and the pooling processing by each of the convolution layers 47 a, 47 b, and 47 c to generate a feature map, and inputs the feature map to the deconvolution processing unit 48.
The deconvolution processing unit 48 executes deconvolution processing so as to enlarge and restore the feature map by each deconvolution layer 48c, 48b, 48a. Here, in each of the deconvolutional layers 48c, 48b, 48a, by executing the processing corresponding to the corresponding convolutional layers 47c, 47b, 47a, and by using the processing data in each of the convolutional layers 47c, 47b, 47a, The position information lost in each convolutional layer 47a, 47b, 47c is complemented. As a result, the output data 49 output from the deconvolution processing unit 48 and the inference result image 56 corresponding to the output data 49 have position information restored in pixel units.

The abnormality inference unit 42 acquires the teacher data corresponding to the input learning image 45 from the learning data storage unit 41, compares the correct value in the teacher data with the value of the output data 49 in pixel units, and The squared error of these differences between pixels is calculated as a cost function.
Then, the machine learning device 43 is machine-learned by adjusting the weight value of each filter by the error back propagation method or the like so as to reduce the cost function.

Next, using FIG. 1 to FIG. 5 and FIG. 6, in the case of inferring whether or not each pixel corresponds to an abnormality in the actual abnormality detection, that is, after the deep learning of the learning model 43 is finished. The abnormality detection method in will be described. FIG. 6 is a flowchart of the method for detecting an abnormality in the linear body 2.

The inspection machine 11 is run on the linear body 2 that is the target of abnormality detection, and a moving image and a plurality of combinations of coordinate position and time are stored in the storage unit 21.
Each of the moving image and the plurality of combinations of coordinate positions and times stored in the storage unit 21 of the inspection machine 11 are transferred to and stored in the data storage unit 30.
The image conversion unit 31 acquires a moving image from the data storage unit 30 and converts the moving image into a plurality of still images (step S1). The image conversion unit 31 acquires a plurality of combinations of coordinate position and time from the data storage unit 30, and for each of the converted still images, based on the plurality of combinations and the time when the still image was captured. The coordinate position where the still image is captured is estimated.
The image conversion unit 31 transmits the still image as an input image to the learning model 43 to the learning unit 40.
The image conversion unit 31 also transmits the input image and the imaged coordinate position corresponding to the input image to the learning unit 40 and at the same time to the determination result storage unit 52 for storage.

The abnormality inference unit 42 of the learning unit 40 receives the input image 55 from the image conversion unit 31.
The anomaly inference unit 42 inputs the input image 55 as an input image to the learning model 43, and infers, for each pixel in the input image 55, whether or not the pixel corresponds to an anomaly, by the learning model 43. (Step S3).
More specifically, when the anomaly inference unit 42 inputs the input image 55 to the learning model 43, the learning model 43 sequentially traces the convolution layers 47a, 47b, 47c and the deconvolution layers 48c, 48b, 48a, Executes convolution processing, pooling processing, deconvolution processing, and the like. Finally, the output data 49 corresponding to the input image 55 is output from the deconvolution layer 48a.

The anomaly inference unit 42 infers that for each pixel, the classification category having the highest probability of being applicable on the output data 49 is the classification category corresponding to the pixel.
The abnormality inference unit 42 also generates an inference result image 56 as shown in FIG. 4 for the output data 49. That is, the inference result image 56 is generated such that each pixel has a color corresponding to the classification section inferred to be applicable as described above.
The abnormality inference unit 42 transmits the input image 55, the output data 49, and the inference result image 56 generated based on the output data 49 to the abnormality determination unit 51 of the abnormality detection unit 50.

The abnormality determination unit 51 receives the input image 55, the output data 49, and the inference result image 56 from the abnormality inference unit 42.
The abnormality determination unit 51 determines whether or not there is an abnormality in the input image 55 based on the inference result of the abnormality inference unit 42, in this case, the inference result image 56 (step S5).
The abnormality determination unit 51 determines that the area of the abnormal portion region (abnormal region) 5R formed by the pixels inferred by the abnormality inference unit 42 corresponding to the abnormal portion 5 being continuous on the inference result image 56 is predetermined, for example. If it is equal to or larger than the threshold of, it is determined that there is an abnormality in the input image 55.
Further, when there is an abnormal portion region 5R having an area equal to or smaller than the above-mentioned predetermined threshold value, the abnormality determination unit 51 corresponds the corresponding portion in the inference result image 56 to the surrounding classification of, for example, the linear body 2. (For example, black corresponding to a dot pattern that appears dark in FIG. 4) is reset.
The abnormality determination unit 51 resets the determination result, the output data 49, and the coloring in association with the input image 55 that is already stored in the determination result storage unit 52 by the image conversion unit 31, thus reducing noise. The inference result image 56 is stored in the determination result storage unit 52.

The display control unit 53 extracts the determination result that is determined to be abnormal from the determination result storage unit 52, and inputs the input image 55 corresponding to the determination result, the coordinate position where the input image 55 is captured, and the coloring is restored. The inference result image 56 in which the noise is reduced by being set is acquired.
The display control unit 53 displays the input image 55 and the inference result image 56 in contrast with each other on the display device 13 and displays that the input image 55 is determined to be abnormal (step S7). The display control unit 53 also displays the coordinate position at which the input image 55 is captured by, for example, displaying a map in the vicinity and then marking the corresponding coordinate position on the map.

Next, effects of the abnormality detecting device 10 and the abnormality detecting method for the linear body 2 will be described.

The linear body 2 abnormality detection device 10 of the present embodiment detects an abnormality of the linear body 2 by analyzing an image of the linear body 2 and includes a learning image 45 regarding the linear body 2. , A learning model 43 deeply learned as learning data for each pixel in the learning image 45 and a correct value of whether or not the pixel corresponds to an abnormality, and an image is input to the learning model 43 as an input image 55. , And infers with respect to each pixel in the input image 55 whether or not the pixel corresponds to abnormality, based on the inference result of the abnormality inference unit 42 and the abnormality inference unit 42. And an abnormality determination unit 51 for determining the presence or absence of abnormality.
Further, the abnormality detecting method of the linear body 2 of the present embodiment is to detect an abnormality of the linear body 2 by analyzing an image obtained by capturing the linear body 2, and a learning image 45 regarding the linear body 2. And the correct value for each pixel in the learning image 45 as to whether or not the pixel corresponds to an abnormality are input as an input image 55 to the learning model 43 that has undergone deep learning as learning data. For each pixel in the image 55, it is inferred whether or not the pixel corresponds to an abnormality, and the presence or absence of an abnormality in the input image 55 is determined based on the inference result.
According to the above-described configuration and method, the learning model 43 sets the learning image 45 regarding the linear body 2 and the correct value for each pixel in the learning image 45 as to whether or not the pixel corresponds to an abnormality. , Deep learning is performed as learning data. An image is input as the input image 55 using the learning model 43, and for each pixel in the input image 55, it is inferred whether or not the pixel corresponds to an abnormality.
That is, for each pixel of the input image 55, it is inferred whether or not an abnormality corresponds to each pixel. Therefore, when it is inferred that the linear body 2 in the input image 55 is abnormal, the inference is performed. The abnormality inference unit 42 can output accurate information regarding the position and shape of the generated abnormality. Therefore, it is possible to increase the accuracy of detecting the shape and position of the abnormality, and it is possible to understand in more detail the basis of the determination as to which position and shape of the object in the input image 55 the shape is inferred as.

If the accuracy of detection of the shape and position of the abnormality is improved and the basis of the abnormality determination can be grasped in detail, for example, when the determination result is stored in the determination result storage unit 52, the input image 55 is automatically classified according to the size of the abnormality. It is also possible to do. In this case, for example, when the inspector of the linear body 2 visually checks the inference result image 56 that is determined to be abnormal on the display device 13, it is possible to confirm the inference result image 56 by prioritizing the abnormalities. .. Alternatively, if there is a serious abnormality during classification, a system may be realized in which the inspection operator is warned by e-mail or the like before the visual confirmation by the display device 13 during storage in the determination result storage unit 52. It will be possible. That is, in the abnormality detecting device 10 and the abnormality detecting method for the linear body 2 as described above, since it is possible to efficiently search for a serious abnormality, it is possible to reduce the work amount of the inspection worker.

Further, the learning model 43 is realized by FCN and infers whether or not each pixel in the input image 55 corresponds to an abnormality by semantic segmentation.
According to the above configuration, the abnormality detecting device 10 and the abnormality detecting method for the linear body 2 can be appropriately realized.

Further, the learning image 45 and the input image 55 are imaged so that the linear body 2 occupies an area of a predetermined ratio or more in the learning image 45 and the input image 55, and the abnormality determination unit 51 infers that the abnormality corresponds to the abnormality. When the area of the abnormal region 5R formed by the continuous pixels is equal to or larger than a predetermined threshold value, it is determined that the input image 55 has an abnormality.
According to the above configuration, the possibility that the input image 55 is determined to be abnormal is reduced due to the abnormal partial region 5R that is regarded as noise. Therefore, the accuracy of abnormality detection can be improved.

The correct answer value indicates which of the background 6, the linear body 2, and the classification category including the abnormality, the pixel in the learning image 45 corresponding to the correct answer value, and the learning model 43 The probability that each pixel in the input image 55 may correspond to each of the classification sections is output, and the anomaly inference unit 42 infers that the classification section having the highest probability in each pixel corresponds to the pixel.
According to the above configuration, the abnormality detecting device 10 and the abnormality detecting method for the linear body 2 can be appropriately realized.

Further, the linear body 2 is an overhead ground wire. According to the above configuration, it is possible to detect an abnormality in the overhead ground wire.

[First Modification of Embodiment]
Next, a first modified example of the abnormality detection device 10 and the abnormality detection method for the linear body 2 shown as the above embodiment will be described with reference to FIGS. 7, 8, and 9. FIG. 7 is a block diagram of a linear body abnormality detection device according to the first modification. 8A and 8B are an inference result image which is an output image of the abnormality inference unit and an inference result which are input to the inference result inspection unit of the abnormality detecting device of the linear body according to the first modified example, respectively. It is explanatory drawing of the detailed examination image which is an output image of a result detailed examination part. FIG. 9 is a flowchart of the linear body abnormality detection method according to the first modification.
The abnormality detecting device 60 of the linear body 2 in the first modified example is different from the abnormality detecting device 60 of the linear body 2 in the above-described embodiment in that the abnormality detecting unit 62 of the control device 61 includes an inference result scrutinizing unit 63. The difference is that
Only matters related to the inference result scrutinizing unit 63 will be described below.

The inference result inspection unit 63 operates to infer whether or not each pixel corresponds to an abnormality in actual abnormality detection after the deep learning of the learning model 43 is completed. Therefore, during deep learning, the abnormality detection device 60 operates in the same manner as the abnormality detection device 10 of the above embodiment.

When inferring whether or not each pixel corresponds to an abnormality in the actual abnormality detection, the abnormality detection device 60 uses the abnormality inference unit 42 in the input image 55 in the same manner as the abnormality detection device 10 of the above embodiment. For each pixel, it is inferred whether or not the pixel is abnormal, and the output data 49 and the inference result image 56 based on the output data 49 are generated.
The abnormality inference unit 42 transmits the input image 55, the output data 49, and the inference result image 56 to the inference result scrutiny unit 63 of the abnormality detection unit 62.

The inference result inspection unit 63 receives the input image 55, the output data 49, and the inference result image 56 from the abnormality inference unit 42.
The inference result image 56 generated by the abnormality inference unit 42 may include erroneous inference that the non-abnormal portion is inferred to be abnormal. For example, in the inference result image 56A shown in FIG. 8A, in the background region 6R, the false inference part 64 that is inferred to be an abnormal part and is colored white is present as in the abnormal part region 5R.
When the pixel inferred to be abnormal in the abnormality inference unit 42 is outside the linear body region 2R, which is the region occupied by the linear body 2, as described above, the inference result inspection unit 63 determines that pixel. Judges that it does not correspond to an abnormality.
As a result, the inference result inspection unit 63 colors the erroneous inference portion 64 in the same manner as the surrounding, in this case, the background region 6R, as shown by the portion 66 in FIG. To generate.
In this way, the inference result inspection unit 63 scrutinizes the inference result by re-determining that the pixel inferred to be abnormal by the abnormality inference unit 42 is not abnormal.
The inference result inspection unit 63 transmits the input image 55 received from the abnormality inference unit 42, the output data 49, and the generated inspection image 65 to the abnormality determination unit 51.

The abnormality determination unit 51 determines the presence/absence of abnormality in the input image 55 based on the inspection result of the inference result inspection unit 63. That is, in the above-described embodiment, the abnormality determination unit 51 determines whether or not there is an abnormality in the input image 55 based on the inference result image 56 received from the abnormality inference unit 42. However, in the present modification, instead of this, a close inspection is performed. Based on the image 65, it is determined whether or not the input image 55 is abnormal.

In the abnormality detecting method for the linear body 2 according to the present modification, as shown in FIG. 9, after the processing in the abnormality inference unit 42 shown as step S3 in the above-described embodiment, the inference result inspection unit 63 causes the abnormality inference to be performed. The inference result is scrutinized by re-determining that the pixel inferred to be abnormal by the unit 42 is not abnormal (step S4).
After that, the processing shifts to the processing by the abnormality determination unit 51, which is shown as step S5 in the above embodiment.

It goes without saying that the first modified example has the same effect as that of the embodiment already described.
In particular, in this modification, it is possible to effectively suppress erroneous inference and erroneous detection of an abnormality.

[Second Modification of Embodiment]
Next, a second modified example of the abnormality detecting device 10 and the abnormality detecting method for the linear body 2 shown as the above embodiment will be described with reference to FIG. 10. FIGS. 10A, 10B, and 10C respectively show an inference result image which is an output image of the abnormal inference unit and an inference result inspection unit which are input to the inference result inspection unit in the second modified example. It is explanatory drawing of the image in process and the detailed examination image which is an output image of the inference result detailed examination part. The second modified example is a further modified example of the first modified example, and differs from the first modified example in the processing content of the inference result inspection unit.

The inference result inspection unit in this modification receives the input image 55, the output data 49, and the inference result image 56 from the abnormality inference unit 42.
As described in the first modification, the inference result image 56 generated by the anomaly inference unit 42 includes an erroneous error such as an erroneous inference part 64 of the inference result image 56B shown in FIG. Reasoning may be mixed. In addition to this, the inference result image 56 generated by the anomaly inference unit 42 may include erroneous inference regarding the linear body region 2R. In the inference result image 56B, in the background region 6R, there is an erroneous inference part 72 that is inferred to be the linear body 2 and colored black (corresponding to a dot pattern that looks dark), like the linear body region 2R.

The inference result inspection unit removes the erroneous inference portion 72 related to the linear body 2 in addition to the erroneous inference portion 64.
For this purpose, the inference result inspection unit first reduces the area other than the background area 6R occupied by the linear body area 2R, the on-line equipment area 4R, and the abnormal portion area 5R. By this reduction processing, when noise having a small area related to the linear body region 2R, the on-line equipment region 4R, and the abnormal portion region 5R is present in the inference result image 56, this is removed.
Next, after enlarging the area occupied by the linear body region 2R, the on-line equipment region 4R, and the abnormal portion region 5R to restore the original size, only the region having the largest area is associated with the linear body, The area other than the background area 6R is considered, and the other area is removed. As a result, erroneous inference of the linear body 2 that is not removed even by the reduction processing is removed.
As a result of the reduction/enlargement processing described above, the inference result inspection unit reconstructs the area 2RA related to the linear body 2, that is, the area 2RA other than the background area 6R, as shown as the linear body re-estimation image 70 in FIG. presume.

The inference result reviewing unit compares the linear body re-estimation image 70 in which the region 2RA related to the linear body 2 is re-estimated with the inference result image 56B, except for the portion corresponding to the region 2RA in the inference result image 56B. The portions 64 and 72 located at are judged to be erroneous inference portions 64 and 72 and removed.
As a result, as shown as a portion 73 in FIG. 10C, a close-up image 71 in which both the false inference portions 64 and 72 have been close-inspected as the background region 6R is generated.
In this way, the inference result inspection unit enlarges each inference result with respect to the inference result image 56 arranged at the position of the pixel corresponding thereto after reducing the area occupied by the linear body 2.

It goes without saying that the second modification has the same effects as those of the embodiment and the first modification described above.
In particular, in this modification, it is possible to effectively suppress erroneous inference and erroneous detection of an abnormality.

The abnormality detecting device and the abnormality detecting method for a linear body of the present invention are not limited to the above-described embodiment and each modified example described with reference to the drawings, and various other within the technical scope thereof. Modifications are possible.

For example, although the learning image 45 and the input image 55 are picked up by the image pickup device 20 mounted on the inspection machine 11 in the above-described embodiment and each modification, the present invention is not limited to this. The learning image 45 and the input image 55 may be picked up by an image pickup device mounted on a helicopter, a drone, or the like. In this case, it is desirable to magnify and image using a telephoto lens or the like so that the area of the linear body 2 in the learning image 45 or the input image 55 occupies an area of a predetermined ratio or more. Alternatively, when a still image is generated in the image conversion unit 31, the existing image processing technique is used to enlarge the portion of the linear body 2 so that the linear body 2 in the learning image 45 or the input image 55 can be displayed. The area of the part may occupy an area of a predetermined ratio or more.
Further, in the above-described embodiment and each modification, the coordinate position information corresponding to each still image is based on the GPS signal acquired by the GPS reception unit 22 mounted on the inspection machine 11, but the present invention is not limited to this. I can't. For example, the inspection machine 11 may be controlled to run at a constant speed, and the coordinate position may be derived by back-calculating from the shooting time of the still image based on the running speed. Alternatively, the coordinate position may be derived by providing a sensor such as an encoder for measuring the number of rotations of the pulley on the portion of the pulley that drives the inspection machine 11, and calculating the moving distance by this.

Other than this, the configurations described in the above-described embodiments and modifications may be selected or changed to other configurations without departing from the gist of the present invention.

2 linear bodies (overhead ground wire)
2R Linear object area 4 On-line equipment 4R On-line equipment area 5 Abnormal area 5R Abnormal area (abnormal area)
6 background 6R background area 10, 60 abnormality detection device 11 inspection machine 12 control device 13 display device 42 abnormality reasoning unit 43 learning model 45 learning image 51 abnormality determination unit 55 input image 56 inference result image 63 inference result inspection unit

Claims (9)

A linear body abnormality detection device, which detects an abnormality of the linear body by analyzing an image of a linear body,
The learning image regarding the linear body, and for each pixel in the learning image, the correct value of whether or not the pixel is abnormal, provided with a learning model deep learning as learning data, the learning model On the other hand, by inputting the image as an input image, for each pixel in the input image, to infer whether the pixel corresponds to an abnormality, an abnormality inference unit,
An abnormality detection device for a linear body, comprising: an abnormality determination unit that determines whether or not there is an abnormality in the input image based on the inference result of the abnormality inference unit. The linear body abnormality detection device according to claim 1, wherein the learning model is realized by a full-layer convolutional network, and it is inferred by semantic segmentation whether or not each pixel in the input image corresponds to an abnormality. .. The learning image and the input image are captured so that the linear body occupies an area of a predetermined ratio or more in the learning image and the input image,
The abnormality determining unit determines that there is abnormality in the input image when an area of an abnormal region formed by continuous pixels inferred to be abnormal is equal to or larger than a predetermined threshold value. Or the linear body abnormality detection device according to 2. The correct answer value indicates whether the pixel in the learning image corresponding to the correct answer value corresponds to a background, the linear body, or a classification category including an abnormality,
The learning model outputs a probability that each pixel in the input image can correspond to each of the classification categories,
The abnormality detecting device for a linear body according to claim 1, wherein the abnormality inferring unit infers that the classification section having the highest probability in each pixel corresponds to the pixel. A pixel further inferred to be abnormal by the anomaly inference unit is re-determined as not to be abnormal, and further comprises an inference result inspection unit for inspecting the inference result.
The linear abnormality detecting device according to any one of claims 1 to 4, wherein the abnormality determining unit determines whether or not there is an abnormality in the input image based on a detailed inspection result of the inference result detailed inspection unit. .. The inference result scrutinizing unit determines that the pixel is not anomalous when the pixel inferred to be anomalous by the anomaly inference unit is outside the region occupied by the linear body. 5. The linear body abnormality detection device according to item 5. 7. The inference result inspection unit enlarges each of the inference results with respect to an inference result image arranged at a position of a pixel corresponding to the inference result after reducing the area occupied by the linear body. Abnormality detection device for the linear body described. The abnormality detecting device for a linear body according to claim 1, wherein the linear body is an overhead ground wire. A method of detecting an abnormality of a linear body, which detects an abnormality of the linear body by analyzing an image of a linear body,
The learning image regarding the linear body and the correct value for each pixel in the learning image regarding whether or not the pixel is abnormal are input as learning data to the learning model deeply learned. Input as an image, for each pixel in the input image, infer whether the pixel is abnormal,
A method for detecting an abnormality in a linear body, which determines whether or not there is an abnormality in the input image based on the inference result.

Images