JP7613058B2 - How to estimate tire measurements - Google Patents

Description

The present invention relates to a method, program, and device for estimating tire measurement values, as well as a method for generating a trained model.

Patent Document 1 discloses a method for quantifying a tire's meridian cross section. In this method, in order to create a tire model for use in simulations, etc., an image of the tire's outline and an image of the tire's internal structure are extracted from an image of the tire's meridian cross section, and then these are overlaid to obtain a composite tire cross section image. From this tire cross section image, the coordinates of the outline of the tire cross section image and internal structures such as belts are obtained. According to Patent Document 1, the "tire meridian cross section" is a cross section of the tire when cut on a plane that is parallel to and passes through the tire's axis of rotation.

JP 2007-179194 A

According to Patent Document 1, the tire cross-sectional image is digitized using a pointing device such as a digitizer or a mouse. In other words, to obtain coordinates on the tire cross-sectional image, a person must trace the composite image and specify a position on the composite image. For this reason, when it is desired to digitize the position information of a specific point on the actual tire cross-section and thereby estimate the measurement value of a specific location, a person must accurately specify a point on the tire cross-sectional image that corresponds to the specific point, which is a heavy burden and cannot be considered efficient.

The present invention aims to provide an efficient method, program, and device for estimating tire measurement values, as well as a method for generating a trained model.

A method for estimating tire measurements according to the present disclosure includes:
Obtaining a tire cross-sectional image representing a cross-sectional structure of the tire; inputting the obtained tire cross-sectional image into a trained machine learning model; deriving an output from the trained machine learning model; and estimating position information of one or more measurement points on the tire cross-section based on the output.

The method may further include estimating the distance between two predetermined points among the plurality of measurement points.

In the above method, the measurement points may be indicated by markers attached to the cross section of the tire represented by the tire cross section image.

In the above method, the tire cross-sectional image may be an image that represents the cross-sectional structure when the tire is cut along a plane that includes the axis of rotation of the tire.

The tire measurement value estimation program according to the present disclosure causes a computer to execute the following:
Obtaining a tire cross-sectional image representing a cross-sectional structure of the tire; inputting the obtained tire cross-sectional image into a trained machine learning model; deriving an output from the trained machine learning model; and estimating position information of one or more measurement points on the tire cross-section based on the output.

The tire measurement value estimation device according to the present disclosure includes an image acquisition unit, a storage unit, a derivation unit, and an estimation unit. The image acquisition unit acquires a tire cross-sectional image representing the cross-sectional structure of the tire. The storage unit stores a trained machine learning model. The derivation unit inputs the acquired tire cross-sectional image to the trained machine learning model and derives an output from the trained machine learning model. The estimation unit estimates position information of one or more measurement points on the tire cross-section based on the output.

The method for generating a trained model according to the present disclosure includes:
Preparing learning data, which is a data set of a tire cross-sectional image representing the cross-sectional structure of the tire, and correct answer data; and Using the learning data, adjusting parameters of a machine learning model so that when a tire cross-sectional image representing the cross-sectional structure of the tire is input, data corresponding to the correct answer data is output. Note that the correct answer data is positional information of one or more measurement points in the tire cross-sectional image.

According to the above disclosure, the position information of one or more measurement points is estimated from a tire cross-sectional image using a machine learning model. This makes it possible to efficiently estimate tire measurement values based on the position information of the measurement points.

1 is a functional block diagram illustrating a configuration of a tire measurement value estimation device according to an embodiment. An example of a cross-sectional image of a tire. 1 is a sample plan view showing a cross section of a tire. 11 is a flowchart showing the flow of an estimation process. FIG. 1 is a diagram illustrating an outline of the structure of a machine learning model according to an embodiment. An example of a screen generated by the estimation device. 11 is a flowchart showing the flow of a learning process.

The following describes a method, program, and device for estimating tire measurement values according to one embodiment of the present disclosure.

1. Overview of estimation method
Fig. 1 is a functional block diagram showing the configuration of a tire measurement value estimation device 1 according to an embodiment of the present disclosure, and Fig. 2 shows an example of a tire cross-sectional image (hereinafter also simply referred to as a "cross-sectional image") G0 showing the cross-sectional structure of a tire. The estimation device 1 according to this embodiment estimates the measurement values of a tire cross section from an image such as the cross-sectional image G0. The cross-sectional image G0 shows the outline of the tire cut at a predetermined plane and the internal structure of the tire.

The cross-sectional image G0 may be an image acquired by, for example, a computed tomography (CT) device. The cross-sectional image G0 may also be an image captured by a camera of, for example, a cross-section of a tire cut at a specific plane, or a cut sample (hereinafter also simply referred to as a "sample") 3 representing a full-size cross-section of a tire, or an image acquired by reading these with a scanner. The cross-sectional image G0 according to this embodiment is image data acquired by reading the sample 3 with a scanner. The gradation of the scanner is, for example, 24-bit color, and the resolution is preferably 300 dpi or more, and more preferably 600 dpi or more.

An example of a plan view of sample 3 is shown in FIG. 3. Sample 3 is a sample that represents the cross-sectional structure of a tire in actual size when the tire is cut on a plane including the tire's rotation axis. Sample 3 shows rubber 30, which is the main material of the tire, and internal structures 31-34 contained in rubber 30. Rubber 30 forms the contour of the tire and defines, for example, grooves 300 in the tread portion. Internal structures 31-34 are members that are configured separately from rubber 30, and are, for example, reinforcing materials for imparting strength to the tire. The internal structures shown in FIG. 3 are, in detail, cord 31, belt 32, ply 33, and bead wire 34. Furthermore, sample 3 may be provided with one or more markers corresponding to predetermined measurement points. In this embodiment, multiple markers 41-46 are provided, and these markers 41-46 represent measurement points P1-P6, which will be described later. Measurement points P1-P6 may be points on rubber 30 and points on internal structures 31-34.

The tire measurement values include the dimensions of a predetermined measurement point on the tire cross section. The dimensions of the predetermined measurement point can be expressed as the distance between two predetermined measurement points on the tire cross section represented by sample 3. There may be one or more predetermined measurement points, and measurement points that serve as the start and end points of the measurement can be determined for each predetermined measurement point. Typically, such measurements are obtained by a person identifying points on sample 3 that correspond to points specified on the tire design data as measurement points, and then manually measuring the distance between the identified measurement points.

However, the series of tasks of identifying measurement points on sample 3, measuring the distances between them, and recording the measurement results each time is time- and labor-intensive, and may ultimately lead to increased tire design and development costs and quality control costs. In addition, with this type of work, it is not possible to digitize the position information of the measurement points identified by hand, making it difficult to refer to which positions were used as measurement points after measurement.

The estimation device 1 according to this embodiment is configured to estimate position information of a measurement point identified by a person on the sample 3 from the output of a trained machine learning model (hereinafter also simply referred to as the "trained model") 130, and measure the distance between two points based on the estimated position information. This digitizes the position information of the points set as measurement points, making it easier to verify whether a valid measurement point has been identified for a specified point, and improving the reliability of the estimated measurement value. Furthermore, the measurement value is efficiently estimated, reducing the burden of human work. The estimated measurement value can be used in various situations, such as feedback to tire design and development and tire quality control. Below, the hardware configuration of the estimation device 1 will be described, and then the flow of the estimation process by the estimation device 1 will be described.

2. Hardware configuration of the estimation device
The estimation device 1 is a general-purpose computer in terms of hardware, and is configured by installing a predetermined program 131 in such a computer. The program 131 is a program that causes the estimation device 1 to execute processing described below, and is obtained, for example, from another device via a communication network 2 such as a LAN or the Internet, or from a computer-readable recording medium 132 such as a CD-ROM or a USB memory.

As shown in FIG. 1, the estimation device 1 includes a control unit 10, a display unit 11, an input unit 12, a memory unit 13, and a communication unit 14. These units 10 to 14 are connected to each other via a bus line 15 and can communicate with each other. In this embodiment, the display unit 11 is configured with a liquid crystal display or the like, and displays various information to the user. The input unit 12 is configured with a mouse, keyboard, touch panel, operation buttons, etc., and accepts operations on the estimation device 1 from the user.

The memory unit 13 is composed of a non-volatile storage device such as a hard disk or flash memory, and stores a trained model 130 and a program 131. The control unit 10 is composed of a processor (Central Processing Unit: CPU, Graphics Processing Unit: GPU), a ROM (Read Only Memory), a RAM (Random Access Memory), etc. The control unit 10 operates as an image acquisition unit 10A, a derivation unit 10B, an estimation unit 10C, and a learning unit 10D by reading and executing the program 131 in the memory unit 13. The operation of each unit 10A to 10D will be described in detail later. The communication unit 14 functions as a communication interface that connects the estimation device 1 to a communication network 2 and external devices such as a scanner.

In this embodiment, the trained model 130 is generated by a learning unit 10D incorporated in the estimation device 1 as described below. However, the learning function of training a machine learning model and generating the trained model 130 may be independent of the estimation function of estimating tire measurement values based on the trained model 130. In other words, the estimation device 1 may be equipped with only the estimation function, and the trained model 130 trained on another computer may be imported into the memory unit 13.

3. Estimation process flow
Next, the flow of the estimation process performed by the estimation device 1 will be described with reference to FIG.

In step S1, the image acquisition unit 10A of the estimation device 1 acquires the cross-sectional image G0 as described above via the communication unit 14. The image acquisition unit 10A stores the acquired cross-sectional image G0 in the memory unit 13. Note that the image acquisition unit 10A may generate an image in which the cross-sectional image G0 is divided into predetermined regions or an image in which the cross-sectional image G0 is resized to a predetermined size so as to be suitable for input to the trained model 130, and store the image in the memory unit 13. In the following, the cross-sectional image G0 includes an image that has been appropriately processed prior to input to the trained model 130. In addition, the xy coordinate system for the cross-sectional image G0 is defined as shown in FIG. 2, with the upper left vertex of the cross-sectional image G0 as the origin (0,0).

In step S2, the derivation unit 10B inputs the cross-sectional image G0 to the trained model 130 and derives an output from the trained model 130. The trained model 130 is configured to detect each of the markers 41-46 by extracting features from the input cross-sectional image G0, and estimate the most appropriate points (pixels) corresponding to the markers 41-46 as measurement points P1-P6. The trained model 130 outputs the x-y coordinates of the points estimated as measurement points P1-P6 as position information.

The type of the trained model 130 is not particularly limited, but in this embodiment, it is a convolutional neural network (CNN). FIG. 5 is a diagram for explaining the structure of a CNN. As shown in FIG. 5, the trained model 130 according to this embodiment includes an input layer, alternating convolutional layers and pooling layers, a fully connected layer, and an output layer. The input layer has H×W×(number of channels) units, where H is the number of pixels in the y-axis direction of the input image, W is the number of pixels in the x-axis direction of the input image, and the number of channels is 3 in this embodiment.

The convolutional layer applies a smaller-sized filter to the input three-dimensional data and outputs a feature map. There are no particular limitations on the number, size, or stride of the filters, and they can be selected as appropriate. The filters are automatically generated, and the parameters are adjusted by the learning process described below to more appropriately extract features. The pooling layer extracts and outputs characteristic values from the feature map output by the convolutional layer. There are no particular limitations on the size of the pooling, and they can be selected as appropriate.

The fully connected layer combines all the outputs from the final pooling layer into a vector format. The number of output units in the output layer can be set according to the number of estimated measurement points, and is 12 in this embodiment. In other words, a vector with 12 elements representing the coordinates of the six measurement points P1 to P6 is output from the output layer.

The above structure of the trained model 130 is merely an example, and the number of layers and the number of units are not limited to this. Furthermore, the convolutional layers and pooling layers do not necessarily need to be arranged alternately, and these layers can be increased or decreased as appropriate. Furthermore, dropout layers and activation layers may be added as necessary. The activation layer is a layer that performs calculations using an activation function, and an identity function, for example, is used as the activation function. However, the activation function is not limited to this, and any appropriate function can be used.

In step S3, the estimation unit 10C calculates the distance on the xy plane corresponding to the specified measurement location based on the coordinates output as the coordinates of the measurement points P1 to P6. For example, the estimation unit 10C can calculate the distance on the xy plane between the combinations of measurement points P1 and P2, measurement points P3 and P5, and measurement points P4 and P6, which are combinations of the start and end points of the measurement, respectively. The estimation unit 10C may also calculate a distance defined by a combination other than the above-mentioned combinations. Furthermore, the estimation unit 10C may calculate the distance between a point separately identified by known image processing or the like and any of the estimated measurement points P1 to P6, without relying on the trained model 130.

In step S4, the estimation unit 10C converts the calculated distances on the xy plane into distances on sample 3. This estimates the tire measurement values. The constants used for the conversion are calculated in advance and stored in the memory unit 13.

In step S5, the estimation unit 10C outputs the estimation result. The output mode is not particularly limited, but the estimation unit 10C may generate a screen showing the estimated measurement value for each predetermined measurement point and display it on the display unit 11. The estimation unit 10C may also generate an image in which pixels corresponding to the estimated measurement points P1 to P6 are displayed in a predetermined color, or an image in which figures N1 to N6 indicating the positions of the estimated measurement points P1 to P6 are displayed, and the image may be superimposed on the cross-sectional image G0 and displayed on the display unit 11. FIG. 6 is an example of an image in which images in which cross-shaped figures N1 to N6 corresponding to the coordinates of the estimated measurement points P1 to P6, respectively, are displayed superimposed on the cross-sectional image G0. This allows a person to intuitively judge whether the estimation by the estimation device 1 is valid by looking at the image displayed on the display unit 11.

4. Learning process flow
Below, with reference to Figure 7, we will explain a method for generating the trained model 130, i.e., a method for training a machine learning model by the training unit 10D.

In step ST1, the learning unit 10D reads N training data sets that have been prepared prior to training. The training data sets can be obtained from another device via a communication network 2 such as a LAN or the Internet, or from a computer-readable recording medium 132 such as a CD-ROM or USB memory.

The learning dataset is a dataset that combines cross-sectional images Gn (n = 1, ..., N) that represent the cross-section of a tire with correct answer data Yn. The cross-sectional images Gn can be acquired in the same manner as the cross-sectional image G0 described above, and have the same size (H x W) as the cross-sectional image G0 that is to be input to the trained model 130. The correct answer data Yn in this embodiment is a vector having 12 elements that represent the x and y coordinates (position information) of the measurement points P1 to P6. The learning dataset is divided in advance into a training dataset and a test dataset, and the ratio of the two can be set appropriately.

The correct answer data Yn is determined by identifying the xy coordinates of a representative pixel among the pixels (points) in the cross-sectional image Gn that correspond to the markers 41 to 46 indicating the measurement points P1 to P6. The representative pixel can be identified, for example, by referring to the measurement value when a person measures the sample 3. The representative pixel can also be identified using a known program that detects the markers 41 to 46, or by a person checking the positions of the markers 41 to 46 on a monitor.

In step ST2, the learning unit 10D randomly selects K sample data sets from the training data set. K is a value called the batch size and can be set appropriately.

In step ST3, the learning unit 10D inputs the cross-sectional images Gk (k = 1, ..., K) of the sample data set into a machine learning model having a structure as shown in Figure 5, and derives the output Tk from the machine learning model. The output Tk is data corresponding to the correct answer data Yk combined with the cross-sectional image Gk of the input sample data set, and is a vector having 12 elements representing the xy coordinates (position information) of points Pk1 to Pk6.

In step ST4, the learning unit 10D adjusts parameters so that the value of the error function between the output Tk derived in step ST3 and the ground truth data Yk combined with the cross-sectional image Gk input in step ST3 is minimized. More specifically, the learning unit 10D adjusts and updates the weighting parameters in the fully connected layer of the machine learning model and the filter parameters in the convolutional layer by the error backpropagation method. The error function in this embodiment is expressed as the sum of squares error. The sum of squares error represents the distance between the measurement points P1 to P6 and the predicted points Pk1 to Pk6.

In step ST5, it is determined whether or not learning for one epoch has been completed. In this embodiment, when the processes from step ST2 to step ST4 have been performed for the same number of sample datasets as the number of training datasets, it is determined that learning for one epoch has been completed. If it is determined that learning for one epoch has not been completed, the learning unit 10D returns to step ST2 after step ST4. That is, the learning unit 10D randomly selects a sample dataset again, and repeats the procedures of steps ST3 and ST4 using the newly selected sample dataset. On the other hand, if it is determined that learning for one epoch has been completed, it is determined in step ST6 whether or not learning for all epochs has been completed.

If it is determined in step ST6 that learning of all epochs has not been completed, the process returns to step ST2 to learn the next epoch. The total number of epochs is not particularly limited and can be set as appropriate. If it is determined in step ST6 that learning of all epochs has been completed, the learning unit 10D ends learning of the machine learning model. The learning unit 10D stores the latest parameters of the machine learning model in the memory unit 13, and sets this as the trained model 130. In other words, the trained model 130 is generated by the above procedure.

The learning unit 10D may input a test data set into the machine learning model each time one epoch of learning is completed, calculate the error between the output and the ground truth data of the test data set, and display the calculation result on the display unit 11. Furthermore, if it is considered that the error of the output of the machine learning model with respect to the ground truth data of the test data set has converged within a predetermined range before the completion of learning for all epochs, learning may be terminated at that point.

<5. Features>
According to the above estimation process, when a person performs a measurement operation, the position information of the measurement point that the person would identify on the sample 3 is efficiently digitized, and further, it becomes possible to refer to the identified measurement point later. This makes it easy to verify the reliability of the measurement, and makes it possible to suppress the measurement variation when the measurement operation is performed by a person. In addition, the measurement value of a specified measurement point is efficiently estimated from the cross-sectional image, reducing the burden and time of the measurement operation performed by a person.

6. Modifications
Although one embodiment of the present invention has been described above, the present invention is not limited to the above embodiment, and various modifications are possible without departing from the spirit of the present invention. The gist of the modified examples shown below can be combined as appropriate.

<6-1>
In the above embodiment, a CNN model is used as the machine learning model, but the machine learning model is not limited to this, and other machine learning models such as a neural network (NN) model may be used. In addition, multiple machine learning models may be combined. For example, the machine learning model may be a model in which a CNN model (decoder) and a CNN inverse implementation model (encoder) are combined, and a trained model 130 may be generated in which an image is output in which figures N1 to N6 indicating estimated measurement points P1 to P6 appear at the positions of the estimated measurement points P1 to P6.

<6-2>
In the above embodiment, the correct answer data Yk of the learning data set and the output of the trained model 130 are vectors having 12 elements, but the correct answer data Yk and the output of the trained model 130 are not limited to this. For example, the correct answer data Yk and the output of the trained model 130 may be data indicating the probability that each pixel of the input cross-sectional images Gk and G0 corresponds to the markers 41 to 46. In this case, the coordinates of the pixel with the highest probability for each of the markers 41 to 46 may be set as the position information of the measurement points P1 to P6, respectively.

<6-3>
The learning method of the machine learning model is not limited to the above embodiment, and any known parameter optimization algorithm such as stochastic gradient descent can be applied. The loss function is also not limited to the above embodiment, and can be changed as appropriate depending on the nature of the output data.

<6-4>
The measurement value estimation method of the above embodiment can be combined with other known image processing and applied to estimating various measurement values on the cross-sectional image G0. For example, the image processing can be applied to generate a virtual line that does not actually exist, connecting the points on both ends of the groove 300 of the sample 3, and to calculate the distance between the points on this virtual line and the points estimated as the measurement points P1 to P6.

1 Estimation device 10 Control unit 10A Image acquisition unit 10B Derivation unit 10C Estimation unit 10D Learning units 41 to 46 Marker 130 Trained model 131 Programs P1 to P6 Measurement points

Claims (7)

Obtaining a tire cross-sectional image representing a cross-sectional structure of the tire;
Inputting the acquired tire cross-sectional image into a trained machine learning model;
Deriving an output from the trained machine learning model; and
estimating position information of one or more measurement points on a cross section of the tire based on the output; and
Including,
the output corresponds to position information of the measurement point or a probability that each pixel of the tire cross-sectional image corresponds to the measurement point.
How to estimate tire measurements. The method further includes estimating a distance between two predetermined points among the plurality of measurement points.
The method for estimating a measurement value of a tire according to claim 1 . The measurement points are indicated by markers attached to the cross section of the tire represented by the tire cross section image.
The method for estimating a measurement value of a tire according to claim 1 or 2. The tire cross-sectional image is an image representing a cross-sectional structure of the tire when the tire is cut along a plane including a rotation axis of the tire.
The method for estimating a measurement value of a tire according to any one of claims 1 to 3. Obtaining a tire cross-sectional image representing a cross-sectional structure of the tire;
Inputting the acquired tire cross-sectional image into a trained machine learning model;
Deriving an output from the trained machine learning model; and
estimating position information of one or more measurement points on a cross section of the tire based on the output; and
on the computer ,
the output corresponds to position information of the measurement point or a probability that each pixel of the tire cross-sectional image corresponds to the measurement point.
A program to estimate tire measurement values. an image acquisition unit that acquires a tire cross-sectional image representing a cross-sectional structure of the tire;
A memory unit that stores a trained machine learning model;
A derivation unit that inputs the acquired tire cross-sectional image into the trained machine learning model and derives an output from the trained machine learning model;
an estimation unit that estimates position information of one or a plurality of measurement points on the cross section of the tire based on the output;
Equipped with
the output corresponds to position information of the measurement point or a probability that each pixel of the tire cross-sectional image corresponds to the measurement point.
A tire measurement estimation device. Preparing learning data, which is a data set of a tire cross-sectional image showing a cross-sectional structure of a tire and ground truth data;
Using the learning data, adjusting parameters of a machine learning model so that when a tire cross-sectional image showing a cross-sectional structure of a tire is input, data corresponding to the ground truth data is output;
Including,
The correct answer data is position information of one or more measurement points in the tire cross-sectional image.
How to generate a trained model.