Learning device, learning method, and program

Abstract

To generate a machine learning model that accurately detects a road area from a far area, out of areas included in an image.SOLUTION: A learning device that trains a machine learning model that receives an image including a plurality of pixels as an input and outputs a determination value indicating whether each of the pixels represents a road, the learning device including: a setting unit configured to set a weight for an error between the determination value output by the machine learning model and learning data indicating whether each of the pixels represents a road; and a learning unit configured to train the machine learning model so as to reduce a value of a loss function calculated on the basis of the error with the weight set thereto. The setting unit is configured to increase the weight in one or more predetermined directions from a lower end center of the image.SELECTED DRAWING: Figure 1

Description

Technical Field

[0005]

[0001] The present invention relates to a learning device, a learning method, and a program.

Background Art

[0002] Conventionally, a technique for detecting a road area included in an image has been known for use in vehicle driving support or autonomous driving. For example, Patent Document 1 discloses a technique for detecting a road area by scanning an input image in the x direction, extracting edges, and detecting a white line based on a color difference in the vicinity of the edges.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, in the technique described in Patent Document 1, the accuracy of scanning may decrease due to a decrease in the number of pixels as the distance is farther in the area included in the image. Thus, in the conventional technique, there may be a case where the road area cannot be accurately detected from a distant area among the areas included in the image.

[0005] The present invention has been made in consideration of such circumstances, and one of its objects is to provide a learning device, a learning method, and a program that can generate a machine learning model for accurately detecting a road area from a distant area among the areas included in an image.

Means for Solving the Problems

[0006] The learning device, learning method, and program according to this invention employ the following configuration. (1) A learning device according to one aspect of the present invention is a learning device that takes an image containing a plurality of pixels as input and learns a machine learning model that outputs a discrimination value indicating whether or not each of the pixels represents a road, and comprises a setting unit that sets weights for the error between the discrimination value output by the machine learning model and learning data indicating whether or not each of the pixels represents a road, and a learning unit that learns the machine learning model to reduce the value of a loss function calculated based on the error to which the weights have been set, wherein the setting unit increases the weights in one or more predetermined directions from the center of the lower end of the image.

[0007] (2) In the embodiment of (1) above, the setting unit increases the weight from the center of the lower end of the image toward the upper end and the left and right ends as one or more predetermined directions.

[0008] (3) Another learning method according to this invention is a learning method for learning a machine learning model that takes an image containing a plurality of pixels as input and outputs a discrimination value indicating whether each of the pixels represents a road, wherein the computer sets weights for the error between the discrimination value output by the machine learning model and learning data indicating whether each of the pixels represents a road, and learns the machine learning model to reduce the value of a loss function calculated based on the error to which the weights have been set, wherein the setting increases the weights in one or more predetermined directions from the center of the lower end of the image.

[0009] (4): A program according to another aspect of the present invention is a program for training a machine learning model that takes an image containing a plurality of pixels as input and outputs a discrimination value indicating whether or not each of the pixels represents a road, wherein the program causes the computer to set weights for the error between the discrimination value output by the machine learning model and training data indicating whether or not each of the pixels represents a road, and trains the machine learning model to reduce the value of a loss function calculated based on the error to which the weights have been set, wherein the setting increases the weights in one or more predetermined directions from the center of the lower end of the image. [Effects of the Invention]

[0010] According to (1) to (4), it is possible to generate a machine learning model that accurately detects road areas from distant regions within the image. [Brief explanation of the drawing]

[0011] [Figure 1] This figure shows the configuration of the learning device 100 according to this embodiment. [Figure 2] This figure shows an example of the structure of training data 130A. [Figure 3] This diagram illustrates how to set the weights using the setting unit 110. [Figure 4] This flowchart shows an example of the processing flow performed by the learning device 100 according to this embodiment. [Modes for carrying out the invention]

[0012] Hereinafter, embodiments of the learning device, learning method, and program of the present invention will be described with reference to the drawings.

[0013] [composition] Figure 1 shows the configuration of the learning device 100 according to this embodiment. The learning device 100 is an information processing device that takes an image containing multiple pixels as input and learns a machine learning model that outputs a discrimination value indicating whether or not each pixel represents a road. The learning device 100 includes, for example, a setting unit 110, a learning unit 120, and a storage unit 130. The setting unit 110 and the learning unit 120 are each realized by a hardware processor such as a CPU (Central Processing Unit) executing a program (software). Furthermore, some or all of these components may be realized by hardware (including circuitry) such as an LSI (Large Scale Integration), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array), or GPU (Graphics Processing Unit), or by the cooperation of software and hardware. The program may be stored in advance in a storage device such as the HDD or flash memory of the learning device 100 (a storage device equipped with a non-transient storage medium), or it may be stored in a removable storage medium such as a DVD or CD-ROM, and installed in the HDD or flash memory of the learning device 100 when the storage medium (non-transient storage medium) is mounted on a drive device. The storage unit 130 stores, for example, learning data 130A and machine learning model 130B. The storage unit 150 is implemented by, for example, RAM, flash memory, SD card, etc.

[0014] Figure 2 shows an example of the configuration of the training data 130A. The training data 130A is, for example, a group of pixels PX of one or more images IM, to which a label value indicating whether or not the pixel represents a road is associated. For example, in Figure 2, the area enclosed by road markings LRL and RRL represents a road area, so pixel PX1 is associated with a label value of 0, indicating that the pixel does not represent a road, while pixels PX2 to PX11 are associated with a label value of 1, indicating that the pixels represent a road. The training data 130A is generated, for example, by the administrator or operator of the learning device 100 specifying each pixel of the image IM and its label value in advance on their terminal, and stored in the storage unit 150. Alternatively, the learning device 100 may download the training data 130A stored on an external server to the storage unit 130 via a network at the timing of the learning execution by the learning unit 120. The label value is an example of a "discrimination value" in the claims.

[0015] Machine learning model 130B is a machine learning model that takes an image IM as input and outputs a label value indicating whether or not the pixels in the image IM represent a road. Machine learning model 130B is, for example, a convolutional neural network (CNN), and outputs a label value indicating whether or not the pixels represent a road by extracting the feature quantities of the pixels (or groups of pixels) in the input image IM. The image IM input to machine learning model 130B may be a raw image captured by a camera mounted on a vehicle, or it may be a planar projection image or cylindrical projection image of a road obtained by transforming the raw image.

[0016] In conventional techniques, the machine learning model is trained to reduce the value of the loss function, which is calculated based on the error between the label value output by the machine learning model and the training data 130A indicating whether each pixel represents a road, thereby generating a final trained model. Here, the loss function is defined, for example, as the absolute sum of the errors for each pixel. On the other hand, the further away the image IM is from the point where it was taken, the fewer pixels are used to represent an area of ​​the same size. For example, as shown in the right part of Figure 2, although areas A1 and A2 have the same area in a real road, area A1, which is further from the point of capture, is represented by a single pixel PX2, while area A2, which is closer to the point of capture, is represented by multiple pixels PX3 to PX11. As a result, the trained model generated by conventional techniques tends to have high output accuracy for areas of the image IM closer to the point of capture, but lower output accuracy for areas further away from the point of capture.

[0017] Against this backdrop, the setting unit 110 sets weights (for example, multiplies) for the error between the label values ​​output by the machine learning model 130B and the training data 130A indicating whether each pixel represents a road or not. More specifically, the setting unit 110 increases the weight values ​​set for the error in one or more predetermined directions from the center of the lower edge of the image IM. In other words, the setting unit 110 prevents a decrease in the accuracy of machine learning for distant areas by increasing the error values ​​for areas further from the shooting point compared to the error values ​​for areas closer to the shooting point using weights.

[0018] FIG. 3 is a diagram for explaining a method of setting weights by the setting unit 110. The setting unit 110 increases the weight from the center of the lower end of the image IM toward the upper end and the left and right ends as one or more predetermined directions. For example, as shown in the pattern of FIG. 3(a), the setting unit 110 may increase the value of the weight set for the error from the center of the lower end of the image IM toward the upper end. Or, as shown in the pattern of FIG. 3(b), the setting unit 110 may increase the value of the weight set for the error from the center of the lower end of the image IM toward the right end. Or, as shown in the pattern of FIG. 3(c), the setting unit 110 may increase the value of the weight set for the error from the center of the lower end of the image IM toward the left end. These patterns may be applied simultaneously (for example, when a pixel is located in the upper left direction of the image IM, both the weight of the pattern of FIG. 3(a) and the weight of the pattern of FIG. 3(c) may be set), or only a specific pattern may be selected and applied from these patterns. In the present embodiment, the setting unit 110 increases the value of the weight set for the error at least from the center of the lower end of the image IM toward the upper end.

[0019] After the weights are set by the setting unit 110 for the error between the label value output by the machine learning model 130B and the learning data 130A indicating whether each pixel of the image IM represents a road, the learning unit 120 learns the machine learning model 130B so as to reduce the value of a loss function (for example, the absolute sum of errors) based on the error with the weights set. The learning in this case may use any learning method such as, for example, the back propagation method. The learning unit 120 acquires the machine learning model 130B for which the learning is completed as a learned model.

[0020] The obtained learned model may be used by a driving support device or a vehicle control device mounted on a vehicle. For example, the driving support device may input a front image of the vehicle captured by a camera into the learned model and display the road area output by the learned model on the HMI. Further, for example, the vehicle control device may input a front image of the vehicle captured by a camera into the learned model and control the vehicle to travel within the road area output by the learned model.

[0021] As described above, in this embodiment, the setting unit 110 increases the value of the weight set for the error in one or more predetermined directions from the center of the lower end of the image IM, and the learning unit 120 learns the machine learning model 130B so as to reduce the value of the error function based on the error for which the weight is set. Thereby, a machine learning model that accurately detects a road area from a distant area among the areas included in the image can be generated.

[0022] [Flow of processing] Next, referring to FIG. 4, the flow of processing executed by the learning device 100 according to this embodiment will be described. FIG. 4 is a flowchart showing an example of the flow of processing executed by the learning device 100 according to this embodiment.

[0023] First, the learning unit 120 acquires an image from the learning data 130A stored in the storage unit 130 (step S100). Next, the learning unit 120 calculates an error between the correct label of the learning data 130A and the output of the machine learning model 130B for each pixel of the acquired image (step S102).

[0024] Next, the setting unit 110 sets a weight according to the distance from the center of the lower end of the image for the error calculated by the learning unit 120 (step S104). Next, the learning unit 120 calculates a loss function based on the error for which the weight is set (step S106). Next, the learning unit 120 learns the machine learning model 130B so as to reduce the value of the calculated loss function (step S108). Thereby, the processing according to this flowchart ends.

[0025] As described above, in this embodiment, a machine learning model takes an image containing multiple pixels as input and outputs a discriminant value indicating whether each pixel represents a road. A weight is set for the error between the discriminant value output by the machine learning model and the training data indicating whether each pixel represents a road. The machine learning model is trained to reduce the value of the loss function calculated based on the weighted error. This setting increases the weight from the center of the bottom edge of the image in one or more predetermined directions. This makes it possible to generate a machine learning model that accurately detects road regions from the distant regions within the image.

[0026] The embodiments described above can be expressed as follows. A memory device that stores the program, Equipped with a hardware processor, The hardware processor executes the program, A machine learning model takes an image containing multiple pixels as input and outputs a discriminant value indicating whether each pixel represents a road. A weight is set for the error between the discriminant value output by the model and the training data indicating whether each pixel represents a road. The machine learning model is trained to reduce the value of the loss function calculated based on the error with the weights set, The above setting increases the weight in one or more predetermined directions from the center of the lower end of the image. A learning device configured in such a way.

[0027] Although embodiments for carrying out the present invention have been described above using examples, the present invention is not limited in any way to these embodiments, and various modifications and substitutions can be made without departing from the spirit of the present invention. [Explanation of symbols]

[0028] 100 Learning Devices 110 Setting section 120 Learning Department 130 Storage section 130A Training Data 130B Machine Learning Model

Claims

1. A learning device for training a machine learning model that takes an image containing multiple pixels as input and outputs a discriminant value indicating whether or not each of the pixels represents a road, A setting unit sets weights for the error between the discriminant value output by the machine learning model and the training data indicating whether each pixel represents a road or not. The machine learning model is trained to reduce the value of the loss function calculated based on the error with the weights set, The setting unit increases the weight in one or more predetermined directions from the center of the lower end of the image. Learning device.

2. The setting unit increases the weight in one or more predetermined directions, from the center of the lower end of the image toward the upper end and the left and right ends. The learning device according to claim 1.

3. The setting unit increases the weight from the center of the lower end of the image toward the upper end as one or more predetermined directions. The learning device according to claim 1.

4. A learning method for training a machine learning model that takes an image containing multiple pixels as input and outputs a discriminant value indicating whether each of the pixels represents a road, wherein the computer A weight is set for the error between the discriminant value output by the machine learning model and the training data indicating whether each pixel represents a road or not. The machine learning model is trained to reduce the value of the loss function calculated based on the error with the weights set, The above setting increases the weight in one or more predetermined directions from the center of the lower end of the image. Learning methods.

5. A program for training a machine learning model that takes an image containing multiple pixels as input and outputs a discriminant value indicating whether each of the pixels represents a road, wherein the computer... A weight is set for the error between the discriminant value output by the machine learning model and the training data indicating whether each pixel represents a road or not. The machine learning model is trained to reduce the value of the loss function calculated based on the error with the weights set, The above setting increases the weight in one or more predetermined directions from the center of the lower end of the image. program.

Images