Learning methods, learning devices, and programs

Abstract

In this learning method, a first image is generated by applying a noise to a first region, a second image is generated by applying a noise to a second region, a composite image is generated by performing weighted addition of the first image and the second image, a first teacher label (y1) with respect to the first image is generated, a second teacher label (y2) with respect to the second image is generated, a composite teacher label (y) is generated by performing weighted addition of the first teacher label (y1) and the second teacher label (y2), and a learning model is generated by using the composite image and the composite teacher label (y) to perform machine learning.

Claims

1. A learning method performed by a computer, which generates a learning model used for image recognition, By adding noise to the first region of the original image, the first image is generated. A second image is generated by adding noise to the second region of the original image, excluding the first region. A composite image is generated by performing a weighted addition of the first image and the second image at a first ratio. A first training label for the first image is generated by performing a weighted addition of a first basic label corresponding to the correct label of the original image and a second basic label corresponding to the incorrect label of the original image, using a second ratio which is the ratio of the size of the first region to the size of the second region. By performing a weighted summation of the first basic label and the second basic label at a ratio inverse to the second ratio, a second training label for the second image is generated. A composite teacher label for the composite image is generated by performing a weighted addition of the first teacher label and the second teacher label at the first ratio. The learning model is generated by performing machine learning using the synthesized image and the synthesized teacher label. Learning methods.

2. For each of the multiple first regions, a first image, a second image, a composite image, a first teacher label, a second teacher label, and a composite teacher label are generated, thereby generating multiple composite images and multiple composite teacher labels. The learning model is generated by performing machine learning using the plurality of composite images and the plurality of composite teacher labels. The learning method according to claim 1.

3. By generating the composite image and the composite teacher label for each of the multiple first ratios, multiple composite images and multiple composite teacher labels are generated. The learning model is generated by performing machine learning using the plurality of composite images and the plurality of composite teacher labels. The learning method according to claim 1 or 2.

4. The first region is, [Math 1] Determined according to, W represents the width of the original image, and H represents the height of the original image. r ｘ１ indicates the left edge of the first region, r ｙ１ indicates the upper end of the first region, r ｘ２ indicates the right edge of the first region, r ｙ２ This indicates the lower end of the first region, a to U[b, c] indicates that a is determined according to a uniform distribution from b to c. The learning method according to claim 1 or 2.

5. The aforementioned first ratio is determined according to the beta distribution of B(α, α), B represents the beta function, α represents a positive real number. The learning method according to claim 1 or 2.

6. A learning device that generates a learning model used for image recognition, Equipped with a processor and memory, The processor uses the memory to: By adding noise to the first region of the original image, the first image is generated. A second image is generated by adding noise to the second region of the original image, excluding the first region. A composite image is generated by performing a weighted addition of the first image and the second image at a first ratio. A first training label for the first image is generated by performing a weighted addition of a first basic label corresponding to the correct label of the original image and a second basic label corresponding to the incorrect label of the original image, using a second ratio which is the ratio of the size of the first region to the size of the second region. By performing a weighted summation of the first basic label and the second basic label at a ratio inverse to the second ratio, a second training label for the second image is generated. A composite teacher label for the composite image is generated by performing a weighted addition of the first teacher label and the second teacher label at the first ratio. The learning model is generated by performing machine learning using the synthesized image and the synthesized teacher label. Learning device.

7. A program for causing a computer to execute the learning method described in claim 1 or 2.

Images