|
内容記述 |
The emergence of deep learning, the accuracy of object recognition has been dramatically improved. Currently, upon performing object recognition by deep learning, not only computational resources such as a computer but also a lot of learning data are required. In particular, it has been found that even if this learning data is learned by the same procedure using the same neural network, it significantly affects the learning result (correct answer rate of recognition) due to a bias in the number of data. In addition, although learning data that can be used for deep learning is being disclosed on the Internet, learning data available for a task to be recognized is not always disclosed. In that case, it is necessary to create the data used for learning manually. But if the collected learning data is small, the method to inflate the learning data by parallel shift, rotation, inversion, deterioration processing of the image, etc. has been conventionally proposed. However, all of these methods are mathematically determined processes, and if they are used frequently, learning may be hindered. In this research, we focused on generative adversarial networks and examined whether learning data to improve recognition accuracy could be generated with this. |
3_2020_31-38.pdf (1.3 MB)
