Study of Giving Training Images in SegNet

Abstract

When using machine learning, it is necessary to prepare training data and set various parameters. In recent years, machine learning has attracted attention, and the number of cases where a person who does not have specialized knowledge or experience does machine learning has increased. SegNet, which is the target of this research, needs training images that annotated for recognition targets. Therefore, preparing training images requires an enormous amount of time and effort. Previous studies have shown data sets and parameters for learning each recognition target. However, there is no case where the investigation about how to give training images such as the number of training images and the setting of parameters required when using SegNet for the first time was conducted. The larger the number of training images, the higher the recognition accuracy can be expected, but the recognition accuracy does not necessarily increase in proportion to the number of the prepared training images. In this paper, we report the effects of the number of training images and the setting values of batch size on recognition accuracy as a way of giving training images.