A Study on Autonomy of Image Segmentation Model by Reinforcement Learning

Abstract

Management related to social infrastructure covers a wide range, such as daily monitoring and maintenance of river and road facilities, grasping the situation in the event of a disaster, and providing evacuation information. As a supporting technology, we are developing an image segmentation model using deep learning, such as a water surface detection model. However, since these models mainly learn camera images of limited locations, there are many cases where the accuracy decreases when applied to other locations, and they lack versatility. In addition, when re-learning for improvement, there are problems such as creating a huge amount of training data and tuning work of hyperparameters. In this study, we developed a model that autonomously continuously learns for new image data by utilizing reinforcement learning (Deep-Q-Network). As a result, we showed the possibility that the accuracy of the model can be secured at the same level as the conventional supervised learning.