Deep Learning based Facility Drain Inspection using Class Balancing with Synthesized Image Generation

While aging facilities are rapidly increasing, there is a problem that safety management is disrupted due to insufficient manpower and time required for facility inspection. Currently, research on the development of a system that inspects facilities based on deep learning algorithms is increasing. In this paper, we propose an object-detection based inspection algorithm for gravity and siphonic drains on the roof of industrial complexes. It is difficult to find appropriate abnormal class samples for training. Therefore, a problem of class imbalance arises between normal and abnormal samples. To solve this problem, we generate abnormal class synthetically using image in-painting technique. The results show the effectiveness of approach used in this project.