Development of a Faster R-CNN-based Marine Debris Detection Model for an Embedded System
Younghun Byeon, Eunju Kim, Hyeon Jun Lim, Han Sol Kim · Journal of Institute of Control Robotics and Systems · 2021
In this study, we propose a Faster R-CNN-based marine debris detection algorithm for an embedded system. First, the trash annotations in context (TACO) dataset, which is an open image dataset of waste in the wild, is used to build a training image dataset of marine debris. To enhance the amount of our training dataset, we included images of the TACO unofficial dataset, which has not been reviewed by the TACO research team. To this end, we manually screened the appropriately annotated images from the TACO unofficial dataset. In addition, only seven most frequently discovered classes in the ocean are selected from the TACO datasets to enable efficient learning. The utilization of MobileNet as the backbone network of the proposed Faster R-CNN model enables a faster inference time compared to those of conventional models. In addition, the backbone network was fine-tuned on the TACO dataset to improve the feature extraction performance of the model. Lastly, the real-time operability of the proposed algorithm was verified by porting the model to Jetson Xavier NX.