Using Machine Learning to automate trash detection on a limited dataset with the YOLOv5 object detection network

Waste in the ocean is one of the most demanding environmental challenges in today’s world. Marine debris poses an extensive list of threats to aquatic life and places financial burdens on local governments that may be responsible for funding its removal. In recent years, there have been increased efforts to develop technology that can automate the process of waste removal from rivers to prevent the flow of litter into oceans. However, because the process of automatically removing trash requires highly advanced technology, these methods can be very costly to implement and are therefore only used in select areas. In order to widen the impact of financial resources, it is more effective to automate the process of trash detection and notify the appropriate authorities to take action accordingly. In this paper, we propose a mode of trash detection that uses the object detection architecture YOLOv5 to detect trash in rivers. To supplement the insufficient datasets for use in our model, we manually photographed and annotated images of surface-polluted rivers to enable us to train, validate, and test the model. Even with the lack of extensive data, the performance of our model is similar to that of other adequate waste detection networks.