Design a Robust Real-Time Trash Detection System Using YOLOv5 Variants

The amount of garbage produced by individuals living the modern lifestyle is significantly higher than in the past. To preserve the planet, it is necessary to recycle this garbage; one of the first steps is to classify the garbage. Individuals have done the majority of the classification so far. Thus, automating this process will make it much faster and safer, and more trash will be recycled rather than polluting the environment. To detect and separate trash automatically, a large, balanced, and labelled dataset is required, which is currently not publicly available. As a result, a labelled and balanced trash dataset was collected for this study. The dataset includes 2,500 images of trash from five different categories (glass, paper, plastic, metal, and cardboard). Furthermore, the dataset was used to create a robust trash detection model using YOLOv5, a well-known real-time object detection algorithm. As a result, various YOLOv5 variant models were created. On the collected dataset, the model achieved a precision of 90.6%, a recall of 83.3%, a mAP (0.5) of 90.8%, and a mAP (0.5, 0.95) of 81.08%. Furthermore, to demonstrate the validity of the proposed system, a collection of 500 new images was collected and then external testing was performed to evaluate the system. The system produced the following results: 80.8%, 79.2%, and 74.2%, with confidence levels of 0.3, 0.5, and 0.7, respectively.