Small Waste Detection on Land Using fine-tuned YOLOv9 in Unmanned Aerial Vehicle Images

Trash collection in inaccessible areas can be labor-intensive and requires excessive resources. Using Unmanned Aerial Vehicles (UAVs) to gather high-resolution imagery for automated environmental monitoring is one such approach that has shown a lot of potential in this arena. However, due to the significant variation in their attributes and the background scenarios, it becomes a challenging task to discover and classify small-sized waste materials that are also highly variable in UAV photos. In this study, deep learning-based different You Only Look Once (YOLO) models have been exploited to enhance UAV-based small object trash detection. A crucial contribution of the research is the evaluation of different novel and state-of-the-art activation functions aiming to address common limitations of legacy ReLU and sigmoid functions along with the effect of data augmentation on the model’s performance. The best model proven to be efficient at detecting trash on land after applying Data Augmentation is YOLOv9 with Swish activation function. It achieved the best mean Average Precision (mAP) of 71% approximately which was increased by 4% Furthermore, it is found that the best-performing activation function is Swish due to its model sensitivity and accuracy.