Solid Waste Detection with YOLOv9 Using Preprocessing Techniques and Experimental Evaluation, Improving Generalization to Real-World Data via a Data-Centric Preprocessing Pipeline

Advancements in deep learning have enabled the development of computer vision models for automating complex tasks. In the domain of solid waste management, these models have been implemented to improve the efficiency of solid waste detection. This study presents a deep learning-based method to automate the detection of solid waste while addressing the significant performance gap between controlled and real-world environments. A YOLOv9 model was developed to accurately classify waste into five categories. The model was trained on the TrashNet benchmark dataset and evaluated on a custom-collected real-world validation set to measure its generalization capabilities. To improve performance, a comprehensive data preparation pipeline was implemented. This pipeline included the integration of background negative samples, the development of a custom copy-paste data synthesis process for contextual diversity, and extensive offline pre-augmentation. The baseline model, while achieving 97.9% [email protected] on the benchmark, performed poorly on the real-world set with only 24.5% [email protected]. The final selected YOLOv9c model obtained a final [email protected] of 91.7% on the real-world validation set. The results of the study demonstrate that a comprehensive, data-centric approach is crucial for addressing the domain gap. To further improve model performance, it is recommended that the diversity of the training data be expanded and that advanced augmentation techniques continue to be explored.