The culture of unfiltered waste disposal results in mixed and contaminated waste streams, significantly complicating recycling efforts. This practice impedes the sorting process and reduces material quality, making efficient recycling challenging. In this work, we investigated various versions of the YOLO object detection model to identify waste in industrial belt scenarios, utilizing the WaRP dataset for its comprehensive annotations and realistic representation of waste scenarios. We conducted several experiments leveraging sequential transfer learning, hyperparameter tuning, and data augmentation to enhance YOLO models’ waste detection capabilities. The YOLOv9c model achieved the highest performance, with a precision of 75.47%, recall of 63.12%, mAP(50) of 67.22%, and mAP(50–95) of 54.74%, significantly outperforming the baseline YOLOv8m model. Finally, we employed Quantization Aware Training to optimize the best model for deployment, which yielded promising results with a precision of 71.32%, recall of 59.36%, mAP(50) of 63.83%, and mAP(50–95) of 50.93%. This multi-stage approach significantly enhances the accuracy and efficiency of waste detection for sorting facilities, contributing to more effective and sustainable waste management practices in industrial environments.