Real-Time Waste Detection System Using YOLOv12 with Transfer Learning

Waste sorting at the source remains a major challenge in Indonesia due to limited public awareness and the absence of accessible tools for waste classification. While YOLO-based object detection has been widely applied for waste detection, the adoption of the latest YOLO architecture in web-based, real-time public-oriented systems remains limited. This study aims to develop and experimentally evaluate a web-based waste detection system using YOLOv12 with a transfer learning approach to classify waste into organic, inorganic, and hazardous (B3) categories along with their subcategories. The system was developed using the Flask framework and supports image upload and real-time camera-based detection. A real-world dataset was annotated and divided into training, validation, and testing sets for experimental evaluation. The proposed model achieved a precision of 0.86, recall of 0.74, mAP@0.5 of 0.83, and mAP@0.5:0.95 of 0.68, with an average inference time of 0.0187 seconds per image (53.40 FPS). Overall, these results indicate that YOLOv12 with transfer learning provides an effective balance between accuracy and inference speed for web-based real-time waste detection systems, supporting its applicability for practical waste sorting solutions.