Underwater Plastic Waste Detection with YOLO and Vision Transformer Models

• Autores: Cárdenas Rondoño, Jonathan Bruce, Vasquez Espinoza, Ners Armando, Escobedo Cárdenas, Edwin Jonathan
• Formato: artículo
• Fecha de Publicación: 2025
• Institución: Universidad de Lima
• Repositorio: Revistas - Universidad de Lima
• Lenguaje: inglés
• OAI Identifier: oai:ojs.pkp.sfu.ca:article/7868
• Enlace del recurso: https://revistas.ulima.edu.pe/index.php/Interfases/article/view/7868
• Nivel de acceso: acceso abierto
• Materia: object detection; deep learning; plastic waste; object detection model; underwater images; detección de objetos; aprendizaje profundo; residuos plásticos; modelos de detección de objetos; imágenes submarinas

This study addresses the global issue of marine pollution, with a particular focus on plastic bag contamination, by leveraging real-time object detection techniques powered by deep learning algorithms. A detailed comparison was carried out between the YOLOv8, YOLO-NAS, and RT-DETR models to assess their effectiveness in detecting plastic waste in underwater environments. The methodology encompassed several key stages, including data preprocessing, model implementation, and training through transfer learning. Evaluation was conducted using a simulated video environment, followed by an in-depth comparison of the results. Performance assessment was based on critical metrics such as mean average precision (mAP), recall, and inference time. The YOLOv8 model achieved an mAP50 of 0.921 on the validation dataset, along with a recall of 0.829 and an inference time of 14.1 milliseconds. The YOLO-NAS model, by contrast, reached an mAP50 of 0.813, a higher recall of 0.903, and an inference time of 17.8 milliseconds. The RT-DETR model obtained an mAP50 of 0.887, a recall of 0.819, and an inference time of 15.9 milliseconds. Notably, despite not having the highest mAP, the RT-DETR model demonstrated superior detection performance when deployed in real-world underwater conditions, highlighting its robustness and potential for practical environmental monitoring.