Swin Transformer-Based Framework for Floating Marine Debris Detection and Classification

This paper presents an advanced deep learning approach for detecting and classifying floating debris in aquatic environments, with a particular focus on challenging accumulation zones such as rivers, shorelines, and harbor areas. We explore the Swin Transformer (Swin-T) architecture and conduct systematic hyperparameter optimization to maximize its capability in capturing both local and global visual features. Through extensive Bayesian optimization and fine-tuning, we benchmark the optimized Swin-T against state-of-the-art convolutional neural networks, including ResNet-50, InceptionV3, VGG-19, and Vision Transformers (ViT), using a newly curated dataset. To address data limitations, we compiled a comprehensive dataset by augmenting existing public collections with newly captured RGB images obtained under diverse hydrodynamic and lighting conditions. Our optimized Swin-T model achieves 98.5% classification accuracy and 52.3% mAP in debris detection, representing significant improvements over both baseline transformers and conventional CNNs. The study demonstrates that systematic hyperparameter optimization substantially enhances transformer-based models for marine debris monitoring, offering a more robust and scalable solution for environmental protection efforts.