CNNs vs. transformers: A benchmark for multi-class marine debris identification.

Area of Science:

• Environmental Science
• Computer Science
• Artificial Intelligence

Background:

• Marine debris, predominantly plastic waste, poses significant threats to marine ecosystems, biodiversity, and human economies.
• Effective classification of marine debris is crucial for mitigation and management strategies.

Purpose of the Study:

• To evaluate and compare the performance of deep learning models, including Convolutional Neural Networks (CNNs) and Vision Transformer (ViT), for classifying marine debris.
• To identify the most effective model for fine-grained classification of surface marine debris.

Main Methods:

• A custom dataset of marine debris images was utilized.
• Performance of InceptionV3, ResNet-50, and VGG-16 (CNNs) was compared against the Vision Transformer (ViT-Base/16).
• The study involved robust preprocessing, feature extraction, and evaluation metrics to analyze model performance.

Main Results:

• InceptionV3 and ResNet-50 achieved classification accuracies of 99.32% and 98.80%, respectively.
• Vision Transformer (ViT-Base/16) demonstrated comparable performance, achieving 98.80% accuracy.
• The performance differences were marginal, requiring statistical significance testing for definitive conclusions.

Conclusions:

• Deep learning models show high efficacy in classifying marine debris into fine-grained categories.
• Vision Transformer (ViT-Base/16) presents a promising alternative to conventional CNN architectures for marine debris classification.
• Further research is recommended for large-scale and real-world deployment of ViT models for marine debris detection.