Vision Transformer beats CNNs in maritime object detection with 100% accuracy

A new study from researchers led by Ismet Gocer evaluates AI architectures for maritime security, comparing convolutional neural networks (CNNs) with Vision Transformers (ViTs) for detecting ships in diverse real-world conditions. The team used a dataset of 6,468 images covering cloudy, foggy, rainy, and sunny environments. They tested six models: a base CNN, four transfer learning models (Xception, VGG16, MobileNetV2, EfficientNetV2L), and a Vision Transformer. Performance metrics included accuracy, Type I/II errors, model size, and video processing time. Results showed the ViT achieved perfect 100% accuracy with the lowest error rates and fastest processing, while lightweight architectures like MobileNetV2 excelled for edge devices with limited computational resources.

The findings demonstrate that ViTs outperform traditional CNNs for maritime object detection when sufficient compute is available, but the trade-off between accuracy and efficiency is critical. The ViT's ability to process video frames rapidly makes it ideal for real-time surveillance and autonomous navigation systems. The study also notes that transfer learning models (Xception, VGG16, EfficientNetV2L) performed well but lagged behind ViT. This research has direct implications for maritime border protection, illegal fishing monitoring, and search-and-rescue operations, where reliable AI-driven detection under varied weather is essential.