AI for Maritime Security: Comparative Evaluation of CNN and Vision Transformer Architectures for Maritime Object Detection

Technical Summary: AI for Maritime Security – Comparative Evaluation of CNN and Vision Transformer Architectures

Problem Statement
Maritime security faces escalating threats, including illicit trafficking, unregulated fishing, and covert vessel operations. While traditional surveillance relies on radar, human patrols, and satellite overviews, these methods often struggle with real-time processing, small target detection, and varying environmental conditions. Although deep learning-based Computer Vision (CV) offers a pathway to automate detection and tracking, challenges remain regarding data scarcity, domain adaptation, and the deployment of complex models on resource-constrained platforms. This study addresses the need to evaluate and compare different deep learning architectures to determine their suitability for real-world maritime object detection under diverse weather conditions.

Methodology
The research utilized a custom maritime image dataset comprising 6,468 images labeled for "Object" (ship) and "Background," covering cloudy, foggy, rainy, and sunny environments. The dataset was split into 70% training, 15% validation, and 15% testing, with data augmentation applied exclusively to the training set.

Six deep learning architectures were evaluated:

1. Base CNN: A manually designed Convolutional Neural Network with six convolutional blocks, batch normalization, dropout, and a sigmoid output for binary classification.
2. Transfer Learning Models: Four pre-trained architectures (Xception, VGG16, MobileNetV2, and EfficientNetV2L) initialized with ImageNet weights. The study emphasized strict adherence to model-specific preprocessing protocols (e.g., specific normalization ranges and channel ordering) to ensure compatibility with original training distributions.
3. Vision Transformer (ViT): A ViT-Base (Patch16, 224×224) model fine-tuned on the maritime dataset, leveraging self-attention mechanisms to capture global contextual dependencies.

Performance was assessed using training/testing accuracy, Type I (false positive) and Type II (false negative) errors, model size (MB), and computational processing time on a 37-second video sequence.

Key Contributions
The paper outlines four primary contributions:

1. Systematic Architectural Comparison: A comprehensive evaluation of six distinct models (base CNN, four transfer learning variants, and ViT) using a multi-metric framework that includes error types and computational efficiency, not just accuracy.
2. Methodological Framework for Preprocessing: A detailed, reproducible guide on applying model-specific preprocessing pipelines in transfer learning, addressing a common gap in literature where preprocessing is often insufficiently described or inconsistently applied.
3. Scenario-Based Evaluation: An analysis of model performance from a practical deployment perspective, jointly considering model size, processing speed, and error rates to determine suitability for edge devices versus high-capacity systems.
4. Practical Guidelines: Evidence-based recommendations for model selection in maritime monitoring, distinguishing between lightweight needs for resource-constrained environments and high-accuracy needs for centralized surveillance.

Results
The experimental results indicated significant performance variations based on architectural design and computational constraints:

• Vision Transformer (ViT): Achieved the highest overall performance with 100% accuracy on both training and testing sets. It recorded the lowest error rates (1 Type I, 3 Type II errors) and the fastest video processing time (30.97 seconds for a 37-second clip).
• Transfer Learning Models:
  • Xception and VGG16 demonstrated strong performance (99–100% accuracy) with moderate processing times (~32 seconds).
  • MobileNetV2 offered the smallest model size (10 MB) and reasonable speed (34.08 seconds) but exhibited slightly higher error rates compared to ViT.
  • EfficientNetV2L, despite having the largest parameter count (117M) and model size (462.5 MB), resulted in the slowest processing time (183.23 seconds).
• Base CNN: Performed well (99% accuracy) with a moderate model size (23.2 MB) but was slower than the ViT and Xception models.

Significance and Claims
The paper claims that while no single architecture is universally optimal, the Vision Transformer (ViT) offers superior reliability and speed for maritime object detection when computational resources are sufficient. The study highlights that transformer-based architectures effectively capture global contextual relationships, providing an advantage in complex maritime scenes with varying scales and lighting.

Conversely, the authors note that lightweight architectures like MobileNetV2 remain valuable for edge-based systems with limited hardware capacity, despite potentially higher error rates. The research underscores that model selection must be guided by a trade-off between detection accuracy, computational efficiency, and operational constraints. By providing a structured preprocessing framework and a multi-metric evaluation, the study aims to improve the reproducibility and practical deployment of AI-driven maritime surveillance systems.