AI for Maritime Security: Comparative Evaluation of Convolutional Neural Network and Vision Transformer Architectures for Maritime Object Detection

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

Problem Statement
Maritime security faces escalating threats, including illicit trafficking, unregulated fishing, and covert vessel operations, necessitating robust surveillance capabilities. While traditional methods rely on radar, human patrols, and satellite overviews, they often struggle with real-time processing of multimodal data and detecting small or anomalous targets in challenging sea states. Although deep learning-based Computer Vision (CV) offers a pathway to automate detection and classification, significant 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 the efficacy of various deep learning architectures—specifically Convolutional Neural Networks (CNNs) and Vision Transformers (ViT)—for maritime object detection under diverse environmental conditions.

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

Six distinct deep learning architectures were evaluated:

1. Base CNN: A manually designed architecture featuring six convolutional blocks, Batch Normalization, Dropout, and a fully connected layer, serving as a controlled baseline.
2. Transfer Learning Models: Four pre-trained architectures (Xception, VGG16, MobileNetV2, and EfficientNetV2L) were fine-tuned. A critical methodological component involved adhering to model-specific preprocessing protocols (e.g., channel-wise mean subtraction for VGG16, specific normalization for Xception and MobileNetV2, and internal preprocessing for EfficientNetV2L) to maintain consistency with their original ImageNet training distributions.
3. Vision Transformer (ViT): A ViT-Base (Patch16, 224×224) model pre-trained on ImageNet was fine-tuned. This model treats image patches as sequences, utilizing self-attention mechanisms to capture global contextual dependencies rather than relying solely on local convolutional operations.

All models were trained for up to 500 epochs with early stopping. Performance was assessed using training/testing accuracy, Type I and Type II errors, model size, and video processing time on a 37-second test clip.

Key Contributions
The paper outlines four primary contributions:

• Systematic Comparison: A comprehensive evaluation of six architectures (one base CNN, four transfer learning models, and one ViT) across multiple performance metrics, moving beyond simple accuracy to include error types and computational costs.
• Methodological Framework for Preprocessing: A detailed exposition of the specific preprocessing pipelines required for different transfer learning models. The authors emphasize that inconsistent preprocessing is a common source of error in comparative studies and provide a structured approach to ensure reproducibility.
• Scenario-Based Evaluation: An analysis of model performance from a practical deployment perspective, jointly considering model size, computational requirements, and processing speed to determine suitability for edge devices versus high-capacity monitoring systems.
• Practical Guidelines: Evidence-based recommendations for model selection in maritime monitoring, distinguishing between scenarios requiring lightweight efficiency and those demanding maximum detection reliability.

Results
The experimental results indicated that model performance is highly dependent on computational constraints and deployment conditions:

• Accuracy: The Vision Transformer (ViT) achieved the highest performance, reaching 100% accuracy on both training and testing sets. It recorded the lowest error rates (1 Type I error and 3 Type II errors).
• Efficiency: Contrary to the assumption that larger models are slower, the ViT model demonstrated the fastest video processing time (30.97 seconds for the 37-second clip), outperforming even lightweight models like MobileNetV2 (34.08 seconds) and significantly faster than EfficientNetV2L (183.23 seconds).
• Model Size: The ViT model (335.2 MB) was larger than most CNNs but smaller than EfficientNetV2L (462.5 MB). MobileNetV2 remained the most compact (10 MB), suitable for resource-limited environments.
• Convergence: The ViT model showed stable convergence with near-zero loss and an Area Under the Curve (AUC) of 1, indicating perfect classification capability on the test data.

Significance and Claims
The authors claim that while lightweight architectures (e.g., MobileNetV2) are preferable for edge devices with limited computational resources, the Vision Transformer offers superior overall performance for maritime surveillance when computational capacity allows. The study asserts that ViT's ability to capture global contextual relationships via self-attention mechanisms provides a significant advantage in distinguishing vessels from complex sea backgrounds compared to traditional CNNs.

The paper concludes that no single architecture is universally optimal; the choice depends on the specific operational context. However, for systems prioritizing detection accuracy and reliability, such as centralized border protection or autonomous navigation, transformer-based architectures like ViT represent the state-of-the-art solution. The research underscores the importance of rigorous preprocessing in transfer learning and provides a validated framework for deploying AI-driven computer vision systems in real-world maritime security applications.