Energy-Efficient and Adversarially Resilient Underwater Object Detection via Adaptive Vision Transformers

Area of Science:

• Computer Vision
• Marine Technology
• Artificial Intelligence

Background:

• Underwater object detection faces challenges like optical degradation, high energy use, and adversarial threats.
• Existing methods struggle with image quality and computational efficiency in diverse marine environments.

Purpose of the Study:

• To develop an Adaptive Vision Transformer (A-ViT)-based framework for enhanced underwater object detection.
• To improve image quality, detection accuracy, and system efficiency for marine applications.
• To enhance system resilience against adversarial perturbations and ensure operational feasibility.

Main Methods:

• Implemented a power-modeling and endurance-estimation scheme for hardware feasibility.
• Utilized Hybrid Attention Transformer (HAT) for super-resolution and DICAM for staged image enhancement.
• Employed an improved YOLOv11-Coordinate Attention-High-order Spatial Feature Pyramid Network (YOLOv11-CA_HSFPN) for detection.
• Integrated an Adaptive Vision Transformer (A-ViT) with Region of Interest (ROI) pooling for efficiency.
• Introduced an Image-stage Attack QuickCheck (IAQ) module for defense against adversarial attacks.

Main Results:

• Achieved significant improvements in image quality metrics: PSNR (+74.8%), SSIM (+375.8%), UIQM (3.00 to 3.85), and UCIQE (0.550 to 0.673).
• YOLOv11-CA_HSFPN reached 56.2% mAP@0.5, surpassing the baseline YOLOv11 by 1.5%.
• A-ViT + ROI reduced inference latency by 27.3% and memory usage by 74.6% with YOLOv11-CA_HSFPN.
• Demonstrated up to 48.9% latency reduction and 80.0% VRAM savings with other detectors.
• The IAQ module reduced adversarial-attack-induced latency growth by 33-40%.

Conclusions:

• The proposed A-ViT framework effectively addresses key challenges in underwater object detection.
• The system demonstrates superior performance in image enhancement, detection accuracy, and computational efficiency.
• The framework offers robust defense mechanisms against adversarial attacks, ensuring reliable operation in critical marine missions.