A Lightweight Underwater Instance Segmentation Method Based on YOLOv8 and RFAHead

Abstract

Instance segmentation is vital for unmanned systems and robotics, enabling precise object identification and interaction. In underwater environments, especially for Autonomous Underwater Vehicles (AUVs), challenges like poor visibility and limited computational resources complicate real-time applications. This paper presents a lightweight instance segmentation method combining YOLOv8_seg and Region Feature Aggregation Head (RFAHead)., designed to overcome these challenges. This method enhances feature extraction and aggregation, achieving accurate segmentation with optimized performance in diver recognition, wrecks and ruins identification. Tested on the NVIDIA Jetson TX2, it delivers 7.6 frames per second, significantly improving AUV capabilities in underwater exploration.