YOLOv8-MU: An Improved YOLOv8 Underwater Detector Based on a Large Kernel Block and a Multi-Branch Reparameterization...

Area of Science:

• Computer Vision
• Marine Biology
• Robotics

Background:

• Accurate underwater target recognition is vital for marine exploration and monitoring.
• Existing deep learning models face challenges in underwater environments due to factors like low visibility and complex backgrounds.
• There is a growing demand for more robust and accurate underwater visual detection technologies.

Purpose of the Study:

• To develop an innovative deep learning architecture, YOLOv8-MU, for enhanced underwater visual detection.
• To improve the accuracy, robustness, and generalization capabilities of underwater target recognition models.
• To address specific challenges in underwater organism detection, such as localization accuracy and boundary clarity.

Main Methods:

• The proposed YOLOv8-MU architecture integrates the large kernel block (LarK block) for an optimized backbone.
• It incorporates C2fSTR (Swin transformer with C2f module) and SPPFCSPC_EMA (SPPFCSPC with attention) for improved feature extraction.
• A fusion block from DAMO-YOLO enhances multi-scale feature extraction, and MPDIoU loss optimizes localization accuracy.

Main Results:

• YOLOv8-MU achieved an mAP@0.5 of 78.4% on the URPC2019 dataset, a 4.0% improvement over YOLOv8.
• The model reached 80.9% on URPC2020 and 75.5% on the Aquarium dataset, outperforming YOLOv5 and YOLOv8n.
• On an improved URPC2019 dataset, YOLOv8-MU demonstrated state-of-the-art performance with an mAP@0.5 of 88.1%.

Conclusions:

• YOLOv8-MU significantly enhances underwater visual detection accuracy and robustness.
• The model exhibits strong generalization capabilities across diverse underwater datasets.
• The proposed architectural improvements and loss function offer a superior solution for marine exploration and monitoring applications.