SFL-YOLO: an improved YOLOv11-based model for underwater object detection.

Area of Science:

• Computer Vision
• Marine Robotics
• Artificial Intelligence

Background:

• Underwater object detection faces challenges like low contrast, scattering, and noise.
• Small objects, weak textures, and blurred boundaries are difficult to detect accurately.
• Existing models struggle with preserving fine-grained details in complex underwater imagery.

Purpose of the Study:

• To develop an improved underwater object detection model, SFL-YOLO, enhancing YOLOv11n.
• To boost the representation capability for small, weak-textured, and blurred objects.
• To achieve high accuracy and real-time performance in challenging underwater conditions.

Main Methods:

• Proposed SFL-YOLO based on YOLOv11n with three novel modules: Spatial Channel Transform Convolution (SCTC), Feature-Aware Reassembly Upsampling (FARU), and Lightweight Detail Modeling Detection Head (LDMDH).
• SCTC replaces strided downsampling to preserve edge and texture information.
• FARU mitigates semantic misalignment and noise diffusion during upsampling.
• LDMDH enhances fine-grained features while reducing parameters.

Main Results:

• SFL-YOLO achieved 85.4% mAP@0.5 on URPC2020, a 1.6% improvement over YOLOv11n.
• Achieved 85.7% mAP@0.5 on RUOD, outperforming YOLOv11n by 1.1%.
• Model has 4.50M parameters, 11.4G FLOPs, and 142 FPS inference speed, demonstrating real-time capability.

Conclusions:

• SFL-YOLO effectively improves object detection accuracy in complex underwater environments.
• The proposed modules (SCTC, FARU, LDMDH) enhance the detection of challenging objects.
• The model maintains robust performance and real-time processing for underwater applications.