Skip to content

Limitations With Existing Approaches for Undersea Imaging

1. Multi-Scale Marine Object Detection in Side-Scan Sonar Images

Model: BES-YOLO (modified YOLOv5 optimized for side-scan sonar)

Dataset: 1,584 labeled images across five categories

  • Pipeline
  • Shipwreck
  • Rock
  • Seabed mound
  • Anchor

Performance:

  • mAP0.5: 92.4%

  • Baselines:

    • YOLOv8n: 87.1%
    • YOLOv5s: 71.4%

2. Underwater Garbage Detection with Improved YOLOv5

Model: Enhanced YOLOv5 for marine debris detection

Dataset: Trash_ICRA19

  • 5,720 training images
  • 820 testing images
  • Categories: plastic debris, ROV/man-made objects, biological material

Performance:

  • Accuracy: 88.7%

  • mAP: 90.6%

3. Rethinking General Underwater Object Detection (RUOD)

Model: YOLOF baseline

Dataset: Large-scale, diverse underwater imagery

  • 14,000 high-resolution images
  • 74,903 labeled objects
  • Ten classes (fish, turtle, diver, coral, jellyfish, scallop, cuttlefish, etc.)

Performance:

  • mAP0.5:0.95: 50.1%

Common Limitation

Most existing datasets are narrowly scoped—either focusing on specific object types or constrained by limited resolution or size—which hampers model generalizability.

Potential Improvements

  1. Optical Imaging Expansion Merge multiple optical datasets (e.g., Trash_ICRA19, RUOD) to create a broad, diverse training corpus capable of detecting varied underwater anomalies.

  2. Sonar-Based Extension Enhance the BES-YOLO framework by aggregating additional side-scan sonar data—ideal for seabed inspection scenarios where optical imagery is unavailable.