sagar-sonar-detector

SAGAR SonarVision - Underwater Anomaly Detection

SAGAR (SonarVision) is a state-of-the-art YOLOv8-based object detection model optimized for Side-Scan Sonar (SSS) imagery. It is designed to autonomously detect and classify underwater anomalies and objects of interest with high precision.

This repository hosts the ONNX exported version of the model, enabling highly efficient, cross-platform inference on edge devices, autonomous underwater vehicles (AUVs), and marine monitoring stations.

📊 Model Details

• Architecture: YOLOv8 (Ultralytics)
• Format: ONNX (.onnx)
• Input Size: 640x640 pixels
• Training Epochs: 100
• Primary Use Case: Underwater Sonar Anomaly Detection

🎯 Detected Classes

The model is trained to detect the following 5 distinct underwater classes:

1. human (Divers / Swimmers)
2. ghost_net (Abandoned fishing nets)
3. submarine_pipeline (Underwater infrastructure)
4. shipwreck (Sunken vessels)
5. mine_cylinder (Unexploded ordnance / cylindrical objects)

---

📈 Evaluation Metrics

The model was rigorously evaluated on a synthetic and real-world side-scan sonar validation set.

Class-Specific mAP50 Breakdown:

• Ghost Net: 0.995
• Submarine Pipeline: 0.994
• Human: 0.752
• Shipwreck: 0.545
• Mine Cylinder: 0.533

---

💻 How to Use (Inference)

You can easily run inference on this ONNX model using the ultralytics package or raw onnxruntime in Python.

Prerequisites

pip install ultralytics onnxruntime huggingface_hub

Python Snippet (Ultralytics API)

The easiest way to perform inference and utilize automatic non-maximum suppression (NMS):

from ultralytics import YOLO
from huggingface_hub import hf_hub_download

# 1. Download the model from Hugging Face
model_path = hf_hub_download(repo_id="Narayan-nkj/sagar-sonar-detector", filename="sonar_detector.onnx")

# 2. Load the ONNX model
model = YOLO(model_path, task='detect')

# 3. Run Inference on a sonar image
results = model.predict("path/to/your/sonar_image.jpg", conf=0.50)

# 4. Display Results
for result in results:
    result.show()  # Displays the image with bounding boxes

---

⚠️ Limitations & Intended Use

• Acoustic Shadows: The model relies heavily on the acoustic shadow signatures present in side-scan sonar. Performance may degrade if the sonar altitude or grazing angle drastically alters the shadow profiles.
• Environmental Noise: High sea states, thermoclines, or excessive water column noise may induce false positives. A confidence threshold of >= 0.60 is recommended for production deployments.

🤝 Citation & Acknowledgements

Developed as part of the NetraSonar project for autonomous underwater surveillance and environmental cleanup initiatives.