Samyukta31/sonar_yolo

SONARINTEL — Side-Scan Sonar Object Detection

Model Summary

SONARINTEL is an AI-powered object detection system designed to detect underwater objects and anomalies in Side-Scan Sonar (SSS) imagery.

The project focuses on automated sonar-image analysis for applications such as marine debris detection, underwater inspection, shipwreck detection, and identification of anomalous sonar contacts.

The primary model is a YOLO11M teacher model trained to detect five underwater target categories. A smaller YOLO11N student model was also produced using knowledge distillation for lightweight inference and deployment.

---

Model Details

Property	Details
Architecture	YOLO11M
Task	Object Detection
Input Size	640 × 640
Number of Classes	5
Training Experiment	EXP-01
Framework	Ultralytics YOLO
Ultralytics Version	8.4.142
PyTorch Version	2.11.0+cu128
GPU	NVIDIA RTX 4060 Laptop GPU 8 GB
Batch Size	4
Random Seed	42
Early Stopping Patience	15

---

Detection Classes

The model detects five underwater object and anomaly categories:

ID	Class
0	crab_pot
1	submarine_pipeline
2	shipwreck
3	ghost_net
4	mine_like_contact

---

Training Dataset

The SONARINTEL dataset combines multiple sources of Side-Scan Sonar imagery to create a unified multi-class object-detection dataset.

Primary Dataset

AI4Shipwrecks

Used as the primary annotated sonar dataset.

Secondary Sources

SCTD

Used as an additional Side-Scan Sonar source containing annotated sonar targets.

KLSG

Used as an additional sonar imagery source containing ship and aircraft imagery.

The different sources were quality-checked, processed, and organized into a common dataset structure for YOLO-based object detection.

---

Data Preprocessing

The preprocessing pipeline was designed to provide consistent model input while preserving relevant sonar target characteristics.

Main steps include:

• Image quality checking
• Image validation
• Intensity normalization
• Image tiling and resizing
• YOLO-format annotation preparation
• Multi-source dataset organization
• 640 × 640 input preparation

---

Training Procedure

The primary model was trained using Ultralytics YOLO11M.

Parameter	Value
Model	YOLO11M
Image Size	640 × 640
Batch Size	4
Random Seed	42
Early Stopping Patience	15
GPU	NVIDIA RTX 4060 Laptop GPU 8 GB
CUDA	12.8
PyTorch	2.11.0+cu128
Ultralytics	8.4.142

The training experiment is identified as EXP-01.

---

Evaluation Results

YOLO11M Teacher — EXP-01

The primary YOLO11M teacher model achieved the following validation results:

Metric	Score
Precision	0.8275
Recall	0.7286
mAP@50	0.6736
mAP@50–95	0.5816

These metrics represent the validation performance recorded for the EXP-01 YOLO11M teacher model.

---

Knowledge Distillation

To obtain a more lightweight model, SONARINTEL uses knowledge distillation to transfer useful detection knowledge from the larger YOLO11M teacher to a smaller YOLO11N student.

The process can be summarized as:

YOLO11M Teacher → Knowledge Transfer → YOLO11N Student

The teacher provides the reference detection behavior, while the smaller student model is designed to require fewer computational resources and provide a more lightweight deployment option.

YOLO11N Distilled Student

The distilled YOLO11N model achieved the following evaluation results:

Metric	Score
Precision	0.9020
Recall	0.8172
F1	0.8575
mAP@50	0.8484
mAP@50–95	0.8080
Parameters	2,590,815
Mean Inference	17.78 ms
P95 Inference	20.89 ms
Peak VRAM	848 MiB

The distilled student is provided as:

best_distilled_yolo11n

The student model provides a lightweight alternative to the larger YOLO11M architecture.

Note: Teacher and student metrics should be interpreted according to their respective evaluation runs and validation configurations.

---

Available Model Files

The repository provides multiple model formats for research and deployment:

File	Description
best.pt	Trained YOLO11M teacher in PyTorch format
best_fp32.onnx	YOLO11M exported in ONNX FP32 format
best_fp16.onnx	YOLO11M exported in ONNX FP16 format
best_distilled_yolo11n	Distilled YOLO11N student model

PyTorch

best.pt is the primary trained YOLO11M model and can be used with the Ultralytics framework for inference and further development.

ONNX

The FP32 and FP16 exports provide deployment-oriented formats for environments supporting ONNX Runtime or compatible inference frameworks.

Distilled Student

best_distilled_yolo11n is the smaller YOLO11N student model produced through knowledge distillation.

---

Inference

Using Ultralytics

from ultralytics import YOLO

model = YOLO("best.pt")

results = model.predict(
    source="sonar_image.jpg",
    imgsz=640,
    conf=0.25
)

for result in results:
    result.show()

The confidence threshold can be adjusted according to the intended application and validation requirements.

---

Intended Use

SONARINTEL is intended for research and development involving:

• Side-Scan Sonar image analysis
• Underwater object detection
• Marine debris detection
• Ghost fishing net detection
• Shipwreck detection
• Submarine pipeline detection
• Sonar anomaly detection
• Automated underwater inspection

The model is intended to assist human analysis, rather than replace expert interpretation.

---

Limitations

Performance may vary depending on:

• Sonar sensor characteristics
• Acquisition conditions
• Underwater environment
• Image quality
• Sonar noise and clutter
• Acoustic shadows and reflections
• Target appearance
• Differences between training and deployment domains

The model may produce false positives and false negatives, particularly on sonar imagery that differs significantly from the training data.

The reported teacher metrics represent the validation performance of EXP-01 YOLO11M. The student metrics represent the corresponding evaluation of the distilled YOLO11N model.

---

Safety Considerations

SONARINTEL is a research and decision-support system.

Model predictions should be treated as automated detections rather than definitive ground truth.

The model should not be used as the sole basis for:

• Safety-critical decisions
• Navigation-critical decisions
• Autonomous operational decisions
• Definitive identification of real-world underwater hazards

Appropriate human review and operational validation should be performed before using predictions in real-world applications.

---

Project Information

Project: SONARINTEL

Full Project Title:

AI-Powered Automated Underwater Marine Debris and Anomaly Detection using Side-Scan Sonar

Primary Model: YOLO11M
Student Model: Distilled YOLO11N
Task: Side-Scan Sonar Object Detection
Framework: Ultralytics YOLO
Training Experiment: EXP-01

---

Disclaimer

This model is provided for research and development purposes.

Model predictions may contain errors and should be validated for the intended deployment environment before operational use.