Solar Panel Defect Detection

YOLOv11-based object detection model for solar panel surface anomaly detection. Identifies 6 defect categories in RGB images. Trained on a custom labeled dataset using both RGB and thermal modalities; this repo contains the RGB variant.

GitHub: 4keles/Solar-Panel-AI-Analysis

Classes

Class	Description
`bird_drop`	Bird dropping contamination
`bird_feather`	Feather debris on panel surface
`physical_damage`	Cracks, chips, physical panel damage
`dust_partical`	Dust and particle contamination
`leaf`	Leaf debris
`snow`	Snow coverage

Performance — v1.2.1 (test split)

Metric	Value
mAP@50	0.546
mAP@50-95	0.241
Precision	0.569
Recall	0.582
F1	0.575

Per-Class Breakdown

Class	mAP@50	Precision	Recall
bird_feather	0.995	0.832	1.000
leaf	0.752	0.668	0.813
physical_damage	0.552	0.543	0.565
snow	0.467	0.567	0.494
dust_partical	0.408	0.590	0.373
bird_drop	0.100	0.214	0.246

bird_drop performance is low due to limited labeled samples in the dataset — planned improvement in v1.3.

Model Versions

Version	Format	Size	Notes
`v1.2.1/best.onnx`	ONNX	37.9 MB	Recommended — CPU/GPU portable
`v1.2.1/best.pt`	PyTorch	6 MB	Fine-tuning / training
`thermal-v1.0.4/best.onnx`	ONNX	37.9 MB	Thermal camera variant

Usage

Download (Python)

from huggingface_hub import hf_hub_download

model_path = hf_hub_download(
    repo_id="4keles/solar-panel-od",
    filename="v1.2.1/best.onnx"
)

Or use the project download script:

python scripts/download_model.py --version v1.2.1

Inference

from ultralytics import YOLO

model = YOLO("best.onnx", task="detect")
results = model.predict("solar_panel.jpg", conf=0.25)
results[0].show()

Class names (ordered)

CLASSES = ["bird_drop", "bird_feather", "physical_damage", "dust_partical", "leaf", "snow"]

Hardware

Trained on NVIDIA GeForce RTX 3050 Laptop GPU (4 GB VRAM). ONNX export runs on CPU or any CUDA device without recompilation.

License

MIT

Downloads last month: 255