Solar Panel Defect Detection — YOLO11 Multi-Modal

End-to-end solar panel anomaly detection: 6-class fault classification from RGB and thermal imagery, with a PyQt6 live-stream dashboard and an active learning annotation pipeline.

---

Demo

Live stream detection with dynamic inference and FPS controls.

Dashboard user interface and thermal mode.

---

Features

• 6-class fault detection — physical damage, dust particles, bird drops, bird feathers, leaf, snow
• Multi-modal support — RGB and thermal camera input with live mode toggle and thermal contrast/brightness controls
• ONNX + TensorRT export — ONNX for CPU/GPU portability (37.9 MB); TensorRT .engine for edge deployment
• PyQt6 live-stream dashboard — multi-source input (USB cam, IP cam, RTSP/RTMP, MP4), ByteTrack object tracking, clean recording without OSD overlay
• Active Learning pipeline — Label Studio integration for human-in-the-loop annotation from drone footage
• Modular design — inference, streaming, training, and annotation pipelines are independent and composable

---

Architecture

Input (RGB / Thermal camera, RTSP, MP4)
            │
            ▼
    Pre-processing  ◄── configs/
            │
            ▼
  YOLO11 Inference ──── models/v1.2.1/best.onnx  (37.9 MB)
            │
     ┌──────┴──────┐
     │             │
     ▼             ▼
  PyQt6         scripts/
  Dashboard     (train / augment / active-learn / export)
     │
     ▼
  output/  (recordings, snapshots — no OSD overlay)

---

Installation

Requirements: Python 3.10+, CUDA 12.4, uv

git clone https://github.com/4keles/Solar-Panel-AI-Analysis.git
cd Solar-Panel-AI-Analysis

uv sync

Environment variables:

cp .env.example .env
# Edit .env — set HF_TOKEN if downloading from a private HF repo,
# set LABEL_STUDIO_URL if using the active learning pipeline

TensorRT (optional): uv sync installs the Python bindings (tensorrt-cu12), but the runtime must be installed system-wide. Follow the NVIDIA TensorRT install guide for CUDA 12.4.

---

Quick Start

1 — Download the model

uv run python scripts/download_model.py --version v1.2.1 --format onnx
# Downloads best.onnx (37.9 MB) into models/v1.2.1/

All model versions are hosted on HuggingFace Hub.

2 — Launch the dashboard

uv run python streaming/main.py

Select a source (webcam, MP4, or RTSP URL) and model version from the sidebar. Switch to Thermal mode for infrared input.

3 — Headless inference (scripted)

from ultralytics import YOLO

model = YOLO("models/v1.2.1/best.onnx")
results = model.predict("path/to/image.jpg", conf=0.25)
results[0].show()

4 — Export to TensorRT

uv run python scripts/export_engine.py --model models/v1.2.1/best.pt --imgsz 640
# Outputs models/v1.2.1/best.engine (CUDA 12.4 specific)

---

Model Performance — v1.2.1

Evaluated on held-out test split. Source: reports/v1.2.1/val_summary.json.

Class	mAP@50	mAP@50-95	Precision	Recall
Overall	0.546	0.241	0.569	0.582
bird_feather	0.995	0.498	0.832	1.000
leaf	0.752	0.302	0.668	0.813
physical_damage	0.552	0.251	0.543	0.565
snow	0.467	0.202	0.567	0.494
dust_partical	0.408	0.160	0.590	0.373
bird_drop	0.100	0.030	0.214	0.246

All versions with model cards: huggingface.co/4keles/solar-panel-od

---

Developer Guide

Active Learning (human-in-the-loop annotation)

# Start Label Studio in a separate terminal
label-studio start

# Run the auto-annotation pipeline on raw drone footage
uv run python scripts/active_learning_pipeline.py \
  --image-dir data/raw_data/unlabeled \
  --model models/v1.2.1/best.pt

Training

uv run python scripts/train.py --config scripts/schemas/train_config.yaml

Data augmentation

uv run python scripts/augment.py \
  --source data/processed_data/rgb_master/train \
  --target-count 5000

Run tests

pytest tests/

---

Project Structure

solar_panel_od/
├── configs/              # YAML configs for training and UI
├── data/                 # Datasets and labels (gitignored — ~25 GB)
├── docs/                 # Active learning guide and research notes
├── models/               # Model weights — downloaded via download_model.py
│   └── v1.2.1/           # best.pt (PyTorch) + best.onnx (37.9 MB)
├── reports/              # Per-version val_summary.json
│   └── v1.2.1/
├── scripts/              # Training, augmentation, export, download utilities
│   ├── download_model.py # Pull weights from HuggingFace Hub
│   ├── train.py
│   ├── augment.py
│   ├── export_engine.py  # ONNX → TensorRT
│   └── active_learning_pipeline.py
├── streaming/            # PyQt6 live dashboard source
├── tests/                # Pytest test suite
├── tools/                # Utility scripts
├── main.py               # Dashboard entry point
├── pyproject.toml        # Dependencies (uv)
└── .env.example          # Env variable template

---

License

MIT — see LICENSE.

---

Citation

@software{solar_panel_od_2026,
  author    = {4keles},
  title     = {Solar Panel Defect Detection — YOLO11 Multi-Modal},
  year      = {2026},
  url       = {https://github.com/4keles/Solar-Panel-AI-Analysis},
  note      = {Model weights: https://huggingface.co/4keles/solar-panel-od}
}