UTIS-HeliostatBeamCharacterization

UNet-Based Target Image Segmentation for Camera Target Method in Solar Tower Plants

Overview

This project presents a robust implementation of the UNet3+ framework designed for background separation and flux determination in calibration target images used in solar tower plants. The model is trained using artificially generated target images, which are created by combining background images with simulated focal spots. By leveraging these synthetic images with known properties, the UNet model is trained to accurately separate the background from the focal spots, enabling precise flux prediction. For an in-depth explanation and methodology, please refer to this paper. The model used to achieve the results in this paper is placed in the train models folder.

Directory Structure

``` UTIS-HeliostatBeamCharacterization/

data/ # Directory containing image datasets used for training and inference emptyTargetImages/ # Images of empty targets used for artificial data generation fluxImages/ # Images representing flux used for artificial data generation realSamples/ # Real images used to test the model's inference capabilities

logs/ # Directory to store training logs trained_models/ # Directory to store trained model checkpoints tests/ # Directory for test scripts and inference predict.py # Script for running inference with a trained model results.png # Example output from the inference script

utis/ # Core package directory containing model and utility code init.py # Makes this directory a Python package dataset.py # Script for loading and processing image data architecture/ init.py UNet3Plus.py # UNet3+ model architecture layers.py # Additional layers used in the model unetplmodule.py # PyTorch Lightning module for UNet3+ train.py # Script for training the UNet3+ model

setup.py # Script for installing the package README.md # Project description LICENSE # License for your package ```

Usage

Training

Run the training script:

sh python -m utis.train

Inference

Run the inference script:

sh python -m tests.predict

Results

The predicted images and a sample result grid are saved in the specified output directory as results.png:

The result grid is organized as follows: - First row: Input image - Second row: Predicted focal spot - Third row: Predicted background

Acknowledgments

This work is supported by the Helmholtz AI platform grant.