Geometry transfer of deep neural networks for heliostat detection

Reliable object and keypoint detection on heliostats is essential for advancing automation and practical deployment of airborne condition-monitoring methods in concentrated solar thermal (CST) tower plants. Yet the wide variety of heliostat collector geometries in real systems limits the applicability of deep-learning approaches, and their transferability between plants remains unexplored. To address this gap, we compiled a database of representative real-world heliostat geometries and generated a large synthetic dataset using our previously published rendering framework. With this data, we evaluated three training strategies: geometry-specific baseline models trained from scratch, a universal model intended to generalize across all geometries, and fine-tuning approaches initialized from either the baseline or the universal model. Baseline models perform well on their respective geometries, while the universal model shows inconsistent performance and may not be sufficient as a stand-alone solution. Fine-tuning, however, consistently adapts the model to new geometries and achieves performance comparable to geometry-specific baselines, as shown by suitable metrics on real-world test datasets of three distinct collector types. In practice, an effective model for a new geometry can be obtained by rendering as few as 100 synthetic images and fine-tuning an available baseline or universal model for 2000 iterations. This procedure requires 20 h on a single GPU and scales efficiently with additional computational resources. Overall, the results demonstrate that fine-tuning an existing heliostat detection model to the target domain provides a practical and reliable strategy for deploying deep models across the diverse heliostat collector geometries present in CST plants.