Domain-randomised instance-segmentation benchmark for soot in PIV images

Abstract Access to high-level statistical information from scientific image-based diagnostics is facilitated by the capacity to segment quantity-of-interest (QoI) signal accurately from measurement noise and interferences. In particular under conditions with diverse nuisance signatures, deep learning (DL) pipelines can outperform algorithm-based computer vision (CV) strategies and simultaneously enhance domain-invariance. DL-pipelines aim to overcome the generalisation bottleneck through augmentations and learnable parameters, supplemented by huge amounts of data. Yet, manual pixel-accurate annotation for scientific tasks is prohibitively expensive and the robustness of the trained models is often impeded by the underlying narrow and domain-specific set of training data. In the current study, a previously developed domain-randomised pipeline for automatic annotation and synthetic training data generation is benchmarked using a physics-aware composite score. An ablation study for background variance, QoI placement strategy and QoI object source proportions is conducted. The benchmark, consisting of 72 synthetic training data generation recipes, highlights the optimal set of domain-randomisation parameters to strike the balance between domain-invariance and segmentation accuracy. Synthetic training datasets with a balanced mix of QoI objects, fused onto realistic background instances, are found to provide the most accurate models for segmentation. The best model is subsequently used for the inference on semantically challenging Mie scattering images containing particle-image velocimetry (PIV) tracer particles and soot filaments. The model’s ability to detect soot structures with good accuracy is demonstrated and high-level soot filament area and contour statistics (e.g. curvature and fractal dimension) are used to delineate the effects of turbulent flow on soot filament structures. The present study highlights key parameters to tune domain randomisation strategies for DL-training pipelines and the pipeline usability and transferability is proved for autonomous semi-supervised learning. This eases access to high-level statistics in scientific image-based diagnostics.