Data Augmentation Study Improves Laser Speckle Material Classification
A new study, shared on arXiv (2607.22725), explores how controlled data augmentation affects the classification of materials using laser speckle, specifically through the SensiCut dataset. The team applied a framework for parametric augmentation to train both ResNet18 and EfficientNet-B0. They used various techniques, including rotation, Gaussian blur, independent Gaussian noise, and others. To measure the effectiveness, they assessed the macro F1-score across three random seeds. They also created ordinary least squares models to link augmentation methods with performance for each network, finding that Gaussian blur was particularly influential. This research highlights that typical augmentation strategies fall short for coherent imaging, as laser speckle patterns result from coherent interference and contain unique structured statistics.
Key facts
- arXiv paper 2607.22725 studies data augmentation for laser speckle material classification
- Uses SensiCut dataset
- Trains ResNet18 and EfficientNet-B0
- Augmentation framework includes rotation, Gaussian blur, Gaussian noise, speckle-aware noise, intensity jitter, spatial masking
- Performance measured by macro F1-score averaged over three random seeds
- Ordinary least squares models link augmentation parameters to performance
- Gaussian blur is a key factor across both models
- Standard augmentation policies are poorly matched to coherent imaging
Entities
Institutions
- arXiv