Shearographic anomaly detection dataset (SADD)

Shearography is an emerging optical method for non-destructive testing and is gaining increasing attention in industrial inspection scenarios. However, the development and systematic comparison of machine learning methods to allow higher degrees of shearographic inspection are limited by the lack of publicly available, well-characterized datasets. This paper introduces a curated shearography dataset designed specifically for machine vision research. The dataset comprises systematically designed defect geometries with controlled sizes and orientations, complemented by defect-free samples and annotated measurement artefacts that frequently occur in practical measurements. All annotations were created by a domain expert and cross-checked by a second researcher. They are supported by a detailed description of the underlying deformation physics, which explains characteristic shearographic signatures and class ambiguities. This physical context motivates the experimental design and supports informed interpretation of learning-based results beyond purely statistical correlations. The dataset enables learning-based methods across unsupervised, supervised, and zero-shot paradigms, demonstrated through three representative use cases: defect detection, multiclass object detection, and text-based automated labeling of shearographic measurements. It provides a standardized and reproducible benchmark for systematic machine-vision research and for evaluating advanced methods under controlled shearographic inspection conditions. All data and code are publicly available.