Baseline Study on Damage Recognition for EPDM Diaphragms Using Four-Light Photometric Stereo

The objective of this study was to establish a fundamental framework for automated damage recognition in EPDM rubber diaphragms under industrial conditions. This was achieved using a four-light photometric stereo technique. In this study, we utilized a production dataset comprising 313 MD65 components with expert annotations to assess four binary tasks: kidney deformation, warp out-of-tolerance, wrinkle presence, and crack presence. Photometric-stereo normals are integrated into a height map, divergence-based features, and morphology-yield class decisions. The findings indicate an F1 score of 0.81 for both kidney and warp, with recall rates of 0.93 and 0.92, respectively, and precision rates of 0.72 for both. Crack detection is recall-oriented (recall 0.88) but precision-limited (precision 0.19) owing to systematic confusion with wrinkles. Wrinkle detection exhibits perfect recall but degrades specificity at the current operating point. This was addressed as a design choice for recall-first quality assurance and as a target for thresholding improvements. Our error analysis identified crack–wrinkle ambiguity as the predominant failure mode. We delineate a comprehensive methodology, comprising polarization imaging, regulated deformation during acquisition, and a lightweight CNN cascade, to enhance selectivity while maintaining industrial robustness. This study establishes a transparent, reproducible baseline and evaluation protocol for future iterations on elastomeric surfaces.