SurfDet3D: Surface-Aware Geometric Representation and Quality-Aware 3D Detection of Small Defects on Ceramic Package Substrates

Accurate 3D detection of surface defects is essential for quality control in the manufacturing of ceramic package substrates. However, this task faces two major challenges: the reliable representation of subtle local surface geometric anomalies and the accurate localization and reliable candidate ranking of small defects with diverse morphologies. To address these challenges, we propose SurfDet3D, which preserves within-column relative geometry across multiple stages of a sparse backbone and introduces bi-scale surface statistics to enhance subtle geometric features. It further combines whole-box geometric constraints with localization quality estimation to improve bounding-box regression and candidate ranking. Experiments on the public CPS3D-Det benchmark show that SurfDet3D achieves AP3D,R11, APBEV,R11, AP3D,R40, and APBEV,R40 values of 48.13%, 46.45%, 46.38%, and 45.42%, respectively. These results demonstrate the effectiveness of SurfDet3D in representing subtle local surface geometric anomalies and improving the localization and candidate ranking of small defects with diverse morphologies. Ablation studies further verify the contribution of each proposed component.