This paper presents a large‑scale multi‑SKU benchmark, Bottle‑SKU, for shampoo‑bottle surface defect detection, together with a Prototype‑Guided Dynamic Domain Modulation (ProDMM) method. Bottle‑SKU comprises 13,119 images, 20,419 instance‑level bounding boxes, seven defect categories, and 48 bottle SKUs, providing both defect annotations and SKU identities to evaluate shared defect recognition and SKU‑specific appearance variation.
To tackle SKU‑specific appearance differences, ProDMM first builds a visual prototype for each SKU from the training set and predicts a soft SKU distribution for every input image. The prototype‑conditioned representation then dynamically generates low‑rank Query/Value modulation for the visual backbone and applies residual conditioning to the defect‑text‑to‑visual attention, while preserving the pretrained Grounding DINO detection pathway. The model is trained jointly with detection, SKU‑domain classification, and modulation‑regularization objectives.
Experiments on Bottle‑SKU and the public GC10‑DET‑SKU metallic surface defect dataset demonstrate that prototype‑guided domain conditioning substantially improves industrial surface defect detection performance.
Review