Generative Inpainting for Privacy-Oriented Industrial CCTV Dataset Preparation: Cross-Domain Hard-Hat Detection with Temporal Separation

Hard-hat compliance detection is central to vision-based industrial safety monitoring. Public hard-hat datasets support model development but rarely reflect industrial closed-circuit television (CCTV) conditions, such as fixed elevated viewpoints, small objects, temporal correlation, occlusion, and privacy-sensitive content. We investigate (i) whether LaMa inpainting, alongside face blurring, affects detection, (ii) whether models trained on public data generalize to industrial CCTV, and (iii) why temporal separation matters for evaluation. We introduce the Industrial CCTV-HH Dataset, an annotated resource prepared from industrial surveillance recordings using privacy-oriented preprocessing, containing 2413 images and 9608 originally annotated head-region instances in two classes (with and without a helmet). A day-level temporal split reduces the risk of near-duplicate leakage and yields deployment-oriented evaluation. A preprocessing pipeline removes on-screen overlays and textual identifiers with generative inpainting (LaMa) and adds localized face blurring designed to preserve head structure. Cross-domain YOLO11n experiments reveal a large domain gap: a public-only model reaches mAP@0.5 ≈ 0.97 on public images but only ≈0.11 on industrial CCTV, whereas target-domain samples and higher input resolution raise industrial-domain mAP@0.5 to ≈0.93 (target-only) and ≈0.90 (balanced training, which also retains ≈0.97 on public images). YOLO11n mAP varied modestly; OCR and face tests indicated reduced detectability, not anonymity.