HiD-CLIP integrates hierarchical semantic scoring, dual-stream localization and training-time pixel refinement for cross-domain industrial anomaly detection

Abstract Industrial inspection models may degrade when they are applied to product categories, materials, or imaging conditions that differ from the training data. We introduce HiD-CLIP, a CLIP-based anomaly detection framework that keeps the pretrained backbone frozen and adds adapters for product-level anomaly scoring and pixel-level defect localization. The method combines information from multiple visual depths, fuses an adapted similarity map with an attention-based map, and applies a training-only auxiliary pixel-refinement loss for multi-scale localization. In source-supervised experiments, no target labels or masks are used for adaptation, and HiD-CLIP is evaluated on seven public target domains. Using the public or corresponding training procedure for each method, HiD-CLIP achieves Image AUROC/AP of 89.76/89.24 and Pixel AUROC/PRO of 95.89/85.47. It gives the highest average Image AUROC and Pixel AUROC in this comparison, while the Image AP and Pixel PRO results indicate remaining room for better score calibration and region overlap. The reported accuracy is accompanied by higher inference cost and peak GPU memory than several lighter CLIP-based alternatives.