鏽蝕檢測是工業管線維護與安全評估的重要工作。實際檢測時，人工目視判讀容易受到檢測人員經驗、判斷標準及鏽蝕邊界認定差異影響；真實場域影像亦常受污漬、鏽水與背景紋理干擾，使鏽蝕範圍難以穩定判定。此外，像素級標註成本高，限制了分割模型的訓練資料規模，而模型推論結果仍可能存在局部誤判，難以直接滿足實際檢測需求。因此，本研究建立一套適用於真實管線影像的鏽蝕分析方法，以提升鏽蝕判讀的一致性，並提供後續面積比例計算與分布型態分析的基礎。本研究提出Appearance–Structure Collaborative Distillation with Interaction Network（ASCDI-Net），結合RGB影像與Pseudo Normal影像進行鏽蝕分割與互動式修正。RGB影像提供顏色、亮度與紋理資訊，Pseudo Normal影像補充局部結構與表面方向變化，使模型同時利用外觀與結構線索進行判斷。模型採用編碼器－解碼器架構，並加入Wavelet-based Shallow Structure Refinement（WSSR）模組強化淺層結構特徵。ASCDI-Net包含Establish、Refine、Extend與Interact四個階段。Establish 階段建立RGB與Pseudo Normal的基礎表徵；Refine階段利用人工標註資料訓練分割模型；Extend階段結合未標註資料進行半監督學習，以提升模型對真實場域影像的泛化能力；Interact階段則透過使用者框選感興趣區域（Region of Interest， ROI）及正負提示點修正局部錯誤，使模型輸出更符合實際判讀需求。 實驗結果顯示，ASCDI-Net在鏽蝕分割任務中達到IoU 0.64與Dice 0.78，較比較方法最佳結果分別提升0.21與0.16，顯示本研究能更穩定地分割真實場域中的鏽蝕區域。互動修正結果亦顯示，使用者僅需少量提示即可有效修正局部錯誤，提升最終分割品質。此外，本研究建立鏽蝕種類分類模型，將影像分為Clean、Spot、General與Pinpoint四類，並結合分割結果計算鏽蝕面積比例，以提供鏽蝕區域、面積比例及分布型態等完整分析資訊，作為工業管線鏽蝕檢測與後續維護決策之輔助依據。;Rust inspection is an essential task in industrial pipeline maintenance and safety assessment. In practical inspections， manual visual assessment is easily influenced by inspectors′ experience， evaluation criteria， and differences in rust boundary identification. In addition， real-world pipeline images are often affected by stains， rust runoff， and complex background textures， making rust regions difficult to identify consistently. Furthermore， the high cost of pixel-level annotation limits the availability of labeled training data， while segmentation models may still produce local prediction errors that fail to satisfy practical inspection requirements. Therefore， this study aims to develop a rust analysis framework for real-world pipeline images to improve the consistency of rust assessment and provide a reliable basis for subsequent rust area ratio estimation and distribution pattern analysis. This study proposes the Appearance–Structure Collaborative Distillation with Interaction Network (ASCDI-Net)， which integrates RGB images and Pseudo Normal images for rust segmentation and interactive refinement. RGB images provide appearance information， including color， intensity， and texture， whereas Pseudo Normal images provide complementary structural cues by describing local surface geometry and orientation variations. The proposed network adopts an encoder–decoder architecture and incorporates a Wavelet-based Shallow Structure Refinement (WSSR) module to enhance shallow structural features. ASCDI-Net consists of four stages: Establish， Refine， Extend， and Interact. The Establish stage learns the fundamental correspondence between RGB and Pseudo Normal representations. The Refine stage trains the segmentation model using manually annotated data. The Extend stage incorporates unlabeled data through semi-supervised learning to improve the model′s generalization to real-world scenarios. The Interact stage allows users to refine local segmentation errors by specifying a Region of Interest (ROI) together with positive and negative prompts， producing segmentation results that better satisfy practical inspection requirements. Experimental results demonstrate that ASCDI-Net achieves an Intersection over Union (IoU) of 0.64 and a Dice coefficient of 0.78 for rust segmentation， outperforming the best comparison method by 0.21 in IoU and 0.16 in Dice. These results indicate that the proposed method can segment rust regions in real-world pipeline images more accurately and consistently. Interactive refinement further shows that local segmentation errors can be effectively corrected using only a small number of user prompts， resulting in improved segmentation quality. In addition， a rust-type classification model is developed to categorize images into four classes: Clean， Spot， General， and Pinpoint. Combined with the segmentation results for rust area ratio estimation， the proposed framework provides comprehensive rust analysis information， including rust regions， area ratios， and distribution patterns， thereby serving as a useful tool for industrial pipeline rust inspection and subsequent maintenance decision-making.
