本研究針對工業管線鏽蝕檢測中的多樣化挑戰，提出了一套創新的解決方案，結合資料擴增技術與雙向強化學習機制，旨在提升檢測模型的精度、穩定性和泛化能力，以應對工業環境中複雜多變的鏽蝕特徵。 在資料擴增方面，本研究基於擴散模型（Diffusion Model）與其延伸技術（ControlNet）技術，生成高解析度且涵蓋多種形態的鏽蝕影像，並結合自建篩選模組，根據指定的鏽蝕等級進行特徵篩選與優化。此方法顯著擴充了訓練數據集的多樣性，並有效解決了原始數據集中樣本不足的問題。在多樣化場景下，生成的影像與實地拍攝影像相結合，使模型在小型、不規則及背景複雜的鏽蝕區域中表現出色。實驗結果顯示，結合此資料擴增技術後，YOLOv8 和 YOLOv11 模型的檢測準確度顯著提升，漏檢測率大幅降低，特別是在形狀不規則與背景干擾嚴重的鏽蝕區域中，檢測性能實現突破性改善。 此外，本研究引入了雙向強化學習機制，透過獎懲回饋機制實現模型策略的動態優化。在訓練初期，人為干預對模型檢測結果進行校正，幫助模型快速聚焦於鏽蝕區域。隨著訓練的進行，雙向強化學習逐漸取代人為干預，通過自適應調整檢測策略，增強了模型在多樣化工業場景中的適應能力，同時顯著提升了檢測結果的穩定性與準確性。綜合而言，本研究成功結合資料擴增技術與雙向強化學習機制，顯著提升了工業鏽蝕檢測模型的整體性能，並展現了兩者相結合的創新潛力與應用價值。 ;This study addresses the challenges in industrial pipeline rust detection by proposing an innovative solution that combines data augmentation techniques with a bidirectional reinforcement learning mechanism. The goal is to enhance the precision, stability, and generalization capability of detection models to handle diverse and complex rust charac-teristics in industrial environments. For data augmentation, this study utilizes Diffusion Models and ControlNet tech-nology to generate high-resolution rust images. A customized filtering module is integrat-ed to select and optimize features based on specified rust levels. This approach signifi-cantly expands the diversity of the training dataset and effectively resolves the issue of insufficient samples in the original dataset. By combining generated images with re-al-world captured images, the model demonstrates exceptional performance in detecting small, irregular, and complex rusted regions. Experimental results show that with this data augmentation technique, the detection accuracy of YOLOv8 and YOLOv11 models im-proved significantly, with a notable reduction in missed detections, particularly in irregular shapes and highly noisy backgrounds. In addition, this study introduces a bidirectional reinforcement learning mechanism that dynamically optimizes model strategies through reward and penalty feedback. During the initial training phase, human intervention plays a critical role in correcting model de-tection results, helping the model focus on rust areas. As training progresses, the bidirec-tional reinforcement learning system gradually replaces human intervention by adaptively adjusting detection strategies. This enhancement strengthens the model’s adaptability to diverse industrial scenarios while significantly improving detection stability and accuracy. In summary, this study successfully integrates data augmentation techniques with a bidi-rectional reinforcement learning mechanism, substantially enhancing the overall perfor-mance of industrial rust detection models. It demonstrates the innovative potential and practical value of this combined approach in addressing complex rust detection tasks.
