2026-04-02
【學術亮點-頂級期刊論文】基於弱標註的影像語意分割之少樣本學習
Font Size
Small
Middle
Large
AI core Technology: Governance Framework for Trustworthy Agricultural AI Agents
【Department of Management Information Systems / Josh Jia-Ching Ying / Associate Professor】
核心技術：可信任農業 AI 代理人治理框架【資訊管理學系英家慶副教授】
上架日期2026-03-16
【Department of Management Information Systems / Josh Jia-Ching Ying / Associate Professor】
核心技術：可信任農業 AI 代理人治理框架【資訊管理學系英家慶副教授】
| 論文篇名 | 英文：A few-shot learning for image semantic segmentation with weak annotations 中文：基於弱標註的影像語意分割之少樣本學習 |
| 期刊名稱 | Engineering Applications of Artificial Intelligence (指標清單期刊) |
| 發表年份, 卷數, 起迄頁數 | 2026,174, no.114525 |
| 作者 | Josh Jia-Ching Ying(英家慶)*, Jin-Qun Liao, Ji Zhang |
| DOI | 10.1016/j.engappai.2026.114525 |
| 中文摘要 | 儘管基於卷積神經網路的影像語意分割已取得顯著進展，高精度模型的訓練仍需大量像素層級標註，而此過程既耗時又費力。雖然少樣本學習可降低對大規模標註資料的依賴，但現有多數少樣本語意分割方法仍需強標註的支援集。為解決此限制，本文提出一種以邊界框標註作為弱監督的少樣本語意分割框架，此類標註方式在實務上更易取得且已被廣泛採用。為抑制邊界框區域中固有的非目標干擾，我們引入一個基於輪廓偵測的預遮罩（pre-mask）生成模組，以產生接近像素層級標註品質的支援遮罩。在 PASCAL VOC 2012 基準資料集上的實驗結果顯示，本方法可達到 55.44% 的 Mean-IoU，在經策展的 1-shot 評估設定下，相較於現有最先進的弱監督方法最高提升 17.27%；在未策展（支援影像隨機抽樣且未經品質篩選）的設定下，仍可提升 5.81%。這些結果證實，本方法的效能提升在不同評估條件下具有一致性，並不依賴於精心挑選的支援影像；同時，相較於像素層級標註，採用邊界框標註可大幅降低標註成本。 |
| 英文摘要 | Despite significant advances in convolutional neural network-based image semantic segmentation, training high-accuracy models continues to demand extensive pixel-level annotation, which is both time-consuming and labor-intensive. Although few-shot learning mitigates the reliance on large labeled datasets, most existing few-shot semantic segmentation methods still require strongly annotated support sets. To address this limitation, we propose a few-shot semantic segmentation framework that accepts bounding box annotations as weak supervision, a labeling modality that is more accessible and widely adopted in practice. To suppress non-target interference inherent in bounding box regions, we introduce a pre-mask generation module based on contour detection, which produces support masks of quality approaching that of pixel-level annotations. Experiments on the PASCAL VOC 2012 benchmark demonstrate that the proposed method achieves a Mean-IoU of 55.44%, outperforming state-of-the-art weakly supervised methods by up to 17.27% under a curated 1-shot evaluation protocol, and by 5.81% under an uncurated protocol where support images are randomly sampled without quality-based filtering. These results confirm that the performance gains are consistent across evaluation conditions and are not contingent on careful support image selection, while the use of bounding box annotations substantially reduces annotation overhead compared to pixel-level labeling. |
| 發表成果與AI計畫研究主題相關性 | AI Agent在on-line learning時，使用者很難提供強標註的資料作為訓練素材，並且往往能提供的資料量也十分有限，因此如何讓模型再少樣本且只有弱標註的情況下依然能進行on-line learning，這會是AI Agent能否快速適應的重要議題，本研究成果將可促成本研究計畫所建構的AI Agent在影像語意分割任務上能有較好的適應性。 |
