在現代硬體技術快速發展下，許多人工智慧的研究也得到了突破性的進展，愈來愈多的領域紛紛朝向機器取代人類或輔助人類的方向發展與研究。但大多數的人工智慧或深度學習都需要大量的訓練資料並且只能應用於單一任務目標，然而在取得這些大量訓練資料都是相當困難的，例如醫學圖像。在圖像處理領域中，少樣本圖像切割便是其中的一個研究。近期的研究都透過深度學習與元學習的方法，使訓練的模型只需要少量的訓練資料就可以切割出圖像中的目標並使模型可以快速適應於新任務目標。 本論文提出的以元學習分類權重轉移網路生成遮罩於少樣本圖像分割架構做為元學習少樣本圖像切割訓練的網路架構，透過預訓練好的分類權重轉移架構去生成良好的先驗遮罩，並利用預訓練好的特徵提取架構進行query image與support image的特徵提取，然後利用特徵增強模塊中自上而下的路徑自適應地將信息從更精細的特徵傳遞到粗糙的特徵來進行query image的特徵提取，最後再通過分類模塊去進行query image的切割預測。實驗結果表明，以聯合平均交集(mIOU)為評估機制與Baseline相比，在1-shot的實驗結果中準確度上升了1.7%，而在5-shot的實驗結果中準確度也上升了2.6%，所以與Baseline相比下證明了以元學習分類權重轉移網路生成遮罩於少樣本圖像分割架構有最佳的少樣本圖像切割表現。 關鍵字：元學習、少樣本圖像切割、語意分割、少樣本學習 ;With the fast development of the hardware technology in today′s world, many artificial intelligence researches have also made a breakthrough, and the more and more fields are developing and researching in the direction of replacing humans with machines or aiding humans. However, most artificial intelligence or deep learning requires large amounts of training data and can only be applied to a single task, but it is very difficult to obtain these large amounts of training data, such as medical images. In the image processing field, one of the studies is few-shot image segmentation. Recent studies have used deep learning and meta-learning approaches to enable trained model to segment targets in images with only a few training data and to adapt the model to new tasks quickly. In this thesis, we proposed the meta-learning classifier weight transformer network generation masks for few-shot image segmentation architecture. The top-down path in the feature enrichment module is used to transfer the information from finer features to coarser features for query image feature extraction, and finally the classification module is used for query image segmentation prediction. The experimental results show that the mean intersection over union(mIOU) ratio as the evaluation mechanism, proposed method increased the accuracy by 1.7% in the 1-shot experiment and 2.6% in the 5-shot experiment compared with Baseline, so it proves that the meta-learning classifier weight transfer network generation masks for few-shot image segmentation architecture has the best performance. Keywords – meta learning, few-shot image segmentation, semantic segmentation, few-shot learning.
