近年來，深度學習在電腦視覺領域取得重大進展，但其優異表現往往高度依賴大規模且標註精確的訓練資料。在有限資料 (Limited Data) 的情境下，模型容易產生過度擬合 (Overfitting) 並缺乏泛化能力。 為解決此問題，本研究提出了一種結合生成對抗網路 (Generative Adversarial Networks) 與線性樣本混合的兩階段資料增強 (Two-Stage Data Augmentation) 策略。首先，在第一階段引入專為少量資料設計的 StyleGAN2-ADA 作為基底架構，並將其修改為支援連續軟標籤的條件生成模型，藉此合成具備跨類別特徵的虛擬樣本。接著，為了避免生成樣本與標籤產生語意偏移，本研究設計了語意一致性過濾機制 (Semantic Consistency Filtering)，利用預訓練分類器與簡森-夏農散度 (Jensen-Shannon Divergence) 客觀地篩選出高品質的生成影像。最後，在第二階段中，將過濾後的生成資料與原始真實資料結合，再次應用標準線性樣本混合來訓練下游分類器，使非線性的結構多樣性與線性的正則化效益得以互補。 實驗結果顯示，本研究提出的兩階段增強策略在 CIFAR-10、CIFAR-100 與 Mini- ImageNet 等標準資料集的有限資料設定下，均提升了影像分類的準確度，顯示出結合生成對抗網路與樣本混合能夠提升少資料情況下模型的泛化能力。;In recent years， deep learning has made significant breakthroughs in computer vision， but its outstanding performance heavily relies on large-scale， accurately labeled training data. In limited-data scenarios， models are prone to overfitting and lack generalization ability. While traditional Linear Mixup smooths decision boundaries， its pixel-level interpolation often generates artifacts lacking real semantic meaning. Conversely， existing generative augmentation methods rarely explicitly model cross-class manifolds. To address these challenges， this study proposes a novel two-stage data augmentation strategy that integrates Generative Adversarial Networks (GANs) and Linear Mixup. In the first stage， a conditional StyleGAN2-ADA architecture， modified to support continuous soft-labels， is utilized to synthesize virtual samples with cross-class semantic features. To mitigate the risk of semantic drift between the generated images and their assigned labels， a Semantic Consistency Filtering mechanism is introduced. This mechanism employs a pretrained classifier and Jensen-Shannon Divergence (JSD) to objectively filter out inconsistent generated samples. Finally， in the second stage， the filtered GAN-generated dataset is combined with the original real dataset， and standard Linear Mixup is applied again during the training of the downstream classifier. This approach allows the non-linear structural diversity from the generative model to complement the linear regularization benefits. Experimental results demonstrate that the proposed two-stage augmentation strategy improves image classification accuracy under limited-data settings on benchmark datasets， including CIFAR-10， CIFAR-100， and Mini-ImageNet. The results suggest that combining GAN-based generation， semantic filtering， and second-stage Linear Mixup can improve generalization under limited-data conditions.
