近年來，流形學習(manifold learning)的技術被應用到許多圖形識別的領域中，流形學習的方法就是基於所要分析的物件資料在高維度空間中有平滑流形分佈的假設，再利用轉置（re-embed）的方法將物件資料投影到較低維度的歐氏空間，並區域性（locally）保持其原有在流形上的分佈。由於多媒體應用的蓬勃發展，常需要分析如視訊、聲音、高解析度影像等資料，而這些多媒體資料其特徵向量往往都是採用高維度的方式來描述，因此需要一適當的降維方法使多媒體資料在降維後仍然可以保持其在原始高維特徵空間中的結構關係，而流形學習正是符合這樣概念的方法。在本研究中，我們提出了一個新的最近特徵線子空間學習方法，並且將其應用於車輛顏色辨識及相關性回饋影像檢索系統。透過這樣的子空間學習方法可以保留樣本區域結構拓樸及達到最大邊際投影的效果。在實驗結果中，我們的演算法與其他著名的演算法在比較後都有較佳的辨識結果。;In recent years, manifold learning has attracted a lot of researchers. Manifold learning assumes data set in the high-dimensional feature space are with specific manifold distribution, which could be re-embed into the low-dimensional Euclidean space. The rapid generation on digital information has made the development of efficient multimedia applications become urgently. Abundant multimedia data (video, audio and high-resolution image) need intelligent analysis which can be become useful information for multimedia application, such as video surveillance, audio classification and image retrieval. In addition, these multimedia data are often characterized by integrating high-dimensional features. Therefore, an approach which can effectively reduce high-dimensional features and keep its structural relationship in the low-dimensional feature space is required. As mentioned above, the manifold learning method is consistent with this concept. In this study, we proposed a novel nearest feature-line subspace learning method, and this method is applied to vehicle color classification and relevance feedback image retrieval system. According to our proposed subspace learning method which can be effectively preserved structural locality of samples and maximum margin projection in the new feature space. Experimental results have shown that our proposed method outperformed several state-of-the-art algorithms.
