電腦視覺是一門研究如何令機器理解圖像的科學，隨著現在電腦速度的改進，處理圖像這種大規模資料的應用漸漸地扮演了舉足輕重的腳色，而物件辨識(Object Recognition)，就是其中一種相當熱門的技術，為了達到精準辨識的目的，從傳統的支持向量機(SVM)、稀疏表示分類器(SRC)以及DK-SVD(Discriminative K-SVD)，皆能獲得一定的效果。 基於DK-SVD，本論文提出一種結合局部保留投影(LPP)概念的LP-KSVD，在訓練的目標函示中加入局部保留的限制，使得訓練字典的同時能保留資料間局部的特性，並同時訓練出可用於物件辨識的線性分類器。除此之外，我們也提出將局部保留的目標式做為一種新的特徵，在一筆新的輸入訊號進來時，也能夠直接獲得額外的局部保留資訊；我們也使用核化方法將訓練資料投射至高維特徵空間以增強字典的辨識及重建能力並且提升系統彈性。 在實驗上使用Caltech101與15Scene資料庫，其結果顯示LP-KSVD在辨識率上有更好的表現。 ;Computer vision is a field that understanding images. Along with advances in computer performance, processing large-scale data becomes more common than before. Object Recognition is one of the popular computer vision technology. In order to enhance the recognition performance, we can use the traditional SVM, SRC and DK-SVD. In this paper, we present locality preserving K-SVD based on DK-KSVD. Adding the locality preserving term into the objective function to preserve the neighborhood structure of the data set. The locality preserving term can be used as a new feature. We can gain additional locality information for new input signals. We also apply kernel method on the dictionary learning which efficiently strengthen the reconstructive and discriminative ability. Experiments on Caltech101 and 15Scene databases indicate that the proposed method achieve good recognition performance.
