| 摘要: | 近年來基於影像之地點辨識(Visual place categorization)是研究人員重視的研究議題。然而，一影像辨識系統常受到物件大小、光線條件、物件遮蔽、視角變化影響。此外，地點辨識系統必須具備有效率的訓練及測試過程。相關的地點辨識研究中，其在分類影像內特徵時，大多採用非線性支撐向量機(support vector machine, SVM)，因其相較線性 SVM普遍能獲得較佳的辨識率，但其缺點是訓練階段之時間複雜度為O(N2)?O(N3)，N為訓練樣本數量，如此無法有效率地處理大規模(large-scale)樣本，因此本論文提出基於局部特性之linear SVM之群體分類器(ensemble of local linear SVMs, ELL-SVMs)，可有效降低訓練階段之時間複雜度至O(N1.5)。 本系統分為訓練階段及測試階段。於訓練階段，我們提出一創造群體(ensemble creation)的方法，其可在原特徵空間中找尋近乎線性可分割(linearly-separable)之子集，再利用一linear SVM分割每個子集內的樣本。在測試階段，對於影像中各個所偵測的特徵，我們利用貝式分類器及最近鄰居法則(nearest neighbor rule)個別選擇對每個特徵最佳的分類器。接著，本論文提出基於信心度之加權式一對多方案(confidence-based weighted one-against-all, CW-OAA)以結合影像中各個特徵之分類結果，以決定影像所屬地點。實驗結果顯示，本論文所提出的ELL-SVMs在訓練速度上，相似於FaLK-SVM，且優於BVM，在測試速度上，同時優於FaLK-SVM、BVM以及標準SVM。此外，本論文所提出之整合式地點影像的辨識率，皆優於FaLK-SVM、BVM以及標準SVM搭配OAA之方案。 Recently visual place categorization is an important research topic due to its numerous potential applications. However, such visual categorization system is easily affected by object scale, illumination conditions, object occlusion and viewpoints. In addition, categorization system should be efficiently trained and tested with huge amount of visual cues extracted in a very short period. Relevant researches on visual place categorization rely on non-linear SVM to categorize those visual cues within each image, since non-linear SVM has always shown promising categorization results. However, its major defect is that it suffers from O(N2) to O(N3) in training complexity and O(D?S) in test complexity, where N is the size of training data, D is the dimension of data vector, and S is the amount of support vector. Efficient training and test pro-cesses are demanding for tackling large-scale categorization problems. Therefore, this thesis proposes ensemble of local linear SVMs (ELL-SVMs), lowering training com-plexity to O(N1.5). Our proposed scheme has training and test phases. In training phase, we propose a scheme for generating ensemble of local linear SVM (L-SVM). This idea is derived from discovering the linearly-separable partitions among training data, while such partitions are found, those partitions could be classified by linear SVM instead of non-linear one. In test phase, we impose nearest neighbour rule into Bayes decision rule to assist in identifying the best trained local L-SVM for test sample. Afterwards, we further propose confidence-based weighted one-against-all (CW-OAA) approach to fuse the categorization results of visual cues within an image, and thus to categorize the image. Empirically, the training speed of proposed ELL-SVMs is similar to that of FaLK-SVM and much lower than those of BVM and standard SVM. The test speed of ELL-SVMs is lower than FaLK-SVM, BVM and standard SVM. Moreover, the categorization ability of ELL-SVMs with CW-OAA outperforms three SVM’s variants with OAA. |
