- 學位論文
基於稀疏表示之電子鼻混合氣味辨識之研究
Sparse Representation Based Mixed Odor Recognition by an Electric Nose
摘要
稀疏表示法近年來在影像分析上有很好的辨識效果，且基於稀疏表示的語者識別方法相繼被提出，本論文藉由稀疏表示線性組合的特性進行氣味的分析，目標在於做單一氣味以及混合氣味的辨識，首先，選用生活中常見的20種氣味源(混合物)，建立稀疏表示演算法進行單一氣味的辨識，接著，將兩種單一氣味組合成18種混合氣味，訓練資料庫以單一氣味的數據資料作為基礎，使用稀疏表示法的特性進行混合氣味的辨識，期望可以將混合氣味判斷為單一氣味的線性組合。 在辨識氣味成分前，本論文提出以稀疏表示的特性判斷氣味源為單一氣味或是混合氣味，正確率可達到7成以上。接著，為了獲取有意義的多變量響應作為特徵進行氣味識別，我們分別在資料前處理、特徵選擇、分類器以及辨識方法上進行比較與分析，找到最佳的資料處理以及分類方法。資料前處理方法上，使用不同的壓縮方式以及正規化；特徵選擇包含特徵選取以及特徵萃取，前者採用循序向後的特徵選擇演算法，後者則使用了主成分分析方法以及線性識別分析；分類器選用最近鄰居分類器以及稀疏表示法；辨識方法上，在單一氣味的分析上使用多類別辨識方法，混合氣味則比較多類別辨識方法以及多標籤辨識方法的分析結果。本論文中的混合氣味資料共有18組，每一組是由兩種單一氣味所組成，因此共有36個目標氣味需要判別，目前最多可以判斷出14個目標氣味，其中包含正確地判斷兩組混合氣味。
並列摘要
Sparse Representation Classification (SRC) has performed well in the field of image analysis and speaker identification. In this thesis, we applied SRC in single and mixed odor recognition. First, we chose 20 kinds of odor sources and built an SRC-based algorithm to recognize them. Then, we produced 18 kinds of mixed odors by mixing two of the 20 kinds of odor sources. We attempted to recognize the mixed odors by learning from the training data set which consists of single odor data. A mixed odor could be recognized as a linear combination of the single odors by using SRC. Before performing odor recognition, we detected whether the sample was a single odor or a mixed odor with >70% accuracy using an SR-based method. Then, odors were analyzed by the following steps: data preprocessing, dimension reduction, classification and recognition, respectively. For data preprocessing, various methods were applied to compress and normalize the raw data. Afterwards, dimension reduction was achieved via feature selection and feature extraction. We used Sequential Backward Selection (SBS) for feature selection, and Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) for feature extraction. For data classification, we applied K-Nearest Neighbor Classification (KNNC) and SRC. For recognition, we used multiclass classification for single odors identification, and compared the results of mixed odor identification produced by a multiclass algorithm and a multi-label algorithm. The mixed-odor dataset consisted of 18 pairs of odors, so there were 36 targets to identify. Results show that, as many as 14 targets could be successfully identified.