A Study on Seashell Retrieval Using Image Feature Selection and Classification Methods

[[abstract]]近年來，隨著資訊科技的蓬勃發展及數位典藏計劃的實行，多數的研究以內容為基礎的影像檢索方式來搜尋資料，而這些研究多利用低階的影像特徵包含顏色、紋理和形狀等特徵做為影像分類的依據。本研究利用MPEG-7標準色彩結構描述子之HMMD顏色特徵、紋理特徵之模糊紋理頻譜以及形狀特徵之中心點到輪廓的距離等方法去摘錄貝殼影像的特徵值，然後使用統計量之方法，平均數、標準差以及熵，來取代低階影像以直方圖做為特徵比對的方式來對貝殼進行檢索。接著我們採用資料探勘的方法,K-Nearest-Neighbor、C4.5決策樹、倒傳遞類神經網路以及支撐向量機等用來進行貝殼影像的檢索。將上述的方法來進行貝殼的分類實驗並比較各檢索的方法，實驗結果顯示利用支撐向量機在訓練集為10%,測試集為90%以及訓練集與測試集各50%下,貝類圖片的檢索分別具有98%及99%較高正確率。本研究所提出的以統計量做為特徵值來取代傳統影像特徵的分類方式，可以節省資料儲存的空間成本和資料索引的時間且具有不錯的檢索結果[[abstract]]With the rapid development of information technology and the implement of the National Digital Archives Program recently, many researches adopted content-based image retrieval to search information. Most of these researches utilize low-level image features as the basis of image classification such as color, texture and shape. In the thesis, we try to use Hue-Min-Max-Difference (HMMD) of MPEG7 color structure descriptor as the color feature, the fuzzy texture spectrum as the texture feature and the centroid-contour distance as the shape feature to extract the features from seashell images. Then three statistical measures including Mean, Standard Deviation and Entropy are calculated from the computed image features on seashell retrieval. Finally, different classification techniques, K-Nearest-Neighbor smallest distance, C4.5 decision tree, back-propagation neural networks and support vector machine are applied to compare the classification performance.All of the above methods are carried on classification for seashell images and then compared with various methods. The experimental results show that Support Vector Machine under training set=10%, test set=90% and training set=50%, test set=50% obtains 98% and 99% precision rates for seashell image retrieval respectively. The proposed method which uses statistical measures to replace with the histograms of image feature can save stored cost of large amount data and indexing time. Experimental results present that our proposed method has good performance in image retrieval of seashell