- 學位論文
區域性紋理影像的最佳化特徵之影像查詢系統
Image Retrieval System Using Optimum Adaptive Features for Local - Region Texture of Image
摘要
本研究主要以區域性紋理影像為目標，發展出一套最佳化特徵之影像查詢系統。為了達到高精確度的影像查詢，並避免因影像背景而影響影像查詢的結果，需先將影像進行物件切割處理，接著以物件區域為主萃取其特徵，最後以該物件之特徵進行影像查詢。 本研究於區域性紋理影像切割中，首先將萃取影像之粗糙度(coarseness)、對比度(contrast)與方向度(directionality)等三個紋理特徵，接著再針對此三個紋理特徵以動態區塊(dynamic block)的方式進行自適性平均濾波器處理(adaptive mean filter)，之後運用自適性動態K-mean分群方法，將相同的紋理特徵區域群聚在一起。最後，再將群聚區域做物件合併處理，經由邊界抽取獲得物件之輪廓，以達到物件切割的目的。 本文所進行的區域性紋理影像查詢，主要是針對切割後之物件進行特徵萃取，以萃取出影像之TFVP、CHKM與SFPMI等特徵。其中，TFVP為影像中物件區域紋理之紋理特徵；而CHKM為描述影像中物件區域的相似顏色像素值之顏色特徵，其能克服影像物件的位移與旋轉，以及影像物件雜訊的問題；SFPMI為描述影像物件區域的形狀特徵，以記錄該物件的輪廓及區域資訊，其能克服拍攝角度不同而造成扭曲等情況。此三項特徵可以表示區域性紋理影像之不同特性，並能查詢出與欲查詢影像之紋理、顏色與形狀較相似之影像。 為了能更精確表示本文方法的準確性，本研究分別與其他影像切割與影像查詢方法進行比較，以證明本研究之物件切割方法能準確的切割出該影像物件，且能獲得較高精確的影像查詢結果。若能將原本的關鍵字查詢加上本文提出的影像查詢方法，於未來期望能在龐大的數位影像下的管理及使用者查詢，能達到更多元、更完整且更方便之數位影像的查詢與管理。
並列摘要
This study took local-region texture of image as the targets to develop an image retrieval system with optimum features. In order to retrieve images with high precision and avoid impacts of background image on retrieval results, segmentation of image objects needed to be done first. Then the features were extracted mainly according to the regional part of objects; and finally image retrieval by features of objects could be made. For local-region texture of image segmentation in this study, we extracted three texture features: coarseness, contrast and directionality of images first. Then the three texture features were filtered by adaptive mean filter which adopted dynamic block method. After that, we used adaptive dynamic K-mean clustering method to cluster the same texture feature region together. Finally, the clustered regions were merged into an object, and the contour of object was obtained by contour extraction, in order to achieve the purpose of object segmentation. In this study, the conducted local-region texture of image retrieval mainly extracted object characteristics after segmentation to get TFVP, CHKM, SFPMI, etc features. Among them, TFVP is the texture characteristics of the object local-region texture of image; CHKM is used to describe the color features of the similar color pixel value in the image object region, and it can solve displacement and rotation problems in image objects as well as image object noise problems; SFPMI is used to describe shape features of image object region to record the object's contour and regional information, which can overcome the distortions or other problems caused by different camera angles. These three types of features can express different characteristics of local-region texture of image, and be used to retrieve the images which have similar texture, color and shape with the image intended to be retrieved. In order to more accurately represent the accuracy of this method, method of this study would be compared with other image-segmentation/image-retrieval methods to show the object segmentation in this study can accurately segment the image object, and obtain results with higher precision in image retrieval. If the proposed image retrieval method could be added on the existing keyword query, then more diverse, more complete and more convenient digital image retrieval and management can be expected in the future, which can be applied to management and user queries of large digital image database.