臺灣博碩士論文加值系統

尺(metric)可以用以量度距離；若吾人將此觀念應用到影像處理上，我們可以得到〝影
像尺〞(image metric)。影像尺是一種抽象的觀念，它可用來計算兩張影像間的不相似程
度(dissimilarity)，也因此量度了這兩張影像間的距離(distance, or norm)。在本篇論
文中，我們發明了一種嶄新的影像尺；這支以多重解析度訊號處理(multiresolution di-
gital signal processing)為基礎的影像尺讓全影像檢索及查詢(content-based image
indexing and querying)的程序變得準確(accurate)、快速(fast)，且更為好用(user-
friendly)。
在我們的影像尺模型(model)中，我們的影像母尺事實是由兩支子尺(sub-metric)所組
成的；這兩支子尺隨時可以以不同的百分比例混合，來組成我們的母尺。兩支子尺其中一
支是以影像中的顏色分佈為準則(color distribution)來度量兩張影像間的距離，而另一
支則是以影像內容的形狀資訊(content shape information)來量度。因此，我們的影像
尺可以用來處理不同繪圖習慣的使用者所畫出來各式各樣的查詢影像(query image)。以
這支影像尺所實作的系統不但查詢命中率奇高，並且有著驚人的處理速度。在我們的實作
系統上，從查詢影像的解壓縮(decompression)開始，經過前置處理(preprocessing)、小
波轉換(wavelet transformation)、特徵粹取(feature extraction)、資料庫存取(data-
base access)，以至於最後用影像尺作特徵比對的距離計算(distance measurement)，到
結果排序(ranking)，總共只需不到四秒的時間。
除此之外，我們還進一步把這支影像尺運用到視訊處理的場景變換偵測(video scene
change detection)上，而提出了有別於前人所提六大類演算法的第七種場景變換偵測演
算法。這個以多重解析度為基礎方法(multiresolution-based approach)，配合了我們所
獨創的切贅影像轉換演算法(border-cropping conversion)，更是如虎添翼，既快又準。
為了上述的研究，我們在Microsoft Windows 95作業系統上研發出了一套系統，這系統
我們稱它為 QueryStore II Plus；它含有兩個子系統，第一個叫 PowerIQ (for Power
Image Querying)，它是一個全影像資料庫系統，不但具有基本資料庫管理系統的功能，
例如新增(augmentation)、刪除(deletion)、修改(modification)、瀏覽(browsing)、表
列(listing)等，並且提供了以微小波轉換為基礎的影像特徵粹取(wavelet-based image
feature extraction)、影像註冊(registration)、批次註冊(batch registration)、全
影像特徵檢索(content-based image querying)的高階功能；另一個子系統叫 SmartSCD
(for Smart Scene Change Detection)，它是一個使用上非常有彈性的系統，可對一般的
視訊作相當準確的片斷分割及場景變換偵測。

This thesis presents a new image metric that makes image querying more
friendly and accurately. In fact, in our system, there are two distinct
metrics, each of which is to take considerations on multiresolution transform
domain coefficients rather than to analyse the source image pixels as most
algorithms do. Ideally, one can mix two distances measured by two different
metrics to proceed in a query. Such a mix can be applied to painted querying
in image database systems, in which some user may behave as an impressionist
and emphasize on the color distribution in his query while another one may
like to give a sketch and draw the shape of a figure in monochrome. The result
metric can be used to deal with various kinds of query images, takes little
time in querying, and has a reasonable hit ratio. The experiment shows that it
only takes under four seconds totally to take actions on necessary preceding
operations for a 256x256x24 query, to transform it by wavelet, to operate on
the database, and to search among 60 database images using our metric.
Another contribution of our thesis is to apply our image metric to video
scene change detection, so as to propose a brand-new shot boundary detection
algorithm rather than the other six methods that were proposed previously.
This approach, which is multiresolution-based in nature, has been experiment-
ally validated to be a better one with a higher detection hit rate.
For our research, we have built a system that not only performs traditional
image database manipulation but is capable of content-based image querying.
Basic operations of an ordinary image database such as augmentation, deletion,
browsing, listing and modification are implemented. Enhanced functions like
wavelet-based feature extraction, image registration, multiple registration,
multiresolution image querying and video shot boundary detection are also
presented.