對於CBIR(CONTENT-BASED IMAGE RETRIEVAL)的系統來說，除了一般IR系統所需要的條件外，如何讓使用者簡單明瞭的下達查詢指令是Image Retrieval另一個重要的課題，不似一般文字查詢的方式，影像查詢系統必須要有適當的查詢機制，讓使用者能夠描述其想要查詢的影像特徵，查詢的機制對使用者來說必須要越簡單越明瞭，對系統來說越能夠準確的描述影像特徵越好。
“台灣蝴蝶面面觀”是以「QBF\QBE」架構的蝴蝶搜尋系統，在一個Specific Domain下建立一個對使用者來說自然且易於使用的影像查詢系統，不需要複雜的方式讓使用者描繪查詢影像，而是系統先對影像定義好特徵，讓使用者以選擇特徵的方式來描繪想要查詢的蝴蝶影像，對使用者來說十分的自然且簡單，對系統來說，所有影像特徵皆是對所有的影像所歸納出來的，以這些特徵來做查詢會使的查詢的效果會更好。
在蝴蝶系統中把影像特徵分為3個dimensions，分別為 (color shape pattern)三個dimensions，並建立影像和這三個dimensions上影像特徵之間的『符合度』，以影像在各個影像特徵上的符合度做為查詢時之『索引』(indexing)。
對QBF系統來說建立影像與影像特徵間的”符合度”是最關鍵的工作，在舊系統中以人工觀察的方式，衡量每張影像對每個影像特徵的符合度，以人工的方式可以保證所得的”符合度”符合人所見的感覺，使搜尋結果更準確，不過對於擁有大量影像的系統來說，考慮到效率及擴充性的問題，必須要有自動化的方式才行，如何計算出符合一般人看法的符合度便是自動化最主要的問題。
在蝴蝶系統中有分(color shape pattern)三個dimension上的features，本篇論文主要的目的便是實作蝴蝶影像對color features(features in color dimension)的符合度，再與人工產生之符合度做比較，看自動化的方法是否足以取代人工來計算影像與color features之間的符合度。

For CBIR(Content-Based Image Retrieval) systems, it is important to key the query mechanism easy to use. Instead of querying by key words, a CBIR system needs some appropriate method for users to describe features of the image. The method has to be easy to understand for users and to precisely describe the image for systems.
“Taiwanese butterfly query system” is a specific domain query system which contains QBF(query by feature) and QBE(query by example) query methods. It let users query adult Taiwanese butterfly in a natural way. Without complicated describing methods, users just simply select butterfly image features which are predefined by the system to describe the query image. It is quite natural and simple for users. Because the predefined image features are the analysis of all butterfly images in the system, it is more precise by using these features to query butterfly images.
There are three dimensions of butterfly image features in Taiwanese butterfly query system. They are color, shape, and pattern. The system evaluates the proximity value between each image in the system and each feature in the three dimensions. Then use the proximity values to all features as each image’s “indexing”.
For the QBF query method, it is the most important thing to evaluate proximity values. System got proximity values manually. Watch the images then give the proximity value to each feature in the opinion. In this way, we can guarantee the proximity values matching human perception and also make the query result more precise. But for systems with large number of images, in consideration of scalability and efficiency, we need some automation method to evaluate proximity values. The main problem on automation is how to make the proximity values match human perception.
This thesis describes the design and implementation of a method to automatically evaluate proximity value between each butterfly image and each feature in color dimension. Use some similarity functions in image processing to calculate proximity values then compare the performances. Select the appropriate similarity function as the proximity function. At last, compare the result to ground truth (proximity values evaluated manually) to see if the method is good enough to evaluate proximity values between images and features.

第一章 序 論 1
第二章 蝴蝶影像搜尋系統 4
2-1 系統架構 4
2-2 QBF查詢架構 7
2-3 QBE查詢架構 9
2-4 QBF/QBE查詢流程 11
第三章 低階顏色特徵之擷取 13
3-1 COLOR SPACE 13
3-2 QUANTIZATION 16
3-2-1 UNIFORM QUANTIZATION 17
3-2-2 MEDIAN-CUT 18
3-2-3 MEAN-SPLIT 19
3-2-4 K-MEANS 20
3-2-5 VECTOR QUANTIZATION 21
3-3 IMAGE REPRESENTATION 23
第四章 影像與顏色特徵之符合度的測量 25
4-1 LN-METRIC 26
4-2 QUADRATIC FORM DISTANCE 27
4-3 HISTOGRAM INTERSECTION 30
4-4 QUADRATIC FORM SIMILARITY 31
第五章 蝴蝶系統之符合度的測量 32
5-1 QUANTIZATION 32
5-1-1 UNIFORM QUANTIZATION 33
5-1-2 MEDIAN-CUT 34
5-1-3 K-MEANS 34
5-1-4 QUANTIZATION效能比較 34
5-2 相似度 36
5-2-1 單色特徵之符合度 37
5-2-1-1 Quadratic Form Distance 37
5-2-1-2 Color Moment 38
5-2-1-3 Quadratic Form Similarity 39
5-2-2 雙色特徵之符合度 41
5-2-2-1 Quadratic Form Distance 42
5-2-2-2 Quadratic Form Similarity 42
5-2-1 多色特徵 43
5-3 平均效能比較 44
第六章 結論及未來研究方向 46
參考文獻 49

[1] B.C. Chen, Content-Based Image Retrieval of Butterflies, Master thesis, Dept. CSIE, Univ. of NTU.
[2] Giordano Beretta, Understanding Color, Hewlett-Packard Company Palo Alto
[3] Anil K. Jain, Image Perception, chapter 3 of Fundamentals of Digital Image Processing, Englewood Cliffs, NJ 07632, Prentice-Hall, 1989.
[4] D. Androutsos, K. N. Plataniotis, and A. N. Venetsanopoulos, A Novel Vector-Based Approach to Color Image Retrieval Using a Vector Angular-Based Distance Measure, computer vision and Image Understanding Vol. 75, Nos. 1/2, July/August, pp. 46-58, 1999.
[5] M. J. Swain and D. H. Ballard, Color indexing, Int. J. Comput. Vis., vol. 7, no. 1, pp. 11--32, 1991.
[6] S. J. Wan , S. K. M. Wong , P. Prusinkiewicz, An algorithm for multidimensional data clustering, ACM Transactions on Mathematical Software (TOMS), v.14 n.2, p.153-162, June 1988.
[7] M. Orchard and C. Bouman, Color Quantization of Images, IEEE Transactions on Signal Processing, vol. 29, no. 12, 1991.
[8] P. Heckbert, Color Image Quantization for Frame Buffer Display, Proceedings of SIGGRAPH ''82, July, 1982, pp. 297-307.
[9] Bentley, J. L., Friedman, J. H. Data structures for range searching. Computing Survey s ii, 4 (Dec. 1979), 397-409.
[10] WV, X., AND WITTEN, I. H. A fast k-means type clustering algorithm. Dept. Computer Science, Univ. of Calgary, Canada, May 1985.