Content-Based Image Retrieval of Butterflies

• Main Authors: Chen, Bee-Chung, 陳必衷
• Other Authors: Hsiang, Jieh
• Format: Others
• Language: zh-TW
• Published: 2000
• Online Access: http://ndltd.ncl.edu.tw/handle/01246083829116682000

碩士 === 國立臺灣大學 === 資訊工程學研究所 === 88 === In this thesis, we present a butterfly query system, which is based on the appearances of butterflies and is extremely user-friendly. We propose a mechanism for the users to describe, graphically, the features of a butterfly. This feature query returns a menu of butterflies, from which the user can further query according to similarity or related features. We call this process “QBF/QBE Query Model”, where QBF stands for Query By Features and QBE for Query By Example. The biggest challenge is that different people have different perception. While one may describe a butterfly as a white one with black stripes, another may describe the same butterfly as a black one with white spots. The query system needs to be robust enough to eventually lead to the same butterfly. During 1999, we built a “Taiwanese Butterfly Query System” based on QBF/QBE Query Model. This system successfully overcomes the “perception problem” and provides an intuitive way for querying about butterflies. However it requires plenty of human efforts (to annotate features and to sort butterflies according to similarity) and is not easy to be scaled up. To overcome this problem, we develop a semi-automatic data construction procedure, including “Butterfly Segmentation”, “Feature Extraction”, and “Perceptual Level Annotation”. Besides, we also propose a set of “Similarity Measures” and “Degree-of-Match Measures” for butterfly images. Through experiments, it shows that the measures we propose excel other candidates. This thesis consists of three major parts: (1) QBF/QBE Query Model: A mechanism, which provides an intuitive way to query images and overcomes the “perception problem”. This model is not only suitable for butterfly image retrieval, but also for any “data-constrained” image retrieval. (For example, we constrain our data to be butterfly specimen images.) (2) Taiwanese Butterfly Query System: A small (850 butterfly images) but successful image query system, which is a prototype system to test whether QBF/QBE Query Model works well, and to be the control of experiments on semi-automatic data construction procedure, and on similarity and degree-of-match measures. (3) Improvements: Including semi-automatic data construction procedure, similarity measures and degree-of-match measures, which give our system the ability to be scaled up. Not only for butterfly image retrieval, they are also useful for retrieval of other images.