利用階層式比對之影像內容檢索

CONTENT-BASED IMAGE RETRIEVAL USING HIERARCHICAL MATCH

董志偉(Zhi-wei Dong)

指導教授 ： 楊燕珠

大同大學/經營設計學院/資訊經營學系所/碩士(2005年)

摘要

數位影像內容檢索（Content-Based Image Retrieval，CBIR）在近幾年來不斷的發展，也遇到一些瓶頸，本研究就以此為出發點，希望能設計出一套更適合人類視覺感知的影像內容檢索，並同時提高檢索的正確性。在本研究所設計的影像內容檢索系統中，採用了色彩及形狀影像特徵，並利用色彩特徵去做語意分類、色彩區塊及最後比對所需條件的其中之一，其中語意分析把影像分成紋理（Textured）及非紋理（Non-textured）兩類影像；而形狀特徵則是最後比對所需的另一個條件。根據這些影像特徵作階層式的影像過濾及比對，包括粗略過濾（Coarse Filter）、中等過濾（Medium Filter）和精細比對（Fine Match），在粗略過濾過程中，把與查詢影像在色彩種類和χ2值差異性太大的影像種類先移除，接下來進入中等過濾的處理步驟。進入中等過濾之後，透過色彩區塊及一般區塊的特徵，把與查詢影像差異性太大的影像移除。經過前面兩個階段過濾所剩下的影像則進入最後的精細比對，在進入精細比對之前，檢索影像必須先經過對比度(Contrast Rate)計算，若查詢影像的對比度高於門檻值，則採用以色彩為主的比對模式去檢索出相似影像。反之則採用以形狀為主的比對模式去檢索出相似影像。由以上可知本研究所提出的架構在檢索過程中，並非在影像資料庫中的每張影像都要經過所有檢索的比對流程，因此可提升檢索的效率且接近人的感知，達到提升檢索的正確率。根據實驗結果，本系統的確具有更好的正確率，同時亦較能符合人的感知。

關鍵字

影像內容檢索 ； 階層式比對 ； 語意分類

並列摘要

For almost a decade, Content-Based Image Retrieval (CBIR) has been an active research area. However, some fundamental problem remains largely unsolved until now. For these reasons, we want to design a better CBIR system that is more close to human perceptual and more efficiency. We call this “Hierarchical Match”. The system contains three phases framework (Coarse Filter, Medium Filter and Fine Match) and two kinds of image features (color and shape). We use color feature to semantic classification, color blocks, general blocks and final match. At the same time, we design an algorithm to classify images into the semantic classes, textured and nontextured. The shape feature also used to final match. At Coarse Filter phase, we use “χ2 Statistics” to classify textured and nontextured images. The “Number of Colors” is used to assist “χ2 Statistics” in classifying image more correctly. At Medium Filter phase, we filter out some images that are obvious different with query image. The system will delete about some images that are obvious different with query image at Color Blocks and General Blocks. Before go to Fine Match phase, we need to compute the Contrast Rate (CR) of image in advance. If CR great than threshold (T), we retrieve similar images based on shape feature primarily is better. Otherwise, we retrieve similar images based on color feature primarily is better. According to mention above, we know that not every image in the database needed to do all retrieved processes. Based on the experimental results obtained in this study, the performance of the proposed method and retrieval system is better than one existing system for comparison and more similar to human perceptual.

並列關鍵字

Content-Based Image Retrieval ； Hierarchical Match ； Semantic Classification

參考文獻

1. Gudivada, V. N., and Raghavan, V. V., “Content-Based Image Retrieval Systems,” IEEE Computer, Vol. 28, No. 9, 1995, pp. 18-22.

2. Mehtre, B. M., Kankanhalli, M. S., and Lee, W. F., “Content-based image retrieval using composite color shape approach,” Information Processing and Management, Vol. 34, No. 1, 1998, pp. 109-120.

3. Do, M. N., and MVetterli, A., “Wavelet-based texture retrieval using generalized Gaussian density and Kullback-Leibler distance,” IEEE Transactions on Image Processing, Vol. 11, No. 2, 2002, pp. 146-158.

4. Mehtre, B. M., Kankanhalli, M. S., and Lee, W. F., “Shape measures for content-based image retrieval - A comparison,” Information Processing and Management, Vol. 33, No. 3, 1997, pp. 319-337.

6. Mohamed, A. M., and Santhana, K., “Multimedia Descriptions Based on MPEG-7: Extraction and Applications”, IEEE Transactions on multimedia, Vol. 6, No. 3, 2004, pp. 459-468.