| 摘要: | 對於以內容為基礎的影像搜尋系統而言，常使用到的特徵有形狀(Sharp)、紋理(Texture)、以及顏色(Color)。所以，為了提升以內容為基礎的影像搜尋系統的正確性，找出更好的影像特徵描述方式是必要的。
本論文則是著重於紋理特徵的研究。對於過去已被學者提出幾種不錯的紋理描述方式，我們做了完整的分析與實驗。
過去對影像分析的研究有空間域的分析以及頻率域的分析兩方面，而本論文則是改進在空間域分析的方法–利用小波轉換(Wavelet Transform)之後所得到的高頻資訊的能量(Energy)；以及空間域分析的方法–利用Local Binary Pattern (LBP)可以統計紋理影像結構的特性，結合兩者的優點，提出另一個特徵–Local Energy Binary Pattern (LEBP)。
在實驗中，我們發現只有使用Local Energy Binary Pattern這個特徵時，由於分析的只有高頻資訊的影像，實驗結果沒有明顯的效果。但是，這個特徵若是結合過去學者提出在一般空間域分析的特徵Local Binary Pattern (LBP)或是Local Edge Pattern (LEP)，實驗後發現可以得到更好、更明顯的效果。
本論文實驗的資料庫有兩個紋理資料庫：Bordatz Textures以及Vision Texture(Vistex)。實驗後可以看出我們提出的LBP結合LEBP的特徵與LEP結合LEBP的特徵都有很好的結果。
最後使用兩種不同的距離量測方法–city block distance metric以及Euclidean distance metric，測試不同距離量測的方法得到的效果。
The major features for content-based image retrieval systems are shape, color, spatial relationship, and texture. For developing an effective content-based image retrieval system, it is desired to find good features to describe images.
In this thesis, we concentrate on texture features. Through complete experiments, we analyze and compare some other well-known texture features, such as local binary pattern (LBP), local edge pattern (LEP), wavelet energy, and gradient indexing, and also test and compare the distance measures, City Block distance and Euclidean distance, to measure the distance between two texture images. Finally we propose a texture image retrieval system with higher performance than all the compared ones.
After performing wavelet transform on images, the high frequency information and its energy can be obtained. Each such high frequency information possesses its own spatial property. Since the feature of “local binary pattern” (LBP) possesses the structure information of the texture images, we use the LBP operator to hold this spatial property of image after performing wavelet transform. We call this feature the “local energy binary pattern” (LEBP). We combine this feature of high frequency information, LEBP, with the other features, “local binary pattern” (LBP) and “local edge pattern” (LEP), as to be the feature for our system. The experimental results show that the proposed system has good performance when testing on both the databases of the Brodatz Textures and the Vision Textures (Vistex). |
