Please use this identifier to cite or link to this item:
Texture Analysis Using Filter Bank and Vector Quantization
|Authors: ||Chen-Yuan Chong|
|Contributors: ||NTOU:Department of Electrical Engineering|
texture analysis;Gabor filter;vector quantization;block-matching;chi-square test
|Issue Date: ||2011-07-04
|Abstract: ||本論文提出一基於Gabor特徵與向量量化之區域相異度量測方法以改善非監督式紋理分析的效能。影像中任一點周圍區域之紋理差異表示其位於相異紋理區塊邊緣的可能性。基於Gabor基本函式善於描述空間與頻率的內容，以及藉由參數調設可使Gabor特徵達到位移、旋轉、尺度的不變性，紋理特徵可以不同空間頻率與方向的Gabor濾波器組所計算。而將高維度的Gabor特徵經向量量化為少數的代表向量，不但使區域內的紋理特徵能有效率地以直方圖描述，且區域間的紋理差異可以一改良式的卡方檢定求得。 本論文提出之改良式卡方檢定係在不同方向的兩相切圓形取樣區域間計算其紋理分佈差異，最後以最大卡方值作為此探測點的邊緣特徵。由於圓形區域的統計值具有旋轉不變性，每一個圓形區域的直方圖統計可供所有位於其圓周上的探測點利用，不惟可避免重複性的特徵統計，且比對方向個數愈多愈能有效減少直方圖的運算負擔。實驗結果證實該圓形區域統計資料的可重複利用性，確實能有效改善區域比對時的計算效能。|
This thesis presents an efficient method to measure the dissimilarity between two adjacent blocks for unsupervised texture analysis. The pairwise dissimilarity of texture blocks in a local neighborhood centered at a target point implies the probability of the point being located on a boundary between different texture regions. Because Gabor elementary functions can describe the space-frequency content and the fact that by properly setting filter parameters the Gabor features are invariant in translation, rotation and scale, texture features in this work are computed by Gabor filter bank with multiple frequencies and orientations. Furthermore, Gabor features in high dimension space are quantified to few representative vectors via vector quantization, which makes texture characteristics of local area be efficiently represented by histogram. The dissimilarities of adjacent blocks are then measured by the improved □2-test proposed in this thesis. The improved proposed □2-test employs two mask circles, rather than the conventional single circle, that intersect at a test point, and the block data points sampled in various orientations are compared for their dissimilarities. A point with the maximum dissimilarity value is taken as a boundary point. Because the feature of a circular block is rotational invariant, statistics of a block can be also utilized by test points located on the circumference of that circle. Such re-utilization would nicely avoid redundant feature calculations. Better yet, due to the re-utilization inclination of computation saving for histogram gets stronger as more orientations. Experimental results are provided to verify that, based on the rotation invariance and the re-utilization of feature statistics of the pair-circular mask. the proposed method indeed can improve the efficiency of texture analysis.
|Appears in Collections:||[電機工程學系] 博碩士論文|
Files in This Item:
All items in NTOUR are protected by copyright, with all rights reserved.