利用方向性次頻分解之紋理結構影像分類及自動分割

Abstract

自然界中,大部份物體都具有類似紋理(texture)結構般的表面,如何對這 些影像做適當的分類以及分割一直是電腦視覺,影像處理,圖形識別等許多 研究領域上所面臨的一項重要課題。針對這類問題,本文提出利用多頻段 方向性次頻濾波器(multichannel directional subbnad filters)解決之 。在紋理影像分類中,我們把每個經過方向性次頻濾波器的影像假設為馬 可夫隨機場(Markov Random Field)模式,紋理影像特徵即為此模式的參數 。對於紋理影像的自動分割,我們分為四個步驟:特徵萃取,粗略分割,精細 分割,及最後處理。紋理特徵取自於方向性相關係數,在粗略分割中我們提 出一快速而有效率的演算方法來作紋理的分類。如何估計出真正的紋理影 像個數,是影像自動分割中最具挑戰性,也是最困難的工作。文中,我們提 出一有系統的判斷方法來解決這難題。它是結合每個經過方向性次頻濾波 器影像的分類結果,來決定出紋理影像的個數。電腦模擬所得到的結果,驗 證了本文所提出方法的優越性。 This thesis presents both texture classification and segmentation algorithms by multichannel decomposition. The channels are characterized by a bank of directional subband filters that allow a two-dimensional input signal to be represented by a sum of maximally decimated subband images and perfectly reconstructed from these decimated ones. For classification, we model the filtered channel image as Markov Random Field (MRF) and the model parameters are then extracted as texture features. For texture segmentation, four stages are taken, namely, features extraction, coarse segmentation, fine segmentation, and post processing. Correlations in each channel are used as features. At coarse segmentation stage, a fast and efficient clustering algorithm by incorporating the clusters' spatial locations is introduced. To estimate the true number of textures (cluster validility problem), we propose a new cluster- number decision algorithm by integrating each channel's clustering result. Simulation results demonstrate the effectiveness of our proposed texture classification and segmentation algorithms.