王瑞祥
【摘要】本文基于花崗巖和砂巖數(shù)字圖像特征,利用小波分析理論及Bayes決策理論建立起巖石中幾種成份(云母、石英、長石)的頻譜圖。首先利用巖石圖像灰度統(tǒng)計函數(shù)存在多個極小值的特點,將其灰度級劃分成若干個子區(qū)間,并利用迭代算法對區(qū)間進行優(yōu)化,根據(jù)優(yōu)化所得區(qū)間來建立起各類的樣本集及其分布域。然后用小波理論對圖像進行多重分解,按塔式原則將其各級系數(shù)矩陣還原成與原圖像大小一致的矩陣,并對各矩陣進行均一化處理,經處理之后的小波系數(shù)矩陣為圖像的波段。最后,以樣本集為基樣本,求出小波分解的各級分解系數(shù)與對應點的坐標集及其分解系數(shù)集,利用Bayes算法建立花崗巖和砂巖中各成份的頻譜圖。本文中頻譜圖是建立在先驗基礎之上的,在對頻譜圖的應用時,只需將一幅圖片進行小波分解,同時對分解系數(shù)做還原及均一化處理,根據(jù)先驗所得的頻譜對樣本進行計算,便可確定出被分析圖像的各種成份及其分布情況。
【關鍵詞】小波分析;巖石圖像分類;頻譜;波段;樣本;樣本集;聚類中心
【Abstract】Based on the characteristics of granite and sandstone digital images, this paper builds up the spectrum of several components (mica, quartz and feldspar) in rock by wavelet analysis theory and Bayes decision theory. Firstly, the gray level is divided into several subintervals by using the gray level statistical function of the rock image. The iterative algorithm is used to optimize the interval. According to the optimized range, the sample sets are set up. Its distribution domain. Then the wavelet is used to decompose the image, and the matrix of the coefficients is reduced to the same size as the original image according to the tower principle, and the matrix is processed uniformly. The wavelet coefficients matrix after processing is the band of the image The Finally, the spectral set of granite and sandstone is established by Bayes algorithm, and the spectral set of the decomposition coefficient of the wavelet decomposition and the corresponding coordinate set and its decomposition coefficient are obtained. In this paper, the spectrum is based on the transcendental basis, in the application of the spectrum, only a picture of the wavelet decomposition, while the decomposition factor to do the reduction and uniform processing, according to a priori spectrum pairs The samples are calculated to determine the various components of the image being analyzed and their distribution.
【Key words】Wavelet analysis;Rock image classification;Spectrum;Band;Sample;Sample set;Clustering center
1. 前言
(1)對于圖像的分類,過去有很多學者對此做了很多的研究。在傳統(tǒng)方法上,人們利用對象與圖像背景之間的差別來識別對象,這些差別主要體現(xiàn)在圖像函數(shù)f(x)的一階導數(shù)和梯度沿圖像邊緣切線方向變化的趨勢較緩,而沿垂直圖像邊緣方向的變化趨勢較陡,經典的算法有:Roberts算子、Prewitt算子、Sobel算子、LOG算子、Canny算子等[1]。另外在利用邊緣檢測與圖像的數(shù)學形態(tài)學相接合,也能較好的識別出圖像中的對象[2,3]。
(2)巖土材料是由不同成份的物質組成,它們緊密交織,且其間隙十分小,因此邊緣檢測和數(shù)學形態(tài)學很難將它們分離出來[4,5]。Seungcheol Shin等[6]人利用小波理論對巖土材料進行分解,并用各級分解系數(shù)的能量特征,對土顆粒的尺度進行研究,取得了較好的效果。但是這一方法沒有考慮到圖像中不同物質成份的概率分布。
(3)花崗巖和砂巖圖像的灰度統(tǒng)計函數(shù)存在著若干個極小值,以這些極小值點為分界點,將灰度函數(shù)的定義域分成若干個子區(qū)間,用迭代算法對各區(qū)間進行優(yōu)化;以優(yōu)化所得子區(qū)間為依據(jù),將像素值屬于同一區(qū)間的點歸為一類;求出各類的分布區(qū)域、聚類中心等參數(shù)。同時用小波分解將圖像沿垂直、水平、對角三個方向分解;按塔式放大原則[7]將各級分解系數(shù)矩陣還原成與原始圖像大小相同的矩陣;根據(jù)所得分布域,把所有放大后的系數(shù)矩陣劃分成若干個子域;將某系數(shù)矩陣中位于同一分布域的點集視為類的一個波段。最后利用Bayes算法[8~10]求出每一類的先驗概率及其判別式方程,將方程中未知項視為某一類成份的頻譜。本文對若干個花崗巖和砂巖圖像進行分析,求出各類成份的頻譜,便可基于這些頻譜對花山崗巖和砂巖進行成份分析。