赫夫轉換的延伸和其應用

Abstract

傳統型的赫夫轉換可在有雜訊的影像中偵測直線、圓，或橢圓，而傳統的 通用型赫夫轉換可在有雜訊的影像中偵測和定位任意形狀的二度空間平面 物件。 然而它們仍有一些缺點: 不適合直接處理有灰階的，彩色的和任 意透視轉換的影像。 針對這些缺點，我們提出了四種赫夫轉換的延伸及 其應用，做各種的改進。 第一種方法是直接對有灰階的影像做赫夫轉換 ，事先不必做影像的二元化，細化或邊緣偵測。 並利用轉換的結果，直 接偵測有寬度的灰階粗線條。第二種方法類似第一種方法，可用來直接偵 測有寬度的圓。第三種方法是利用改進的通用型赫夫轉換來辨認彩色二度 空間的平面物件。由於不同的環境光源會引起彩色影像不同的變化， 且 彩色影像原本較不易處理，因此通用型赫夫轉換很少應用於彩色物件的辨 認。在此研究中所提的方法，假設環境光源為白光或一定混合比例的紅、 綠、藍的三原色光，用正規化的紅、綠、藍三原色光可消除因燈光變化所 引起的各種變化, 再利用每一圖素及其周鄰的圖素的正規化的色彩來建立 此圖素的特徵向量。在赫夫轉換累計空間，計算並累加輸入圖及欲辨認的 物件兩者間每對圖素的色彩特徵向量的相似度。然後，檢測最佳累加值。 若累加值大於某個門檻值，則彩色物件被偵測存在輸入圖中，並定位於最 大累加值的位置。第四種方法是提出一種新的通用型赫夫轉換，可直接偵 測任意透視轉換的二度空間之平面物件，簡稱透視轉換不變的通用型赫夫 轉換。此新方法首先需建立一個二度空間的透視轉換參考表含欲偵測的物 件從各種不同的觀測位置所觀察到的透視轉換資訊，再用它進行通用型赫 夫轉換的累加計算。然後，檢測累加值，若累加值大於某個門檻值即獲得 欲偵測的候選位置。取得正確的位置及透視轉換的參數。 Several approaches based on new extensions of the Hough transform (HT) and their applications are proposed for overcoming several problems about processing analyzing gray-scale, color, and perspective images encountered in the use of the conventional Hough transform (CHT) and the generalized Hough transform (GHT). First, a gray-scale Hough transform (GSHT) for thick line detection in gray-scale images is proposed. Second, a method for thick circle detection by an extended HT (EHT) is studied. Third, a modified GHT (MGHT) is proposed for color image detection and matching. Fourth, a perspective transformation invariant generalized Hough transform (PTIGHT) is proposed for perspective planar shape detection and matching. Finally, a new approach to position and orientation tracking using the PTIGHT for virtual reality applications is presented. Usually, the use of the CHT requires the preprocessing steps of thresholding and edge detection (or thinning) before the transform can be performed to detect thick lines (called linear bands) in a gray-scale image. This causes loss of useful gray and position relationship existing among the pixels of a linear band, and requires certain postprocessing step to recover the band in the original image. The proposed GSHT with a gray-scale image as the direct input removes this shortcoming, reqiring neither preprocessing nor postprocessing step in detecting the bands in the image. The MGHT is proposed to remove this weakness. First, lighting changes in an input color image are removed using normalized color values. Next, certain critical pixels of the input image whose neighborhoods have larger variances of color values. For each critical pixel, a feature vector, which includes the normalized color values of the pixel as well as of the pixel's neighbors, is then constructed. A modified voting rule for the GHT is proposed accordingly, which is based on a