應用矩量保持技術於影像改良,壓縮與分析

Abstract

在本篇論文中，提出一組新的彩色矩量保持技術，並提出應用於彩色影像 銳化的方法。將影像中的每個圖點其臨近 n*n 個圖點之彩色矩量保持,由 公式解出兩組代表值，選出其中一組代表值作為銳化處理後的代表值,以 取代該圖點之原先值,就可達到彩色影像銳化的處理效果。其次本論文也 提出三種以矩量保持技術應用於彩色影像壓縮的方法。第一種方法是以預 定的線條或邊線形狀的二進位位元圖之索引編碼及可調式挑選所需的二進 位位元圖來改良區塊切割編碼法的壓縮結果。第二種方法是應用彩色矩量 保持原理和區塊切割編碼法的技術於彩色影像的壓縮。第三種方法則提出 一種新的彩色量化技術,並應用二值化之矩量保持技術,將每個彩色影像區 塊二值化為兩種彩色顏色和二進位位元圖,再利用邊線偵測法來偵測二進 位位元圖的邊線位置，並加以編碼,將可獲得好品值且較高的壓縮結果。 同時本論文也提出新的空間矩量保持技術應用於邊線的偵測法,使邊線的 偵測在不重疊區塊中直接公式解解出且邊線的偵測有次圖點的精密度。在 本篇論文中也提出一種降低彩色空間維度的方法,將區塊中的所有圖點投 射到區塊中的兩色向量差的方向,使三度空間的彩色值投射成一度空間的 存量值。如此 ,現有灰階影像的處理技術皆可應用於此方法中。結合前面 所提空間矩量保持技術應用於邊線的偵測法,可處理彩色影像的邊線偵測 。最後,本論文所提出的各個方法都有實驗證明其可行性與有效性,並附有 實驗結果。 First, a method of color image sharpening by moment- preserving (MP) is proposed. We develop a new set of color monents. By preserving some color moments of the n*n neighborhood of each pixel, analytic formulas are derived to compute two representative sets of tristimulus color values, and one of them is assign to the pixel as the sharpening result. Second, three image compression methods using MP techniques are proposed. The first method is an improvement of the BTC method. A set of line and edge bit planes are predefined and an adaptive bit plane selection scheme is developed. The second method is a color image compression scheme using MP and BTC techniques. We preserve the new set of color moments mentioned previously to obtain two representative colors and use the Euclidean distance measure to generate a bit-map for each image block. We also present a revised version of the proposed method to obtain a higher compression and good reconstructed image quality. In the third proposed method, a new approach to color quantization using MP thresholding and efficient coding of edge detection results on the bit map are included. Beside, a new approach to subpixel edge detection by mass MP is also proposed. The edge location can estimated directly from analytic formulas with subpixel accuracy. Finally, a method for reduction of color space dimensionality is proposed. After obtaining two representative colors in each image block, we employ the difference vector between the two colors as an axis onto which all the data in the block are projected to reduce the color space into one dimensional. A single spectral image is thus obtained, on which the proposed subpixel edge detection by mass MP is implemented to perform color edge detection. Experimental results are given to show the effectiveness of all the proposed methods.