應用多重特徵於提升材質辨識的準確性

Abstract

現今電腦視覺應用越來越廣泛，其中材質辨識是一個重要的課題。實證分析發現材質的紋理會因為光線明亮、角度旋轉、尺度變化以及雜訊干擾降低辨識系統的準確度。本研究選取常見的四種紋理特徵做為辨識依據，分別為灰階共生矩陣、局部二值化模式、局部模式共生矩陣和紋理基元法。根據特徵特性探討，選擇單一特徵只能解決上述部分問題。舉例來說，灰階共生矩陣僅具有辨識角度旋轉問題的能力，當相同紋理因光亮的變化而有所不同時，將無法正確的辨識。因此本論文提出結合數個紋理特徵的方法進行材質辨識，稱之多重特徵法。實驗結果證明多重特徵法確實有效提升材質辨識的準確度。

Recently, the use of computer vision has become more popular. One of the important uses is texture recognition. There are four major problems about the texture recognition: light, scale, angle, and noise. This thesis proposed four texture features to solve the problems: Gray Level Co-occurrence Matrix, Local Binary Pattern, Local Pattern Co-occurrence Matrix and Textons-based Approach. However, using only one of the texture features, we couldn’t solve these problems at one time. For example, using Gray Level Co-occurrence Matrix as the texture feature only solved the problem of rotational change. And the problem of light illumination still remains. The thesis proposes a new method that combines multiple features. The new method could enhance accuracy of texture recognition. Thus the experiment proved that the method could enhance accuracy of texture recognition.