小波理論應用於影像套合之研究

Abstract

本研究藉由小波轉換進行特徵點萃取，並與TDGO興趣運算元比較，測試兩者對於旋轉量及雜訊之反應。續以改良自Hsieh et al. (1997)所發展，應用小波特徵萃取之影像套合流程，與洪偉嘉（1999）所發展，以TDGO運算元配合最小二乘匹配法（LSM）之流程，進行數值測試並分析比較。 經實驗發現，就特徵萃取而言小波方法對雜訊之忍受度較TDGO運算元為高；對影像間旋轉量之表現差異則並不明顯。就套合流程而言，洪偉嘉（1999）之流程雖具備較嚴密之特徵點分佈控制，但仍出現鑲嵌成果精度不均之現象。改良自Hsieh et al. (1997)之流程控制點配對品質不如洪偉嘉（1999）之流程，但經適當門檻值設定後，可達成完全自動化之成果。

In this study, a wavelet-based edge detector is adopted for extracting feature points from image. In comparison with the TDGO interest operator, the performance on noise tolerance and rotation invariance is investigated. Furthermore, the image registration scheme derived from Hsieh et al. (1997) with a wavelet-based edge detector as well as the one developed by Hong (1999) with the TDGO interest operator and the least square matching method are tested with real images. The wavelet-based method is proven to have higher noise tolerance than the TDGO operator. In regard to rotation invariance, the performances of both methods are about the same. The scheme developed by Hong (1999) is inplemented with a strict control over feature point distribution, however, local inaccuracy still can be observed. The scheme derived from Hsieh et al. (1997) provides fully automation with properly given thresholds.