基於小波-熵-區塊處理及最小錯配誤差的影像鑲嵌系統

Abstract

本論文提出一個以特徵為基礎(feature-based)的影像鑲嵌(mosaic)系統。大多數的特徵都在邊緣、線、角落等能量變化較大的地方。但又為了不讓其受到雜訊影響，故應用多解析度小波理論(multi-resolution wavelet theory)的特性處理影像以得到好的特徵區域，再配合熵(entropy)及區塊處理來求得資訊含量較大且具區域性的特徵點，做為特徵。再利用特徵比對(feature matching)的技術，求出對應的轉換矩陣(transformation matrix)。並使用簡單的幾何公式在眾多特徵對中找出適合做為參考基準的特徵對(matching pair)，以減少轉換矩陣受到錯配誤差(mismatch error)的影響。 影像接合的結果，除了受特徵對的精確與否影響外，影像光源亮度和色彩差異的問題也是佔有一定程度的影響。若此問題考慮不周，在影像接合後，也會造成不自然的視覺效果，所以在影像接合前，先利用簡單的色彩特性，幫助我們選定一張彩色影像做為標準影像，然後做色彩校正(color correction)的動作，以期最後結果達成較佳的視覺效果。 由實驗結果可以看出，利用所提出的架構，所選取的可靠特徵對和標準影像的參考色彩，可使接合結果有較佳的視覺效果。

A feature-based image mosaic system is presented in this thesis. Most reliable features are located at places such as edges, boundaries, lines and corners where energy variances are big. To avoid distortion from unpredictable noises, multi-resolution wavelet theory is applied to obtain good feature regions and then combine with entropy and block process to get larger information capacity and more regional feature points. Feature matching is then used to obtain transformation matrix. Simple geometric formula is used to find feature pairs that could reduce mismatch error effects in transformation matrixes. The results of feature-based mosaic not only depend on the accuracy of matching feature pairs, but also greatly affected by luminance and color difference of image. The mosaic image may lead to a few disadvantages in human vision if such problem is not properly solved. Therefore, before combining images, simple color characters are chosen to help us pick a reference image, and then color correction could be performed for get better results. The experimental results shows that the proposed method has better visual effect due to reliably extract matching feature pairs and reference image choice.