| 摘要: | 影像拼接已被廣泛應用於工業視覺、無人機拍攝、醫學成像等領域，影像拼接目的是創建寬視角的影像，但是在視差和場景照明等條件下容易造成嚴重的偽影。因此，提高影像對齊的精度是影像拼接中的一個重要研究問題。對齊的準確性在於如何從獲得的特徵點找到盡可能多的真實匹配，同時最大限度地減少錯誤匹配，這對於計算機視覺中許多基於特徵匹配的任務至關重要。
基本的影像拼接過程通常包括影像灰階化、特徵檢測、特徵匹配、扭曲、透視投影和影像融合。本論文提出了一種新穎的影像處理技術，結合熵和基於歸一化相關係數的模板匹配來搜索相似區域，熵的使用可以作為亂度分析來量化影像中之訊息量。首先，對影像進行分割並計算每個影像塊的熵值，再使用所有熵的平均值作為閾值。如果影像塊的熵低於閾值，則說明此影像塊為弱紋理區域，可以將此影像塊去除。然後，使用保留的影像塊作為模板並將其與另一影像進行匹配以搜索相似區域。最後，根據相似區域的灰度值分佈，計算ORB的自適應閾值，對相似區域進行特徵提取，以提高特徵匹配和影像對齊的準確性。實驗結果表明，我們的方法具有更高的特徵匹配精度和更好的拼接效果。
;Image stitching has been widely employed in industrial vision, aerial photography using unmanned aerial vehicles (UAVs), medical imaging, and other fields. Image stitching aims to create images with a wide viewing angle, but it easily cause severe artifacts under conditions such as parallax and scene lighting. Therefore, improving the accuracy of image alignment is an important research issue in image stitching. The accuracy of alignment lies in how to find as many ground-truth matches as possible from given feature points while minimizing false matches, which is crucial for many feature-matching-based tasks in computer vision.
The basic image stitching process usually includes image grayscale, feature detection, feature matching, warping, perspective projection, and image blending. This thesis proposes a novel image processing technique that combines entropy and normalized correlation coefficient-based template matching to search similar regions of different images. Entropy can be used as a randomness analysis to quantify the amount of information in an image. First, divide the image into blocks and then calculate the entropy value of each image block. Next, we use the average of all the entropy as the threshold value. If the entropy of the image block is lower than the average, it means that the image block is a weak texture area, representing the image block can be removed. Then, use each of the remaining image blocks as templates and match them to another image to search for similar regions. According to the gray value distribution of the similar area, the adaptive threshold value of ORB is calculated, and feature detection is performed in the similar region so as to improve the accuracy of feature matching and image alignment. Experimental results show that our proposed method has higher accuracy in feature matching and better stitching results. |
