以小波轉換為基礎之影像匹配法

Abstract

隨著電腦視覺及影像處理技術之發展，攝影測量共軛點的量測方式已由傳統的人工選點轉變為自動匹配，各種不同的影像匹配演算法也逐漸被提出。Lowe (2004)所提出的SIFT演算法與Bay et al. (2008)提出的SURF演算法即屬於特徵式匹配法，已被廣泛使用於電腦視覺及攝影測量領域中。由於SIFT與SURF在進行特徵點偵測時，皆必須進行大量的計算，並以經驗法則決定相關的門檻值，此外在特徵點偵測與描述元計算為兩個不同的計算步驟，因此運算效能上較為費時。小波轉換法是一種良好且具有完整數學理論的資料轉換方法，其所具有的多重解析度特性與SIFT找尋關鍵點所使用的高斯金字塔概念類似，但小波轉換具有較佳的極值點搜尋效能。過去研究中小波轉換大都用於影像中明顯的角點偵測或邊緣線偵測，本研究則希望建構一套以小波轉換為基礎之影像匹配法，希望能利用一次小波轉換後的數值完成影像匹配的步驟，發展出一套較有效率的方法。在本研究的實驗中分別以模擬影像以及實際影像進行測試與分析，於模擬影像中探討不同特徵點的特性，並於實際影像中檢驗研究流程中各步驟方法之適用性，並以最小二乘匹配法、相對方位解算成果以及核線幾何對應的方法進行成果評估，實驗成果顯示分別在四組影像相比下，除了LSM的評估成果本研究方法約為1個pixel上下而SIFT與SURF的精度小於1個pixel之外，在相對方位以及核線幾何的成果評估當中，本研究之匹配成果可優於SURF的方法，雖然無法優於SIFT的方法，但可達到與SIFT相近的精度。

With the development of technology in both computer vision and image processing, a variety of image matching algorithms have been gradually proposed. The SIFT algorithm proposed by Lowe (2004) and the SURF algorithm proposed by Bay et al. (2008) belong to a kind of feature-based image matching and are widely used in both computer vision and photogrammetry field. In the stage of feature detection, no matter in SIFT or SURF, there is a great amount of calculation in finding extremes and the thresholds for keypoint localization are determined by experience. Besides, it is time-consuming when doing keypoint detection and producing keypoint descriptors in difference methods in both SIFT and SURF. Wavelet transform is one of the most popular analysis tools of the time-frequency transformation. The basic concept of wavelet multiresolution analysis is very similar to the Gaussian pyramid which is used in SIFT, but wavelet transform can detect the extreme points more accurately and has a better efficiency in searching extreme points. Generally, the wavelet transform is commonly used to detect the distinct features such as corners and edges. In this study, the wavelet transform is used to construct a feature-based image matching method for both detecting keypoints and calculating descriptors. In the experiment of this study, the proposed method is applied on both simulated images and a set of aerial images. The simulated image is used to investigate the characteristic of the keypoints in difference methods. And the aerial images are used to verify the performance of each step in the proposed method. The performance are illustrated comparing with the SIFT and SURF algorithm for least square matching (LSM), solving the relative orientation and epipolar geometry. The LSM results show that the error of conjugate point is about 1 pixel in our method and less then 1 pixel in both SIFT and SURF. But in the relative orientation and epipolar geometry result, our method performs better than SURF algorithm and is close to SIFT algorithm.