| 摘要: | 影像匹配已被廣泛應用於機器視覺、無人機、圖像識別等領域，影像匹配的目的在於辨認影像特徵相同處，為後續更複雜的影像技術提供資訊，但在影像視角差距越大時，一般方法所獲取的特徵點越少，甚至沒有，而在實際情況下，匹配影像的視角差距是不可避免的，因此有人提出了針對大視角影像匹配方法，但卻伴隨著計算成本的增加，使得運算時間增加，所以對於如何減少計算成本並保持大視角的影像匹配效果至關重要。
大視角的影像匹配包含對影像做仿射投影變換、特徵點萃取、特徵點匹配。本論文針對仿射投影變換的部分，提出了自適應分段近似演算法，對Lanczos核做低階多項式近似，使得原本Lanczos插值法中三角函數的計算成本降低為分段的低階多項式運算成本，演算法使用到最小平方法來對方程式執行曲線近似，並利用受限制的最小平方法來保持分段曲線間的連續性，也使用到主成分分析法來自適應判斷分段數量，另外也優化了仿射投影變換的經度採樣，以減少計算成本同時保留大視角的影像匹配效果。
實驗結果也驗證了所提出方法的有效性，針對仿射尺度不變特徵轉換的演算法計算成本進行優化，成功的改善仿射投影變換使用Lanczos插值法的計算成本，同時合理的改善仿射投影變換中對不敏感區域的經度採樣，實現仿射尺度不變特徵轉換之加速。
;Image matching has been widely used in machine vision, drones, image recognition and other fields. The purpose of image matching is to identify the same features of images and provide information for subsequent more complex imaging technologies. However, the greater the difference in image viewing angle, the fewer feature points can be obtained by general methods, or even no feature points can be obtained. In actual situations, the viewing angle difference of matching images is inevitable. Therefore, some people have proposed a matching method for large-viewing angle images, but it is accompanied by an increase in computational cost, which increases the operation time. Therefore, it is crucial to reduce the computational cost and maintain the matching performance of a large-viewing angle.
Large-viewing angle image matching includes affine projection transformation, feature point extraction, and feature point matching. This paper proposes an adaptive piecewise approximation algorithm for affine projection transformation, which makes a low-order polynomial approximation to the Lanczos kernel, so that the computational cost of trigonometric functions in the original Lanczos interpolation method is reduced to the cost of piecewise low-order polynomial operations. The algorithm uses the least squares method to perform curve approximation on the equation, and uses the constrained least squares method to maintain the continuity between piecewise curves. It also uses principal component analysis to adaptively determine the number of segments. In addition, it also optimizes the longitude sampling of the affine projection transformation to reduce the computational cost while retaining the large viewing angle matching performance.
The experimental results also verify the effectiveness of the proposed method. The algorithm computational cost of affine scale-invariant feature transformation is optimized, and the computational cost of the Lanczos interpolation method used in affine projection transformation is successfully improved. At the same time, the longitude sampling of insensitive areas in affine projection transformation is reasonably improved to achieve the acceleration of affine scale-invariant feature transformation. |
