SIFT街景辨識之比對效率與準確度提升

Abstract

本研究承續了以SIFT影像技術為基礎的街景辨識系統，針對原系統所使用的全域式比對(Fully Matching)方法與基於K-D Tree所提出修改 Best-bin-first的加速搜尋演算法做延伸與改進，提出特徵點周圍灰階梯度累計主方向的特徵點分類比對方法，重新設計資料庫建置、測試新方法適合的最佳匹配閥值與KD Range參數，以達成匹配加速計算目的與高準確度匹配結果。透過特徵點主方向計算事前分類，當影像進行匹配時，來源影像特徵點僅需與其同方向特徵點資料庫作比對，可降低78%所需比對時間，若再結合原系統提出之K-D Tree加速搜尋比對方法(原可降低46%所需比對時間)，更可達到降低88%比對時間的最佳效果。雖然此方法對於匹配加速表現優良，但因為犧牲了部分來源比對資料，所以準確度較原方法降低2%。為了兼顧快速匹配與高準確度，本研究最後提出先以主方向加K-D Tree搜尋法快速過篩，再以篩出的少數候選圖片與來源影像進行全域式比對，實驗結果顯示此方法可在降低86%所需比對時間情況下，維持與SIFT原始比對法相近的高準確度。

This research presents a novel streetscape recognition system which improves the classical algorithm of SIFT (Fully Matching) and the modified BBF (Best-bin-first) searching algorithm (K-D Tree Searching). Our system can reduce computation by making comparisons only between the features of the same types according to features’ gradients major orientation. Hence we rebuild a database (DB) for features matching which depends entirely on orientation information. We use a lot of experiments to find out the best thresholds which suit relative methods. Our experimental results show that the proposed method can accelerate the performance of features matching without losing its accuracy. By categorizing feature points in advance, feature matching with major orientation (called MO) can effectively save 78% of the processing time. The processing time can save 88% by combining the modified BBF algorithm (where the original modified BBF can only save 46%). However, its accuracy will decrease 2% because some features are sacrificed in matching. To retain high accuracy and efficiency at once, we propose a hybrid method to find out image candidates rapidly by adopting the MO method and the modified BBF algorithm firstly, and to determine the final result by performing fully matching with image candidates. The experiment results show that this method can save 86% processing time with nearly the same accuracy.