以圖形處理器加速尺度不變特徵轉換演算法

Abstract

圖像內容檢索（CBIR）為一種電腦視覺技術，使用圖像的內容在大型資料庫中進行檢索，如顏色、形狀、紋理等而非關鍵字、標籤或其他描述方法。許多圖像運算與電腦視覺的技術皆需要擷取圖像中的內容，大部分皆透過尺度不變特徵轉換演算法（SIFT）來達成。尺度不變特徵轉換被廣泛應用於物件辨識、圖像拼接、立體對照、描述圖像特徵等。在特定的應用如圖像內容檢索中，特徵點擷取被視為預處理程序，之後的特徵點比對便成為運算最密集的程序。圖形處理器（GPU）因其在大量資料平行運算的卓越能力收到關注，因此，本研究提出基於圖形處理器平行化的尺度不變特徵轉換演算法，藉此加速線性搜尋法與k-近鄰搜尋法。於實驗結果中，相較於傳統的最近鄰居搜尋法，本研究得到22倍的加速；而相較於傳統的k-近鄰搜尋法，得到11倍的加速。

Content-based image retrieval (CBIR) is the application of computer vision techniques to the searching for digital images from large databases using image actual contents such as colors, shapes, and textures rather than the metadata such as keywords, tags, and/or descriptions associated with the image. Many techniques of image processing and computer vision are applied to capture the image contents. Among them, the scale invariant features transform (SIFT) has been widely adopted in many applications, such as object recognition, image stitching, and stereo correspondence to extract and describe local features in images. In certain application such as CBIR, feature extraction is a preprocessing process and feature matching is the most computing-intensive process. Graphic Processing Units (GPUs) have attracted a lot of attention because of their dramatic power of parallel computing on massive data. In this thesis, we propose a GPU-based SIFT by accelerating linear search and K-Nearest Neighbor (KNN) on GPUs. The proposed approach achieves 22 times faster than the ordinary Nearest Neighbor (NN) performed on CPUs, and 11 times faster than the ordinary linear search and KNN performed on CPUs.