利用GPU加速SIFT特徵之擷取與比對

Implementation of fast SIFT feature extraction and matching using GPU

王培學(Pei-hsueh Wang)

指導教授 ： 范國清

國立中央大學/資訊電機學院/資訊工程學系/碩士(2011年)

摘要

GPU 於1999 年由NVIDIA 所提出來的一個硬體架構，主要用於輔助CPU做快速的影像平行運算，由於GPU 功能之強大，自從產品發表之後，陸續不斷有學者投入研究，針對現有的演算法進行平行化，以讓原本的系統能更有效率的發揮其功效。
本論文實作出一個海關貨櫃的認證系統。系統一開始由兩台在不同海關關口的攝影機拍攝二張貨櫃影像，二張影像即刻透過實作於GPU 上SIFT 演算法分別擷取出特徵點，並透過這些擷取的特徵點找出兩張影像的比對點。這些比對點基本上可計算出描述兩張影像對應關係的單應性矩陣，透過RANSAC 可以從對應關係中挑選出最佳對應的單應性矩陣。接著將其中一張影像透過單應性矩陣轉換後，將影像切割成多個相同的區塊，並利用區域二元特徵來做影像的比對，其比對的方法採用histogram intersection 來判斷兩個區塊是否相同。
實驗結果顯示在GPU 平台實作的特徵點擷取和比對步驟只需不到0.4 秒的時間，比一般英特爾CPU 實作出的結果快上10 倍多的速度。

關鍵字

CUDA ； 區域二元特徵 ； GPU ； SIFT

並列摘要

Graphic processing units (GPU), announced by NVIDIA Co. in 1999, is a specially designed circuit for parallelization. According to the high computational power of GPU, many researchers have devoted to commercial product designs or
academic researches. Recently, image processing algorithms are developed on this platform to improve the performance.
In this thesis, an authentication system for the Customers’ containers is developed. First, two images are captured from two different cameras in two gateways. The SIFT features are efficiently extracted from the parallel operations implemented on a GPU platform. Basically, the features for each image pixel are independently and simultaneously computed. Using the extracted features, the corresponding points between two container images are matched. Given the
corresponding point set, a homographic matrix is found using the RANSAC algorithm. After finding the homographic matrix, the corresponding point relations are constructed. A container image in a gateway is next separated into several
blocks, and the local binary pattern (LBP) features for each block are extracted.
Similarly, the corresponding LBP features for the image captured from the other gateway are also extracted using the found homographic matrix. The similarity for two images is calculated using the histogram intersection to determine if they are the same container or not.
The experimental results demonstrate the performance of feature extraction for image matching on the GPU platform. Less than 0.4 seconds are needed which is 10 times faster than that of the Intel-based CPU.

並列關鍵字

GPU ； CUDA ； SIFT ； LBP

參考文獻

[ 2 ] D. G. Lowe, “Distinctive image features from scale-invariant keypoints,” Int. J. Comput. Vis., vol. 60, pp. 91–110, 2004.

[ 4 ] T. Ojala, M. Pietikainen, and T. Maenpaa, “Multiresolution gray-scale and rotation invariant texture classification with local binary patterns,” IEEE Trans.

[ 5 ] K. Mikolajczyk and C. Schmid, “A performance evaluation of local descriptors,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 27, no. 10, pp. 1615–1630, Oct. 2005.

[ 7 ] M. A. Fischler and R. C. Bolles, “Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography, “ Comm. of the ACM, vol. 24, no. 6, pp. 381－395, 1981.

[ 8 ] A. Psyllos, C.N. Anagnostopoulos, and E. Kayafas, “Vehicle logo recognition using a SIFT-based enhanced matching scheme”, IEEE Trans. on Intell. Transportation Syst., vol. 11, no. 2, pp. 322－328, June 2010.