近年來，局部影像描述元在各方面獲得了成功的應用，如：寬基線匹配 (Wide baseline matching) 、物件辨識 (Object recognition) 、紋理辨識 (Texture recognition)、影像檢索 (Image retrieval) 、機器人定位(Robot localization) 、影片資料探勘 (Video data mining) 等方面。
但在即時性方面，卻因複雜度較高，以軟體實作方面較不容易達到即時性的要求。本論文提出以SIFT（Scale-invariant feature transform，SIFT）和SURF(Speed Up Robust Feature，SURF)為基礎之SURF硬體設計以及SIFT和SURF的GPGPU平行化，來加速運算。
SIFT主要分成四部分：尺度空間極值偵測（scale-space extrema detection）、特徵點定位與篩選（Keypoint localization）、特徵點方向性（Orientation assignment）與特徵點描述（Keypoint descriptor），其中以尺度空間極值偵測與特徵點描述最占時間，也是無法及時的主要之處。所以硬體設計須分析時間成本與硬體面積成本之間的關係，找出可達即時性下之最小硬體面積成本，而這方面的研究相對SURF已經成熟很多。
SIFT主要分成四部分：積分影像（Integral image）、快速 Hessian 特徵檢測（Fast Hessian detection）、特徵點方向性（Orientation assignment）與特徵點描述（Keypoint descriptor）。而SURF演算法在複雜度上降低了許多，在軟體實作上離即時性更進一步，但仍然沒有達到即時的要求。硬體方面由於快速 Hessian 特徵檢測（Fast Hessian detection）以及特徵點描述（Keypoint descriptor）計算範圍龐大需要的硬體面積也相對龐大，如此本論文在演算法上左了些修改以及在硬體架構上提供了一個新的做法，使得硬體面積大為縮小運算速度上卻幾乎不受影響。

The description of local features of images has been successfully applied to many areas, including wide baseline matching, object recognition, texture recognition, image retrieval, robot localization, video data mining, etc. However, pure software implementations usually cannot achieve the requirement of real-time processing. In this thesis, we present software acceleration of general-purpose computing on graphics processing units (GPGPU) for two popular image feature extraction/description algorithms, Shift-Invariant Feature Transform (SIFT) and Speeded-Up Robust Feature (SURF). Furthermore, several versions of hardware SURF accelerators are also implemented. The four major parts of SIFT are scale-space extrema detection, keypoint localization, orientation assignment, and keypoint description where scale-space extrema detection and keypoint description, the most critical parts, take most of the total execution time. SURF is composed of four major steps: integral image calculation, fast Hessian detection, orientation assignment, and keypoint description. In terms of software implementation, the computation complexity of SURF is significantly reduced compared with that of SIFT. However, hardware acceleration of SURF is still required for real time processing requirement. In this thesis, we slightly modify the original SURF algorithms in order to significantly reduce the hardware complexity for the implementations of fast Hessian detection and keypoint description without sacrificing too much in speed performance. Experimental results of both software and hardware acceleration are also given and compared.

目錄（Table of Contents）
頁
論文審定書 i
中文摘要 ii
英文摘要 iii
圖目錄（List of Figures） vii
表目錄（List of Tables） x
第1章 概論（Introduction） 1
1.1 研究動機 1
1.2 本文大鋼 1
第2章 研究背景與相關研究 3
2.1 文獻回顧 3
2.2 GPGPU架構 4
2.2.1 OpenCL程式模組 6
2.2.2 OpenCL硬體架構 7
第3章 演算法流程 9
3.1 SIFT（scale-space extrema detection） 10
3.1.1 尺度空間極值偵測（scale-space extrema detection） 11
3.1.2 特徵點定位與篩選（keypoint localization） 15
3.1.4 特徵點描述（Keypoint descriptor） 19
3.2 SURF（Speeded Up Robust Features） 21
3.2.1 積分影像（Integral image） 23
3.2.2 快速 Hessian 特徵檢測（Fast Hessian detection） 25
3.2.3 決定主要方向（Orientation assignment） 29
3.2.4 SURF特徵向量方向建立（Keypoint descriptor） 30
第4章 軟體加速(Software SpeedUp) 32
4.1 SIFT（Scale-invariant feature transform） 34
4.1.1 色彩影像轉亮度影像（RGB to Intensity） 34
4.1.2 尺度空間極值偵測與特徵點定位與篩選（keypoint localization and scale-space extrema detection） 36
4.1.3 特徵點主方向及描述（Orientation assignment and Keypoint descriptor） 37
4.2 SURF（Speeded Up Robust Features） 38
4.2.1 積分影像（Integral image） 38
4.2.2 快速 Hessian 特徵檢測（Hessian detection） 39
4.2.3 SURF主方向及特徵向量方向建立（Orientation assignment and Keypoint descriptor） 40
4.3 實驗結果（experimental results） 41
第5章 硬體加速 48
5.1 SURF_V1 48
5.1.1 RGB to Intensity 49
5.1.2 Image Integrator 50
5.1.3 Integral Image Buffer 56
5.1.4 Hessian Operator 58
5.1.5 HOG Buffer& Extrema Operator 64
5.1.6 Descriptor 67
5.2 SURF_2 74
5.2.1 Down Sample 75
5.2.2 Intergral Image Buffer&Sram、HOG Buffer&Sram 77
5.3 SURF_3 80
5.3.1 Merge Octave2&3 80
5.3.2 Integral Buffer 82
5.3.3 Hessian Operator 83
5.3.4 Descriptor Operator 89
第6章 實驗結果分析與比較 94
6.1.1 邏輯數據合成分析 94
第7章 結論與未來展望 98
7.1 結論 98
7.2 未來展望 98

[1] D. G. Lowe, “Distinctive image features from scale-invariant keypoints,” Int. J. Computer Vision, vol. 60, no. 2, pp. 91-110, Jan. 2004.
[2] GPGPU, General-Purpose Computation on Graphics Hardware. [Online].
[3] C. Wu, “SiftGPU: A GPU implementation of scale invariant feature transform (SIFT),” [Online].
[4] H. Bay, A. Ess, T. Tuytelaars, and L. Van Gool, “SURF: speeded-up robust features,” Comput. Vis. Image Underst., vol. 110, no. 3, pp. 346–359, Jun. 2008.
[5] M. Brown and D. G. Lowe, “Invariant features from interest point groups,” in British Machine Vision Conf., pp. 656-665, 2002.
[6] H. Moravec, “Rover visual obstacle avoidance,” in Int. Joint Conf. Artificial Intelligence, pp.785-790, 1981.
[7] C. Harris and M. Stephens, “A combined corner and edge detector,” in Alvey Vision Conf., pp. 147–151, 1988.
[8] J.J. Koenderink, “The structure of images,” Biological Cybernetics, vol. 50, no. 5, pp. 363-396, 1984.
[9] T. Lindeberg, “Scale-space theory: A basic tool for analysing structures at different scales,” . Applied Statistics, vol. 21, no. 2, pp. 224-270, 1994.
[10] K. Mikolajczyk and C. Schmid. “A performance evaluation of local descriptors,” in Proc. Computer Vision and Pattern Recognition, Jun. 2003.
[11] 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.
[12] M. Schaeferling, G. Kiefer, "Object Recognition on a Chip: A Complete SURF-based System on a Single FPGA". International Conference on Reconfigurable Computing and FPGAs, Nov. 2011, pp.49-54.
[13] J. Švab, T. Krajnik, J. Faigl, L. Preucil, "FPGAbased Speeded up Robust Features". IEEE International Conference on TePRA 2009, pp.35-41.
[14] J. Fischer, A. Ruppel, F. Weisshardt, A. Verl, "A Rotation Invariant Feature Descriptor O-DAISY and its FPGA Implementation", IEEE International Conference on IROS, Sept. 2011, pp.2365-2370.
[15] V.Bonato, "A Parallel Hardware Architecutre for Scale and Rotation Inavariant Feature Detection", IEEE Transactions on Circuits and Systems for Video Technology, Vol.18, No.12, 2008