現今人們通過圖像編輯軟件，可以在不留痕蹟的情況下輕鬆修改這些圖像。這對醫學，刑事調查，監視系統，財務文件都造成一定的威脅，因此對於偽裝影像的偵測是一項非常重要的過程，其中以Copy-move forgery detection (CMFD)是最常見的，而copy-move則是將圖像中的一個或多個區塊粘貼在同一圖像中的其他位置。該過程可能伴隨旋轉，調整大小，壓縮和添加雜訊。CMFD中有一種方式是以關鍵點為主(keypoint-based) 的方法，主要是在匹配過程中通常是計算圖像裡關鍵點的距離，而距離小於閾值則進行匹配。但為了提高準確率，會比較最小距離和第二小距離的比，當這個比小於某個值，才視為可以匹配的點，但這容易刪除過多相似的關鍵點，導致難以尋找多個經由複製而成的偽裝區塊，並且當這些區塊中有平滑的區塊，關鍵點就不易取得。為了克服平滑區域不易獲得特徵以及無法對多個偽裝區塊進行偵測，因而提出本論文。
本論文主要目的是解決CMFD檢測偽裝物體時，對兩個以上的偽裝區塊進行偵測，並對平滑區域不易獲得特徵的區域，找出平滑的裝區塊，因而提出一種基於關鍵點與區塊的機器學習分類方式的檢測演算法。主要處理步驟如下：首先第一把圖像轉成Lab圖像與灰階圖像，作為之後關鍵點分類的選項。其次特徵點提取，採用Scale-invariant feature transform (SIFT)，但在(SIFT)在平滑的區塊容易檢測不到關鍵點，於是在使用(SIFT)進行處理時，需降低其不符合需求的關鍵點的閾值。其三執行機器學習K-means演算法，將Lab圖像的a和b通道進行分類，並比對關鍵點的位置，分出相似的關鍵點。其四把同樣類別的關鍵點進行歐式距離的匹配並對匹配過後的匹配點進行灰階值得比對，當匹配點的灰階值差距太大，則刪除該匹配點，其五將匹配點進行Density-based spatial clustering of applications with noise(DBSCAN)，把不夠密集的匹配點做濾除的動作。最後通過圖像的形態學處理把聚集在一起的匹配點，進行膨脹(dilation)，使匹配點成為物件區塊，並通過計算精準度(precision)，召回率 (recall)和F1 值檢測多偽裝物體的準確度。

Nowadays, people can easily modify these images without leaving traces through image editing software. This poses a certain threat to medicine, criminal investigation, surveillance system, and financial documents, Therefore, the detection of camouflage images is a very important issue, in which copy move forgery detection (CMFD) is the most common technology. Copy-move pastes one or more memory blocks in the image to other locations in the same image. This process may be accompanied by rotation, resizing, compression, and adding noise. One method in CMFD is a keypoint-based mechanism, which is mainly to calculate the distance between key points in the image in the matching process, and matches if the distance is less than the threshold. However, in order to improve the accuracy, the ratio of the minimum distance to the second small distance will be compared. When this ratio is less than a certain value, it will be regarded as a matching point. However, this condition is easy to delete too many similar key points, making it difficult to find more than two forgery memory blocks. When there are smooth memory blocks in these memory blocks, key points are difficult to obtain. In order to overcome the difficulty of obtaining the features of the smooth region and the inability to detect more than two camouflaged memory blocks, this thesis is proposed to solve these two problems.
The main purpose of this thesis is to detect more than two camouflage memory blocks. Owing to the CMFD method detects smooth camouflage blocks, it is difficult to obtain the corresponding features. Therefore, a detection algorithm of mechanical learning classification based on key points and memory blocks is proposed. The main processing steps are described as follows. (1) Image preprocessing is used. It is converting the image into Lab color space and gray-scale image as the option of key point classification. (2) Scale invariant feature transform (SIFT) is used for feature point extraction, but it is not easy to detect key points in smooth memory blocks. Therefore, it is necessary to reduce the threshold of key points that do not meet the requirements in the SIFT process. (3) Based on the machine learning K-means algorithm, the a and b channels of Lab image is classified. According to the position of those key points, similar key points are distinguished. (4) Matching the key points of the same category with Euclidean distance, and comparing the gray-scale values of the matched points. When the gray-scale value difference of the matching points is too large, these matching points are detected. (5) Conducting density based spatial clustering of applications with noise (DBSCAN) on the matching points, and filtering the matching points that are not dense enough. (6) The matching points gathered together are dilated through the morphological processing of the image, so that the matching points become object memory blocks, and the accuracy of multi camouflage objects is detected by calculating precision, recall, and F1 value.

摘要...........................................................................i
Abstract.....................................................................iii
誌謝..........................................................................vi
目錄.........................................................................vii
表目錄.........................................................................x
圖目錄........................................................................xi
符號說明 ......................................................................xv
第一章 緒論...................................................................1
1.1 簡介...................................................................1
1.2 研究動機與目的..........................................................1
1.3 論文架構...............................................................2
第二章 文獻探討及基礎理論 ...........................3
2.1 複製移動偽造檢測(copy-move forgery detection)...........................3
2.1.1 基於區塊(block-based)的檢測方式.........................................3
2.1.2 基於關鍵點(keypoint-based)的檢測方式....................................5
2.2 基礎理論說明..............................11
2.2.1 CIELAB color space 色彩空間算法..............................12
2.2.2 Scale-invariant feature transform (SIFT)算法.........................12
2.2.3 K-means算法..............................17
2.2.4 Density-based spatial clustering of applications with noise noise (DBSCAN) 算法..............................19
2.2.5 傳統SIFT特徵點用於CMFD的方法..............................22
第三章 系統架構..............................26
3.1 圖像預處理 (Image preprocessing)..............................28
3.2 K-means 分群 (K-means clustering)..............................28
3.3 Scale Invariant Feature Transform improvement.........................32
3.4 關鍵點之間的匹配 (Keypoints matching)..............................36
3.5 關鍵點匹配的灰階值比對..............................43
3.6 Density-based spatial clustering of applications with noise (DBSCAN) 分群..............................44
3.7 物件區塊組合(Object block combination)..............................51
第四章 實驗結果與討論..............................53
4.1 實驗環境平台..............................53
4.2 實驗視訊測試資料庫..............................53
4.3 實驗參數設定..............................56
4.4 實驗結果與分析..............................57
4.4.1 偽裝區塊為平移偵測結果..............................58
4.4.2 偽裝區塊為旋轉偵測結果..............................61
4.4.3 偽裝區塊為縮放偵測結果..............................68
4.4.4 整體效能評估..............................72
4.5 討論..............................75
第五章 結論與未來展望..............................80
文獻參考..............................81
Extended Abstract..............................84

[1] Yuanman Li and Jiantao Zhou, "Fast and effective image copy-move forgery detection via hierarchical feature point matching," IEEE Transactions on Information Forensics and Security, vol. 14, no. 5, pp. 1307 - 1322, May 2019.
[2] David G. Lowe, "Distinctive image features from scale-invariant keypoints," International Journal of Computer Vision, vol. 60, no. 2, pp. 91-110, January 2004.
[3] Tawheed Jan Shah, and M. Tariq Banday, "Copy-move forgery detection using hybrid transform and K-means clustering technique, " 2017 3rd International Conference on Applied and Theoretical Computing and Communication Technology (iCATccT), INSPEC Accession Number 17858847, pp. 21-23, Dec. 2017.
[4] Edoardo Ardizzone, Alessandro Bruno, and Giuseppe Mazzola, "Copy–move forgery detection by matching triangles of keypoints," IEEE Transactions on Information Forensics and Security, vol. 10, no. 10, pp. 2084 - 2094, Oct. 2015.
[5] Gul Muzaffer, Guzin Ulutas, and Beste Ustubioglu, "Copy move forgery detection with quadtree decomposition segmentation," 2020 43rd International Conference on Telecommunications and Signal Processing (TSP), INSPEC Accession Number 19875957, 7-9 Jul. 2020.
[6] R. Nithiya and S. Veluchamy, "Keypoint descriptor based copy and move image forgery detection system," 2016 Second International Conference on Science Technology Engineering and Management (ICONSTEM), INSPEC Accession Number 16285355, 30-31 Mar 2016.
[7] Jian Li, Xiaolong Li, Bin Yang, and Xingming Sun, "Segmentation-based image copy-move forgery detection scheme," IEEE Transactions on Information Forensics and Security, vol. 10, no. 3, pp. 507 - 518, Mar. 2015.
[8] Badal Soni, Pradip K.Das, and Dalton "Keypoints based enhanced multiple copy-move forgeries detection system using density-based spatial clustering of application with noise clustering algorithm," IET Image Processing, vol. 12, no. 11, pp. 2092 - 2099, Sep. 2018.
[9] Cai-Wei Shu and Xue-Zhong Xiao, "ORB-oriented mismatching feature points elimination," 2018 IEEE International Conference on Progress in Informatics and Computing (PIC), INSPEC Accession Number 18638512, 14-16 Dec. 2018
[10] Amanpreet Kaur, Savita Walia, and Krishan Kumar, "Comparative analysis of different keypoint based copy-move forgery detection methods," 2018 Eleventh International Conference on Contemporary Computing (IC3), INSPEC Accession Number 18249429 ,pp. 518–522, 2-4 Aug. 2018
[11] J. Hartigan and M. Wong, "Algorithm AS 136: A k-means clustering algorithm." Applied statistic, vol. 28, pp.100-108. 1979.
[12] Martin Ester, Hans-Peter Kriegel, Jörg Sander, and Xiaowei Xu,"A density-based algorithm for discovering clusters in large spatial databases with noise," The Second International Conference on Knowledge Discovery and Data Mining (KDD-96), pp. 226-231, 1996.
[13] Badal Soni, Pradip K.Das, and Dalton Meitei Thounaojam, "CMFD: a detailed review of block based and key feature based techniques in image copy-move forgery detection," IET Image Processing, vol. 12, no. 2, pp. 167 - 178, Feb. 2018.
[14] Shaharyar Ahmed Khan Tareen, and Zahra Saleem, "A comparative analysis of SIFT, SURF, KAZE, AKAZE, ORB, and BRISK," 2018 International Conference on Computing, Mathematics and Engineering Technologies (iCoMET), INSPEC Accession Number 17732676, 3-4 Mar. 2018
[15] P. J. Alessi , E. C. Carter , M. D. Fairchild , R.W.G. Hunt ,C.S. McCamy , B. Kranicz ,J.R.Moore ,L. Morren ,J.H. Nobbs , Y. Ohno, M.R. Pointer , D.C. Rich , A.R. Robertson , J.D. Schanda , R. Seve , P.W. Trezona ,K. Witt ,and H. Yaguchi, "COLORIMETRY, 3RD EDITION," CIE 015-2004, ISBN. 9783901906336v, 2004.
[16] Hao Luo, Wei Jiang, Youzhi Gu, Fuxu Liu, Xingyu Liao, Shenqi Lai, and Jianyan, "A Strong Baseline and Batch Normalization Neck for Deep Person Re-Identification," IEEE Transactions on Multimedia, vol. 22, no. 10, pp. 2597 - 2609, Oct. 2020.