Multi-Scale Fusion for Improved Localization of Malicious Tampering in Digital Images Source: IEEE Transactions on Image Processing, Vol. 25, pp. 1312-1326, March 2016. Authors: Paweł Korus, Jiwu Huang Speaker: Jia-Long Lin Date: 2017/04/22
Outline Introduction Related work Proposed Scheme Experimental Results Conclusions
Introduction Digital image Sliding window Detection image 我們有一張圖像 要檢測他有沒有被竄改 在圖像沒有嵌入浮水印的情況下 分析窗口是一個常用的篡改檢測技術
Color space transformation Discrete cosine transform Related work JPEG (Joint Photographic Experts Group) Encoding Color space transformation Downsampling Block splitting Image JPEG壓縮的步驟 Entropy coding Quantization Discrete cosine transform
Related work Compression ratio Quantization matrix of 100% 16 11 10 24 40 51 61 12 14 19 26 58 60 55 13 57 69 56 17 22 29 87 80 62 18 37 68 109 103 77 35 64 81 104 113 92 49 78 121 120 101 72 95 98 112 100 99 量化矩陣決定JPEG的壓縮比 Quantization matrix of 100% Quantization matrix of 50%
Related work Compression ratio Quantized DCT coefficients (100%) -415 -30 -61 27 56 -20 -2 4 -21 -60 10 13 -7 -8 -46 7 77 -24 -28 9 5 -5 -48 12 34 -14 -10 6 1 -6 -13 -3 -1 3 2 -4 -26 -3 -6 2 -1 -2 -4 1 5 JPEG在不同的壓縮比有不同的AC特徵 作者事先使用JPEG圖像資料庫來訓練各窗口的AC特徵分類器 使分類器可以依據AC特徵來分類壓縮率 Quantized DCT coefficients (100%) Quantized DCT coefficients (50%)
Related work Support Vector Machine
Proposed Scheme Window sampling JPEG Image (50%) 64px 62 65 57 60 72 63 82 55 56 108 87 71 58 50 111 148 114 67 66 120 155 68 70 101 122 88 78 64 80 81 75 85 54 83 94 JPEG Image (50%) 作者知道原始圖像的JPEG壓縮率 𝑄 1 分析窗口有不同的大小 窗口將取樣圖像中的內容 並做JPEG encoding Window sampling
Proposed Scheme DCT matrix Detection scores (part) SVM classifier 1 2 3 4 5 6 7 8 64px -26 -3 -6 2 -1 -2 -4 1 5 JPEG Image (50%) 透過分類器 我們得知 𝑄 2 並與 𝑄 1 比對 計算檢測分數 如果完全相同 分數為0, 並將分數記錄在候選圖上涵蓋的認證單位 DCT matrix Detection scores (part) SVM classifier 𝑄 2 ≈ 𝑄 1
Proposed Scheme DCT matrix Detection scores (part) SVM classifier 1 2 3 4 5 6 7 8 9 0.5 64px -24 -3 -7 2 1 -4 -1 -2 6 3 JPEG Image (50%) 窗口移動一個認證單位, 當 𝑄 2 不等於 𝑄 1 時 檢測分數紀錄為1,因為窗口是overlapping 所以分數會平均 DCT matrix Detection scores (part) SVM classifier 𝑄 2 ≠ 𝑄 1
Proposed Scheme DCT matrix Detection scores (part) SVM classifier 2 3 4 5 6 7 8 9 10 0.5 0.3 64px -25 -2 -7 3 2 -1 -4 1 4 -3 JPEG Image (50%) 因為分數會被多重平均, 因此在大尺度的窗口會有明顯的漸層效果, 定位不清楚 DCT matrix Detection scores (part) SVM classifier 𝑄 2 ≈ 𝑄 1
Proposed Scheme (a) (b) (c) (d) (e) (a) Tampered image (b) Factual tampering locations (c-e) Candidate maps {16 px, 64 px, 128 px} 在不同尺度的滑動窗口, 會產生不同效果的檢測圖, 在本文中我們稱之候選圖 小尺度的窗口可以有效的定位篡改區域, 但是容易造成檢測錯誤 大尺度的窗口具有較高的檢測正確率, 但篡改區域的輪廓較模糊
Proposed Scheme Candidate maps (diagram) 0.8 0.9 0.4 0.5 0.9 0.3 1 0.6 0.9 0.4 0.5 0.9 0.3 1 0.6 0.9 1 候選圖中的一個格子為一個認證單位, 一個認證單位一個檢測分數
Proposed Scheme Idealized fusion of 3 candidate maps, corresponding to a small, medium, and large analysis windows, into a single accurate tampering map. 作者有提到融合策略不只一種, 最簡單的為平均融合 計算候選圖的平均檢測分數, 自訂一個門檻值, 當平均檢測分數大於門檻值, 就在最終的檢測圖上設置1, 反之為0 最終檢測圖是一張二位元圖像, 白點代表被篡改區域 1 𝑆 𝑠=1 𝑆 𝑐 (𝑠) The averaging strategy (AV fusion)
Proposed Scheme Threshold Fusion detection map (AV) 0.2 0.4 0.9 0.3 0.8 0.7 0.1 1 Threshold 設定門檻值得到最後的檢測圖 Fusion detection map (AV) Final detection map (binary)
Experimental Results 各種窗口大小與融合後的檢測圖的正確率
Experimental Results 由左至右分別為被篡改圖, 實際上的篡改區域, 3張候選圖, 5種融合策略的檢測圖 Candidate map averaging (AV fusion), energy minimization (EM fusion), bottom-up improvements (BU fusion), supervised learning (SL fusion), clustering analysis (CA fusion)
Experimental Results 候選圖與融合後的檢測圖
Experimental Results 𝐹 1 為平均正確率, 比較單一窗口與多尺度融合的正確率
Comments Advantages - Tampering localization whitout watermarking - High detection accuracy Disadvantages - Must be based on JPEG compression