數位相機及智慧型手機已成為生活中不可缺少的部分,因此我們隨處可見數位影像,但由於數位化資料擁有容易修改的特性,且影像處理軟體功能日益強大的狀況下,往往可以記錄真實事物的特性面臨極大的考驗。 本論文以融合多個不同用途的深度神經網路模型搭配Brute-Force matching、FLANN-based matching等特徵比對演算法之使用來探討多重深度神經網路結合特徵比對的複製移動竄改偵測技術,第一步使用竄改區域偵測神經網路偵測經複製移動竄改的區域;第二步將候選竄改區域與候選竄改區域外的部份做特徵匹配,判定候選竄改區域的可用性;第三步對全圖使用Mask R-CNN偵測與竄改區域重疊的完整竄改物件全貌並得知物件類別,同時找到圖中特徵匹配且相同類別的物件,視為來源物件;在第三步找不到的情況下,使用深度語義協同分割網路,由竄改區域逆向找出來源區域。經實驗結果本論文研究利可以有效降低複製移動竄改偵測的時間。
This paper presents an efficient strategy of applying the deep learning structure to solve the copy-move forgery detection problem. The copy-move forgery detection problem is to detect copy-move image forgery regions in an image. The proposed scheme first detects copy-move replaced regions using SRM filter. The BusterNet method detects the image foreground manipulation for searching the copy-move forgery regions. The presented scheme includes two major topics as foreground copy-move forgery regions detection and background copy-move forgery regions detection. The proposed scheme is a combination of these two schemes. In the proposed scheme, results of the Mask R-CNN and Semantic Co-Segmentation Network can be further utilized for detecting copy-move results.