使用多區域相互關係技術應用於修補影像之偽造偵測快速演算法

張志詮
In the past decades, the issue of digital image forgery becomes more and more important owing to the popularization digital images and the power of image processing software. Digital image forensics concerns about the identification of image authenticity. Recently, many researchers have proposed different methods which could detect forged region effectively.

Most of the detection forged images are generated by image processing software. A few researchers pay attention to exemplar-based inpainting technique which can also generate sophisticated forged image.In this thesis, we focus on the study of detecting exemplar-based inpainting forged image. We propose a fast and effective forgery detection algorithm for exemplar-based inpainting forged image. Our algorithm can be divided into similarity detection and forged region detection. In the similarity detection, we compute similarity among regions in the image to find high similarity regions which are called suspicious regions. According to the high similarity among regions, we obtain high recognition rate for forged image with uniform background by using corresponding vector. In forged region detection, we can correctly identify the forged region by using multi-region correlation (MRC) technique. Besides, in order to improve the performance of similarity detection, we propose a fast searching algorithm based on weight transformation. The experimental results show that the proposed algorithm can precisely locate the forged region of exemplar-based image inpainting in high speed.

過去十年間，由於數位影像的普及和強大的影像處理軟體使得數位影像偽造問題變得越來越重要，數位影像鑑識就是用來識別影像真實性，近來，許多研究者提出不同的方法能夠有效偵測出偽造區域，大部分偵測的影像以影像處理軟體所產生偽造影像為主，較少研究者注意到exemplar-based inpainting technique也能夠產生幾可亂真的偽造影像。在本論文中，我們專注在偵測inpainting偽造影像的研究。我們提出一個快速且有效的偽造偵測演算法針對inpainting偽造影像。我們的演算法分成相似度偵測和偽造區域偵測，在相似度偵測中，我們計算影像中區域之間的相似度尋找高相似度之區域，這些區域被稱為可疑區域 (Region of Suspicious)，另外，依據區域之間的高相似度，我們使用corresponding vector提升辨識具有uniform background偽造影像的準確率。在偽造區域偵測中，我們使用一個多區域相互關係(Multi-Region Correlation) 技術能夠正確地識別偽造區域。此外為了改善similarity detection的效能，我們提出一個快速搜尋演算法基於權重轉換。實驗結果顯示我們的演算法能夠快速且準確找出範例式影像修補的偽造區域。