Detection Methods for Seam Carving and Seam Insertion in Digital Images

• Main Authors: Yi Jing Wu, 吳怡靜
• Other Authors: J. D. Wei
• Format: Others
• Published: 2012
• Online Access: http://ndltd.ncl.edu.tw/handle/03574884364901254583

碩士 === 長庚大學 === 資訊工程學系 === 100 === Seam-based image processing method, including seam carving and seam insertion, has been a famous content-aware image retargeting technique. This method connects pixels with the lowest energy into seams. Repeatedly removing or inserting seams, we can resize an image and retain the most important content of the image in the meanwhile. Seam carving technique can even apply for image tampering -- one can delete particular objects on the image only if he or she assigns the corresponding areas with the lowest energy. Images modified in this way are difficult to detect by both eyesight and current existing detecting algorithms. In this thesis, we propose a novel method, referred to as “patch analysis”, for detecting seam carved and seam inserted images. This method first divide the images into 2x2 sized units, called mini-squares. For each mini-square, there are nine kinds of virtual recovery patterns mapping it to nine patch types. We suggest three kinds of referee patterns and two assessment formula to determine the optimal patch type for every mini-square. Following the concept of Markov process, we calculate the transition probability matrices for the optimal patch types in three directions, i.e., sub-diagonal, vertical, and diagonal. Finally, we use the transition probability as the input feature for a support vector machine (SVM). The well-trained SVM build up a decision model that can identify whether the test image is seam modified or not. Our proposed method are not only very meaningful and readable in operation but also outperforming other previous detecting methods in experimental results.