International Journal of Signal Processing, Image Processing and Pattern Recognition Vol. 10, No. 6 (2017), pp. 75-86 http//dx.doi.org/10.14257/ijsip.2017.10.6.07 ISSN: 2005-4254 IJSIP Copyright © 2017 SERSC Method for Detecting Small Area Forgery by Comparing Two Document Images Daejune Ko 1 and * Eui Chul Lee 2 1 Department of Computer Science, Graduate School, Sangmyung University 20, Hongjimun 2-gil, Jongno-gu, Seoul 03016, KOREA 2 Department of Computer Science, Sangmyung University 20, Hongjimun 2-gil, Jongno-gu, Seoul 03016, KOREA 1 kodaejune@gmail.com, 2 eclee@smu.ac.kr * Corresponding author Abstract In this paper, a method for detecting tiny area with forgery between two images is proposed. The proposed method includes two procedures. Firstly, geometric features such as translation, rotate and scaling between two images are extracted by using SIFT algorithm. The false geometric correspondences are filtered out by cross validation. Then two images are aligned by affine transform using the best three corresponding pairs. Next, aligned images are divided locally and uniformly, the size of each region 20×20 pixels and solving slightly misaligned between two images is used by moving window method for determining the best aligned position. And local similarities between two images are measured by calculating correlation coefficient. To verify proposed method, ten receipts were used. Each receipt was saved as two types such as original and forgery ones, respectively. After training, we defined that the correlation coefficient as an optimal threshold was 0.9 for classifying non-forgery and forgery areas. At result, we acquired 7.391% equal error rate. Keywords: forgery detection, image comparison, affine transform, geometric alignment, correlation coefficient 1. Introduction By the rapid development of IT technology, digital data could be modified by anyone easily. Whereby, forgery of digital data has spread to various fields. In the early days of information technology, forgery of digital data had been used to hide individual identity of news image or augment reality in the field of computer graphics. However, in recent years, it has been used in criminal fields such as counterfeit banknotes or forgery documents. In pervious works, several researches for forgery detection have been performed by using image or watermarking. Firstly, image based methods focused on copy-move forgery [1, 2, 3]. The main purpose of copy-move forgery is to hide or highlight specific area in the image. In these methods, forgeries were detected by analyzing local block features or keypoints. In local block-based method, images are subdivided for extracting features and every region matched another region [2]. In keypoint-based methods, features could be extracted and matched without subdivision [3]. Both, block- and keypoint-based methods include filtering feature process. However, these methods have strong dependencies with the quality of block feature of keypoints. Secondly, invisible watermarking based methods have been widely used [4, 5]. Yoo et al. proposed a hybrid watermarking method [4]. They embed the PN-Sequence using wavelet transform and verified algorithm using various forgeries. Kim and Choi proposed quantization watermarking method. They used discrete cosine transform and adaptive quantization