Abstract
Abstract
This thesis aims to distinguish between fraudulent documents and original ones by analyzing the inherent textural structure present in the papers they are printed on. It is shown that the likelihood of two distinct sections from a paper sharing the same underlying textural structure is extremely low. The primary objective is to determine whether an object exists in the database or does not (i.e., if it is original or fraudulent), which can be framed as a Hypothesis Testing problem. To address this problem, a Siamese Network is utilized to extract discriminative features. By introducing a new coefficient to the loss function of this base network, the identification of mismatched pairs is significantly improved. Subsequently, the learned embeddings from the base network are employed for the Hypothesis Testing problem. The problem can be viewed as comparing image of a new incoming object with those already encountered, using a suggested Meta Learning mechanism in the embedding space. Additionally, an end-to-end network is constructed to facilitate the objectives of both the Siamese Network and the Meta Learner. To demonstrate the effectiveness of proposed method with experiments, a dataset including paper sections is collected and subjected to a data augmentation schema. Additionally, experiments are conducted on a publicly available fabrics dataset. Systematic experiments reveal that the proposed method outperforms the baselines in terms of both accuracy and Type-II error (percentage of frauds predicted falsely as originals). The novel approach showcases improved performance, effectively differentiating between genuine and fraudulent documents based on the textural structure analysis.