Tampered Text Image Detection: Integrating Semantic Segmentation Using Deep Learning and Discrete Cosine Transform (DCT) in Web Applications

With the surge in fraud cases involving manipulated texts and documents, this study seeks to identify the best algorithm for detecting such manipulations specifically the two algorithms, U-Net and modified DTD, to see which is better at detecting these changes. In this case, to help compare which of the model algorithms are more effective in detecting whether or not the texts or documents presented are indeed tempered, two different cross-domains were implemented. In the first cross-domain, the DTD model achieved an Intersection-over-Union (IoU) score of 74.98% and an precision@k of $\mathbf{7 1. 7 7 \%}$, outperforming U-Net, which brought in lower results. These data suggest that the DTD model is a better choice than U-Net for identifying modifications to texts and documents. While in the second cross-domain, DTD had once again excelled with an IoU score of 46.54% and a precision@k of $\mathbf{5 5. 4 8 \%}$. These findings confirm that the DTD model has an overall better generalization and flexibility in detecting text and document manipulation. Therefore, contributing to analyze the comparative result to the field and building a pratical application.