Tampered Text Detection for ID Photos

In recent years, tampered text detection in document images has attracted increasing attention due to its significance in information security. However, detecting tampered text in captured ID photos still faces numerous challenges. In this paper, we propose a novel ID photo tampering detector (PTD), which consists of a frequency-aware head, a multi-resolution feature fusion network and an iterative decoder. The frequency-aware head offsets the limitations of indistinct visual features, while the multi-resolution feature fusion backbone conducts hierarchical aggregation of multi-scale features, subsequently transmitting the global representation into the iterative decoder. Finally, the iterative decoder progressively decouples a binary mask of tampered regions in ID photos and regresses a tampering probability map, significantly improving prediction accuracy. Moreover, we leverage the DocTamper dataset to supplement the training set, and then evaluate our algorithm in novel ID photo samples, which contain 1725 tampered images. Experimental results comprehensively demonstrate the superior performance and generalization capability of our approach.