SALE-SealTamper v2.0: A dataset for document seal tampering localization

It supports research on seal tampering detection, weak-trace tampering localization, cross-synthesis generalization, and false positive analysis on genuine negative seal documents. The dataset provides document images, binary tampering masks, train/validation/test split files, four stepwise ablation branches, independent evaluation samples, metadata, source assets, and reproducibility scripts.The formal dataset contains 30,000 base groups.

Each base group includes four branches, namely S only, S+A, S+A+L, and Full SALE, resulting in 120,000 image-mask pairs. The Full SALE branch is recommended as the main training and testing subset, while the four stepwise branches can be used to analyze the contribution of semantic selection, area constraints, local blending, and explanation records in the dataset construction process. The independent evaluation sets include cross-synthesis testing, real-source positive testing, and genuine negative testing, which can be used to evaluate model generalization under non-identical synthesis settings and real-world scenarios.Image files are mainly stored in JPG format, and mask files are stored in PNG format.