RGB document branch
Document layout, text edges and local visual structure enter the authors’ ASCFormer reference pipeline.
TAFE-ID / 01 / ARCHITECTURE
Two signals.
One pixel map.
The authors’ ASCFormer checkpoint provides the verified reference path. Our separate TAFE-Net experiments combine RGB crops and JPEG frequency inputs. These are distinct model tracks.
MODEL PATH
From a document to a localized region.
JPEG evidence
Real RTM JPEG coefficients and quantization tables were audited for the custom TAFE frequency branch.
Localization output
The reference segmentation output is aligned to the document for inspection. Our custom frequency experiments are reported separately.
PAPER CONTEXT
What came from the source paper?
Luo and colleagues introduced the RTM dataset and ASCFormer in “Toward real text manipulation detection: New dataset and new solution” (Pattern Recognition, 2024). We ran their released source and checkpoint for the reference result. Our RGB/JPEG frequency branch is a separate experiment, not an equivalent implementation or a successful reproduction of their benchmark.
Read the RTM / ASCFormer paper ↗