Muscrat Ink Sep

--- license: mit pipeline_tag: image-segmentation tags: - version2 --- Manuscript Ink Layer Separation — U-Net (ResNet-34) A binary segmentation model for separating ink from parchment in medieval manuscript fragment images. Given a high-resolution manuscript scan, the model produces an ink mask, an isolated ink layer (white background), and a parchment layer (ink removed). Architecture: U-Net with ResNet-34 encoder, trained from scratch (no pretrained weights), 2-class output (background / ink). Inference: Sliding window over full-resolution images (default 512 px window, 256 px stride) with averaged overlapping predictions and morphological post-processing to reduce halo artifacts. Trained on: an annotated collection of medieval liturgical musical manuscripts, vernacular Icelandic saga manuscripts, and fragments featuring various parchment qualities. Validation IoU: .767 Intended use: Research tool for manuscript studies and digital humanities. Not intended for printed documents or modern text. Status: currently one version for ink separation, and so should only be considered as a piece of a work in progress. Running guideline: if you're getting some odd behavior, widen the window size to 900 and decrease confidence.

Source context: 0 downloads · 0 likes · Pipeline image-segmentation · Repo grackle-in-an-HEB-parking-lot/muscrat-ink-sep