Line Remover NN

Note

This is the new v2 version of the project (complete rewrite), to see the old version go to the legacy branch.

Introduction

This repos uses PyTorch to remove ruled lines from an image while reconstructing overlapping characters with lines. The goal of this model is to make easier the word recognition from OCR.

Installation

Prerequisites

• Pixi (cpp dependencies manager)
• UV (python dependencies manager)
• Nvidia GPU with CUDA support (I cant test AMD ROCm)

Setup Environment

# Install CMake, opencv, cairo, compile cpp etc...
pixi install

# Install pre-commit hooks
pixi run hooks

# Generate compile_commands.json
pixi run clangd-setup

Quickstart

Install Datasets

These commands automatically download and extract popular datasets.

IAM

pixi run lineremovernn download-dataset -d iam

Mathwriting

pixi run lineremovernn download-dataset -d mathwriting

AI2D

pixi run lineremovernn download-dataset -d ai2d

Generate synthetic pages

Note

Using IAM's dataset RAM Preloading + Indexing if you can can speed up generation by ~300% (~100ms per page to ~28ms).

pixi run lineremovernn generate-pages [OPTIONS]

Option	Type	Help
-n, --n	INTEGER	Number of images to generate
--datasets	STRING	Space-separated datasets, proportions and flags (p for RAM preloading, i for indexing). E.g. iam:1:ip mathwriting:0.3
-d, --docs	FLAG	Make document like layouts instead of one big chunk of paragraph.
-a, --a	FLAG	Use arcs instead of straight lines
-il, --imperfect-lines	FLAG	Add noise to the lines.
-mw, --max-warp	FLOAT	Maximum perspective warp factor for word crops (0.0 to disable, recommended: 0.15)
-m, --save-metadata	FLAG	Export ground-truth word layout coordinates as XML files.
-w, --workers	INT	CPU threads to use (default: all cores)
-db, --debug	FLAG	Debug how long each process of page generating is.

Train Model

pixi run lineremovernn train [OPTIONS]

Option	Type	Help
-e, --epoch	INT	Number of epochs to train the model for.
-b, --batch-size	INT	Batch size.
-l, --load	FLAG	Continue training of latest model.
-ex, --extended	FLAG	Use extended dataset augmentation & transforms.

Utils commands:

List available models

pixi run lineremovernn ls-models

Model layers

pixi run lineremovernn model-info

Test model

pixi run lineremovernn test

Option	Type	Help
-n, --n	INT	Number of images to test.
-b, --batch-size	INT	Batch size.
-l, --loss	FLAG	Show loss for each image.

Preview dataset

pixi run lineremovernn preview-dataset

Option	Type	Help
-n, --n	INT	Number of images to test.
-d, --dataset	STR	Dataset to preview available: pages.
-t, --transform	FLAG	Add some random transforms.

GUI Infer

No python lib for the moment. You can use the gui:

pixi run lineremovernn gui-infer

Development

To force rebuild of CPP bindings (src/lineremovernn_ext), use:

pixi run build

License

Copyright (C) 2026 PastaLaPate.

This project uses: Barkeep - Licensed under the Apache License, Version 2.0 by Ozan İrsoy. Pugixml - Licensed under the MIT License by Arseny Kapoulkine.

This project is licensed under the GNU Affero General Public License v3 (AGPLv3) - see the LICENSE file for details.