Awesome

LineRelifer

A data-driven line refiner for line-drawing (NPR and hand-drawn) stylization using Keras, which could a simple example to start your illustration related deep learning tour.

Overview

In non-photorealistic rendering, the stylization of line-drawing is an important task. Present commercial software provides such style control through parametric or interactive manipulations. For example, in FinalToon and Pencil+, users are allowed to control the width and opacity of the strokes with maps or values. Futhermore, other effects like noise can be applied to strokes. These systems often rely on the 3D object`s properties, such as UV coordination, normal, curvature. But in real application, painters have more consideration on the expression of drawing, and these styles would be very difficult to describe parametrically with geometric characteristics. To solve these problems, learning based systems have been proposed Learning Style Translation for the Lines of a Drawing [1]. However these approaches use vector graphics as input, and cannot be used for raster images directly. Using vector graphics limits the amount of training data and the diversity of styles. Recent advance in deep learning provides a new method for processing raster image directly and has achieved good results Mastering Sketching: Adversarial Augmentation for Structured Prediction [2].

This work propose a deep-learning based approach to line-drawing (NPR and hand-drawn) stylization problem. We received inspiration from [2] and Learning from Simulated and Unsupervised Images through Adversarial Training [3]. Our approach build upon the framework like [2] and [3], which uses an adversarial network similar to Generative Adversarial Networks (GANs). To achieve production oriented target, it has smaller network than [2] and can run on CPU with acceptable speed. This network is able to generate highly realistic styled line-drawings with HD size when memory is enough. It can also fill gaps between strokes, denoise and correct shapes in line-drawings.

!! This repository is a simple version of LineRelifer in production.

Dependencies

• Keras2 (Tensorflow backend)
• OpenCV3

Usage

1. Set up directories.

2. Download the model from release and put it in the same folder with code, then rename to model.h5.

3. Run predict.py for prediction. Run model.py for train.

Data Preparation

1. Collect line-drawings with LineDistiller.

2. Thin(normalize) the line-drawings with LineNormalizer or tranditional thinning method.

3. Manually processe line-drawings and thinning results(threshold etc.), then crop them into pieces.

4. Put line-drawings into data/line and put thinning results into data/norm.

Models

Models are licensed under a CC-BY-NC-SA 4.0 international license.

• LineRelifer Release Page

Post-processing

To achieve production level, the result should be post-processed. You can use any image editing software. Here shows the method in Photoshop with OLM smoother.

• Convert color mode to RGB

• Use curve to achieve desire style

  <img src="figs/curve.jpg"/>

• Use OLM smoother to do AA.

Notes

• Input should have white background and black lines.
• Prefered stroke width is 1 to 3 px (AA or NoAA).
• Memory cost for HD image is heavy, make sure computer has enough memory (>8 GB).

Examples

• Click to enlarge and view details.
• Result on hand-drawn line-drawing without width change on strokes. Pictures from web. <img src="figs/l1.jpg"/>

• Result on 3d NPR line-drawing. <img src="figs/3d_1.jpg"/>

• Result on filling gap. <img src="figs/one_pixel_gap.jpg"/>

Weibo [Chinese]

From Project HAT by Hepesu With :heart: