Matlab implementation of Efros-Leung texture synthesis by non-parametric sampling.
Switch branches/tags
Nothing to show
Clone or download
Latest commit cb982dd Jan 5, 2016

README.md

Texturesynth

This is a Matlab implementation of the Efros-Leung algorithm for texture synthesis by non-parametric sampling. For details about the algorithm, see the original paper.

Usage

The following is an example of how the synthesize_texture function is used (from synthesize_texture_test.m):

input_image = im2double(imread('161.png'));
output_image = synthesize_texture(input_image, 256, 256, 9); 
imwrite(output_image, 'result.png');

The parameters of synthesize_texture are as follows:

function output_image = synthesize_texture(input_image, output_rows, output_cols, window_size)

Inputs

  • input_image: the image containing a sample of the texture to synthesize
  • output_rows: the desired number of rows in the synthesized image
  • output_cols: the desired number of columns in the synthesized image
  • window_size: the side length of the window used to scan across the input image to look for patches that most closely match the region that is being synthesized

Outputs

  • output_image: an (output_rows x output_cols) image containing the synthesized texture

Some Examples

Custom Textures

Input Output Window Size
Ground Input Ground Output 15
Cloth Input Cloth Ouptut 15
Metal Input Metal Ouptut 15
Moss Input Moss Ouptut 9

Classic Textures

Input Output Window Size
Wood Input Wood Output 15
Blobs Input Blobs Ouptut 15
Knots Input Knots Output 21
Text Input Text Output 21

Format of Examples

The examples folder contains two subfolders, ClassicTextures and MyTextures. ClassicTextures contains examples of this implementation used on some of the texture samples that were used in the original Efros-Leung paper. MyTextures contains examples that use my own texture samples.

For each example, the input file is named after the material pictured, and the corresponding output file contains the name of the input file, plus "result" and "window," followed by the window size used for that example. For example, input file "bark1.png" has output file "bark1_result_window_9.png" which shows that the window size is 9.

Implementation Notes

In order to handle both grayscale and colour images with the same code, this implementation stacks channels vertically, so that calculations are always performed using only column vectors.

In this script, the candidate patches are converted to the stacked representation at the start, and are kept that way for all the distance computations. When looking at an individual pixel to synthesize, the neighbourhood about that pixel is reshaped into a single column with channel stacking, and is replicated as many times as there are candidate patches. It is important to note that even when there are multiple colour channels, the mask showing which pixels have been filled is still two-dimensional. The mask applies across all colour channels, so in order to vectorize the code, the mask is replicated vertically as many times as there are channels:

mask_vec = repmat(mask_vec, size(candidates, 3), 1);