Factorization-based segmentation Python implementation

This is a Python implementation of the factorization-based segmentation algorithm, which fast segments textured images. The algorithm is described in

J. Yuan, D. L. Wang, and A. M. Cheriyadat. Factorization-based texture segmentation. IEEE Transactions on Image Processing, 2015.

Here is a brief introduction of the algorithm. Here is an explanation of computing local histograms based on integral histograms.

There is also a MATLAB implementation. The results from two implementations are similar. Local spectral histogram computation is coded using pure matrix operations, and thus achieves a speed comparable to the mex c code in MATLAB implementation.

OBS : This repository is a fork for the original code

TODO

In this actual stage, this code is only reading gray-scale images. Need to move later to read BGR/RGB images

Also, it's only working for .png and .jpg

Prerequisites

Python 3.9 or above

Numpy>=1.23.3

Scipy>=1.9.1

Scikit-image>=0.19.3

opencv-pytho>=4.6.0.66

Usage

Unsupervised texture segmentation mode

This verison implements the complete algorithm, which segments an image in a fully automatic fashion.

To try the code, run

python3 FctSeg.py -f "img_path" -ws int_value -segn int_value -omega float_value -nonneg_constraint bool_value -save_dir "dir_to_save" -save_file_name "file_name_to_save"

where :

-f is the image path

-ws is the window size value

-segn is the number of segment. if set to 0, the number will be automatically estimated

-omega is the error threshod for estimating segment number. need to adjust for different filter bank.

-nonneg_constraint is a flag to apply the negative matrix factorization

-save_dir is the path with the folder to save the file

-save_file_name is the file name with the extension to save the final result

Example on how to run this code with some parameters

python3 FctSeg.py -f "Images/img_test.png" -ws 25 -segn 0 -omega 0.045 -nonneg_constraint True -save_dir "ResultImages/" -save_file_name "final_result.png"

To try the version with given seeds, run

python FctSeg_seed.py

Each seed is a pixel location inside one type of texture. Note that this version represents the basic form of the algorithm and does not include nonnegativity constraint.

Three test images are provided.