Fast Weighted Median Filter

Weighted median is used in many of computer vision applications for its properties in sparsity representation. We re-implement an efficient algorithm to reduce computation complexity from O(r2) to O(r) where r is the kernel size.

The paper contributes a new joint-histogram representation, median tracking, and a new data structure that enables fast data access. The running time is largely shortened from several minutes to less than 1 second.

Prerequisites

C++11 or higher: The algorithm scheme is implemented in C++.
GCC compiler: For compiling C++ code.

Install gcc

OpenCV 3.0 or higher: For image processing tasks.

Install opencv

Testing

Please, follow the following instructions for testing the code:

Clone this repository

git clone https://github.com/gulshan-mittal/Fast-Weighted-Median-Filter.git

Go to the root directory of the repository.

cd Fast-Weighted-Median-Filter

Compile the files

make

run the output file using the following arguments

./output dataset/ComObj/lena.png results/ComObj/lena_out.png 3 0 1

Arguments:

relative image filepath
output image path
kernel radius
type of weights
- 0: unweighted
- 1: gaussian
- 2: Jacard
comparison with naive implementaion (This argument is not compulsory, you can skip it)
- 0: Yes
- 1: No

Example

Example-1: Removing salt-pepper noise from image

#Kernel size: 3
#Kernel type: Unweighted

./output ./dataset/ComObj/lena.png ./results/ComObj/lena_out.png 3 0

Input Image	Output image

Example-1: Removing Artifacts from image using Gaussian filter

#Kernel size: 10
#Kernel type: Gaussian

./output ./dataset/Artifacts/snail_artifact.jpg ./results/Artifacts/snail_artifact_out.jpg 10 1

Input Image	Output image

Analysis

Plot Details	Plot
Time Vs Kernel Size (Gaussian Kernel)
Time Vs Image Size (Gaussian Kernel)
Time Vs Kernel Size (Unweighted)
Time Vs Image Size (Unweighted)

talsperre / Fast-Weighted-Median-Filter