Fuzzy Image Enhancement

Implementation of the Fuzzy Inference System for image enhancement as described in the paper:

YoungSik Choi & Raghu Krishnapuram. A Robust Approach to Image Enhancement Based on Fuzzy Logic. IEEE Transactions on Image Processing, ( Volume: 6 , Issue: 6 , Jun 1997 )

The repository contains the implementation for the filters A, B, C, R1, R2, R3 and R3-Crisp, as proposed in the paper. Moreover we also add an R4 and R4-Crisp filter which is a small modification of R3. In R4 we change the consequent of Rule 1 to filter B and consequent of Rule 2 to filter A (ie. small -> A and medium -> B). Also to compare against other out-of-box new methods, this implementation has facility for Median, Sharpen, Gauss, TV Chambolle, TV Bregman and Biltaeral filter.

Setup

Prerequisites

• Python 3.7
• pip (for setting up dependencies)

Setting up dependencies

Use

$ pip install -r requirements.txt

Running Experiments

First create the recommended directory structure.

mkdir images/noisy images/enhanced images/tc

Within the images directory, the directory original stores original greyscale images, noisy directory stores the noise added images, and enhanced stores the enhanced images obatined by applying filters.

To add noise to images (note that the image must be stored in images/original directory. Only pass the imagename as IMGNAME not the path):

python noise.py --image IMGNAME

Change the parameters of noise from within noise.py if needed.

To run the filters:

python main.py --image NOISYIMGPATH --method METHOD [--original ORIGINALIMGPATH]

NOISYIMGPATH is relative to images/ typically something like noisy/Cameraman_5_5_100.png, similarly for ORIGINALIMGPATH (eg. original/Cameraman.png). When original is specified the RMSE is reported. Possible options for METHOD are 'A', 'B', 'C', 'Med', 'R1', 'R2', 'R3', 'R3Crisp', 'R4', 'R4Crisp', 'All' and 'Plot'. 'All' applies all the filters. 'Plot' plots the weights of filter A, B and C and saves to images/tc. 'Med' applies the median filter.

To run the other out-of-box filters:

python compare.py --image NOISYIMGPATH --method METHOD [--original ORIGINALIMGPATH]

Possible options for METHOD are 'Sharpen', 'Gauss', 'TVC' (TV Chambolle), 'TVB' (TV Bregman) and 'Bil' (Biltaeral). Other arguments are same as for main.py.

License

This code is provided using the MIT License.

---

This project was a part of the course MA6040: Fuzzy Logic Connectives: Theory and Applications, offered in Spring 2019 at IIT Hyderabad.

Team members: Vaibhav Sinha and Prateek Kumar.