cteckwee / defocus_segmentation

LBP-based segmentation of defocus blur

LBP-Based Segmentation of Defocus Blur

Update2:

Add a python implementation to compute the sharpness metric.

Update1:

The blur maps for the 1000 images in the blur segmentation dataset produced by our algorithm can be found here for easy comparison.

This repo provides the code to reproduce our defocus segmentaion results in our paper.

How to use

Prerequistites

• Matlab 2016a

If you want to try out the python script, you need

• Python 3
• Opencv 3 (pip install opencv-python)

Getting Started

• Clone this repo:

git clone [email protected]:xinario/defocus_segmentation.git

• In Matlab, change your project directiory to <your download path>/defocus_segmentation then run demo.m

• Note that matlab implementation was used by default to compute the proposed sharpness metric. But you can also switch to .mex code to gain some speed boost. The sharpness metric implementation in .mex code was based on integral image and can run in real time on a single core cpu.

Compute the LBP-based sharpness measure using the mex version.

1. Install [mexopencv](https://github.com/kyamagu/mexopencv)

2. Copy lbpSharpness.cpp to <your mexopencv folder>/src/+cv 

3. Run mexopencv.make() in Matlab command line to compile the provided function.

4. Comment out line 14 and uncomment line 18, 19 in localSharpScoreLBP.m, then you are good to go.

All the results reported in the paper were produced by the mex verison of LBP-based sharpness.

• To use python script

python lbpSharpness.py --input ./images/out_of_focus0080.JPG

Citations

If you find it useful and are using the code/model/dataset provided here in a publication, please cite our paper:

Yi, Xin, and Mark Eramian. "LBP-based segmentation of defocus blur." IEEE transactions on image processing 25.4 (2016): 1626-1638.

Acknowlegements

The alpha matting code comes from Levin, 2006:

A. Levin D. Lischinski and Y. Weiss. A Closed Form Solution to Natural Image Matting. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), June 2006, New York.

The multi-scale inference code was adopted form Jianping Shi, 2014:

Jianping Shi, Li Xu, Jiaya Jia. Discriminative Blur Detection Features. IEEE Conference on Computer Vision and Pattern Recognition, 2014.