zlinker / dehazenet Goto Github PK

View Code? Open in Web Editor NEW

112.0 5.0 51.0 635 KB

an end to end dehaze tools

Python 100.00%

dehazenet's Introduction

Reimplement

DehazeNet: An End-to-End System for Single Image Haze Removal

Bolun Cai, Xiangmin Xu, Kui Jia, Chunmei Qing, Dacheng Tao

Requirement

caffe

opencv2

Usage:

simply type

python DehazeNet.py image_path

Demo:

Site:

@article{cai2016dehazenet,
 title={Dehazenet: An end-to-end system for single image haze removal},
 author={Cai, Bolun and Xu, Xiangmin and Jia, Kui and Qing, Chunmei and Tao, Dacheng},
 journal={IEEE Transactions on Image Processing},
 volume={25},
 number={11},
 pages={5187--5198},
 year={2016},
 publisher={IEEE}
}

dehazenet's People

Contributors

Stargazers

Watchers

dehazenet's Issues

Architecture

Hi,

I saw that the architecture seems to be different compared to the original DehazeNet. There seems to be no BRelu units, the stadard Relu units are used instead. Also the architecture alone seems to be a little bit different. For example in the feature extraction phase there is supossed to be maxout unit only. You performed maxout using max pooling after Relu activation etc. Also the filter numbers are a bit different. What are the reasons ?

Where did you use the 4 features ?

Hi zlinker.
I'm curious about the dehazenet but your DehazeNet.py file seems to not have the max satulation, max constract and hue disparity which were described the original paper. How did you use the max out function ?

how to get haze image?

hi, i'm currently reimplementing the experiment, can you upload the code to generate haze patches? I followed the steps by Bolun Cai. But the haze images are so blurring.

When i test a high resolution picture in GTX1080,the program met a problem!

I0307 17:14:24.393541 4778 net.cpp:200] conv2/5x5 does not need backward computation.
I0307 17:14:24.393559 4778 net.cpp:200] conv2/3x3 does not need backward computation.
I0307 17:14:24.393576 4778 net.cpp:200] conv2/1x1 does not need backward computation.
I0307 17:14:24.393594 4778 net.cpp:200] reshape2_reshape2_0_split does not need backward computation.
I0307 17:14:24.393611 4778 net.cpp:200] reshape2 does not need backward computation.
I0307 17:14:24.393627 4778 net.cpp:200] pool1 does not need backward computation.
I0307 17:14:24.393640 4778 net.cpp:200] reshape1 does not need backward computation.
I0307 17:14:24.393649 4778 net.cpp:200] relu1 does not need backward computation.
I0307 17:14:24.393661 4778 net.cpp:200] conv1 does not need backward computation.
I0307 17:14:24.393671 4778 net.cpp:200] input does not need backward computation.
I0307 17:14:24.393680 4778 net.cpp:242] This network produces output ip1-conv
I0307 17:14:24.393713 4778 net.cpp:255] Network initialization done.
I0307 17:14:24.393977 4778 net.cpp:744] Ignoring source layer data
I0307 17:14:24.394007 4778 net.cpp:744] Ignoring source layer ip1
I0307 17:14:24.394019 4778 net.cpp:744] Ignoring source layer loss
/home/fql/anaconda3/lib/python3.6/site-packages/skimage/io/_io.py:49: UserWarning: as_grey has been deprecated in favor of as_gray
warn('as_grey has been deprecated in favor of as_gray')
F0307 17:14:31.605160 4778 syncedmem.hpp:33] Check failed: *ptr host allocation of size 6144000000 failed
*** Check failure stack trace: ***

the picture's pixel is 1200X800 or 4752X3168

Can you share the caffe code and dataset for training

Hello zlinker,

Thanks for sharing the DehazeNet for inference. I was wondering if you can share the training caffe code and dataset with me?

Yunyang