转载于:
https://blog.csdn.net/phoenixtree7/article/details/60769717
本文提到的这种图像曝光增强算法,是文章【1】提及的。具体原理比较简单,文章也是作为去雾算法的一个 postprocessing option 用很小的篇幅介绍。
【1】Tang K, Yang J, Wang J. Investigating haze-relevant features in a learning framework for image dehazing[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2014: 2995-3000. [ download:http://www.juew.org/publication/dehazing_cvpr14.pdf ]
大家都知道,用DCP及其改进等去雾算法都会有曝光不足问题。开始,我也只是用简单的gamma变换等等简单的方法进行曝光改善,但并不理想。这里介绍的方法,感觉要合理很多,大家可以一试。
这里提供自己写的matlab代码。当然,也有python代码,大家可以下载。 [ download: https://github.com/caibolun?page=1&tab=stars, 在该网页找到 haze_2014 ]
MATLAB CODE
function [ exposureOutput,exposureScaling ] = imexposure(J,I,lambda)
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% This is a function to enhance the input image's exposure.
% Image Exposure Algorithm using GIF
%
% Input :
% J The poor exposure image, color image
% I The guidance of the Guided Image Filter, color image
% lambda The regular parameter
% Output:
% exposureOutput The exposure improved image
% exposureScaling The exposure scaling for each pixels
%
% This algorithm is derived by the artical of :
% Investigating Haze-relevant Features in A Learning Framework for Image Dehazing
% Ketan Tang et.al. 2014
%
% Edit by Dong Zhao, 2017
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
hsv_I = rgb2hsv(I);
hsv_J = rgb2hsv(J);
I_gray = rgn2gray(I);
luminance_I = hsv_I(:,:,3);
luminance_J = hsv_J(:,:,3);
%lambda = 0.1;
eps = 10^(-6);
r = 60;
IxI = lambda*luminance_I.*luminance_I;
JxJ = luminance_J.*luminance_J;
IxJ = luminance_J.*luminance_I;
S = (IxJ+IxI)./(JxJ+IxI);
S_GIF = guidedfilter(I_gray,S,r,eps);
%figure,imagesc(S_GIF); colormap jet;
hsv_J(:,:,3) = S_GIF.*hsv_J(:,:,3); %+hsv_J(:,:,3);
exposureScaling = S_GIF;
%figure,imshow(hsv_J(:,:,3) );colormap jet;
exposureOutput = hsv2rgb(hsv_J);
%figure,imshow(exposureOutput)
关于导向滤波器的代码 guidedfilter,网上一搜就能找到。
效果展示:
其中,lambda = 0.1
(a) 去雾图像,曝光不足
(b) 导向滤波器的 guidance,我们选择雾图像
(c) 计算得到的 scaling
(d) 曝光增强图像
Lambda
lambda的调整:
lambda>0, lambda 越小,曝光程度越强。
(e) lambda = 0.3 时的 scaling,发现增强程度不如 lambda = 0.1,即图 (c)。
(f) lambda = 0.3 的输出结果
当然,lambda 不能小于0!