【转】 ISP-镜头阴影校正(LSC)
转自:https://blog.csdn.net/xiaoyouck/article/details/77206505
介绍
镜头阴影校正(Lens Shading Correction)是为了解决由于lens的光学特性,由于镜头对于光学折射不均匀导致的镜头周围出现阴影的情况。
shading可以细分为luma shading和color shading:
luma shading:
由于Lens的光学特性,Sensor影像区的边缘区域接收的光强比中心小,所造成的中心和四角亮度不一致的现象。镜头本身就是一个凸透镜,由于凸透镜原理,中心的感光必然比周边多。如图所示:
chrom/color shading:
由于各种颜色的波长不同,经过了透镜的折射,折射的角度也不一样,因此会造成color shading的现象,这也是为什么太阳光经过三棱镜可以呈现彩虹的效果。如图所示:
此外,还有CRA的原因会导致shading现象的出现,这里不再赘述,这里推荐《What’s CRA》这篇文章,详细讲述了由于镜头的CRA带来的shading。
影响
luma shading:会造成图像边角偏暗,就是所谓的暗角。
color shading:中心和四周颜色不一致,体现出来一般为中心或者四周偏色。如图所示:
校正
lens shading的校正是分别对于bayer的四个通道进行校正,每个通道的校正过程是相对独立的过程。
考虑到芯片设计的成本,因此一般情况下不会存储整幅图像的lut,目前主流的都是存储128*128个点的增益,利用双线性插值的方法计算每个pixel的增益。
算法
由于条件限制,图像仅用于算法验证,不做图像质量评判标准
这里写了一个shading的算法,将图像分为16x16的方块,求取每个交点的增益值,对平面进行四次方拟合,分别计算了luma shading 和 chrom shading,先计算出来一个lut用于存储,校正的世行通过对这个lut进行双线性插值得到每个pixel的值乘以原本像素点。
16x16的分块并非固定,可以对块的大小进行调整,比如中心块偏大,靠近边缘的方块变小,这些都是可以自定义的,本算法由于做演示使用,故不做其他功能。如图所示:
code
由于代码量较大,这里分别附上一部分算法
shading lut caculate:
function [image_r_gain, image_gr_gain, image_gb_gain, image_b_gain] = ... isp_lsc_lut(image_r, image_gr, image_gb, image_b, side_num) [height, width] = size(image_r); side_y = floor(height/side_num); side_x = floor(width/side_num); % figure,imshow(image_r); % hold on; % for k=0:side_num % line_x = side_x * k; % line_y = side_y * k; % if(k==side_num && line_y ~= width) line_y = height;end % if(k==side_num && line_x ~= width) line_x = width;end % line([line_x,line_x],[0,height],'Color','red'); % line([0,width], [line_y, line_y],'Color','red'); % % line(Xd,Yd,'Color','red'); % end % hold off %% compress resolution image_point = zeros(side_num,side_num); for i = 0:side_num for j = 0:side_num x_clip = floor([j*side_x - side_x/2, j*side_x + side_x/2]); y_clip = floor([i*side_y - side_y/2, i*side_y + side_y/2]); if(i==side_num && y_clip(2) ~= height) y_clip(2) = height;end if(j==side_num && x_clip(2) ~= width) x_clip(2) = width;end x_clip(x_clip<1) = 1;x_clip(x_clip>width) = width; y_clip(y_clip<1) = 1;y_clip(y_clip>height) = height; data_r_in = image_r(y_clip(1):y_clip(2), x_clip(1):x_clip(2)); image_r_point(i+1,j+1) = mean(mean(data_r_in)); data_gr_in = image_gr(y_clip(1):y_clip(2), x_clip(1):x_clip(2)); image_gr_point(i+1,j+1) = mean(mean(data_gr_in)); data_gb_in = image_gb(y_clip(1):y_clip(2), x_clip(1):x_clip(2)); image_gb_point(i+1,j+1) = mean(mean(data_gb_in)); data_b_in = image_b(y_clip(1):y_clip(2), x_clip(1):x_clip(2)); image_b_point(i+1,j+1) = mean(mean(data_b_in)); end end % figure,imshow(uint8(image_r_point)); %% caculate lsc luma gain for i = 1:side_num+1 for j = 1:side_num+1 image_r_luma_gain_point(i,j) = mean2(image_r_point(uint8(side_num/2)-1:uint8(side_num/2)+1, uint8(side_num/2)-1:uint8(side_num/2)+1)) / image_r_point(i,j); image_gr_luma_gain_point(i,j) = mean2(image_gr_point(uint8(side_num/2)-1:uint8(side_num/2)+1, uint8(side_num/2)-1:uint8(side_num/2)+1)) / image_gr_point(i,j); image_gb_luma_gain_point(i,j) = mean2(image_gb_point(uint8(side_num/2)-1:uint8(side_num/2)+1, uint8(side_num/2)-1:uint8(side_num/2)+1)) / image_gb_point(i,j); image_b_luma_gain_point(i,j) = mean2(image_b_point(uint8(side_num/2)-1:uint8(side_num/2)+1, uint8(side_num/2)-1:uint8(side_num/2)+1)) / image_b_point(i,j); end end
bilinear interpolation:
image_r_luma_gain_reshape = reshape(image_r_luma_gain_point, [], 1); image_gr_luma_gain_reshape = reshape(image_gr_luma_gain_point, [], 1); image_gb_luma_gain_reshape = reshape(image_gb_luma_gain_point, [], 1); image_b_luma_gain_reshape = reshape(image_b_luma_gain_point, [], 1); for i = 1:17 for j = 1:17 x((i-1)*17+j) = i; y((i-1)*17+j) = j; end end x=x'; y=y'; % scatter3(x,y,image_r_luma_gain_reshape) % hold on Z=[ones(length(x),1),x,y,x.^2,x.*y,y.^2,x.^3,x.^2.*y,x.*y.^2,y.^3]; [x y]=meshgrid(1:17,1:17); A=Z\image_r_luma_gain_reshape; image_r_luma_gain=A(1)+A(2)*x+A(3)*y+A(4)*x.^2+A(5)*x.*y+A(6)*y.^2+A(7)*x.^3+A(8)*x.^2.*y+A(9)*x.*y.^2+A(10)*y.^3; A=Z\image_gr_luma_gain_reshape; image_gr_luma_gain=A(1)+A(2)*x+A(3)*y+A(4)*x.^2+A(5)*x.*y+A(6)*y.^2+A(7)*x.^3+A(8)*x.^2.*y+A(9)*x.*y.^2+A(10)*y.^3; A=Z\image_gb_luma_gain_reshape; image_gb_luma_gain=A(1)+A(2)*x+A(3)*y+A(4)*x.^2+A(5)*x.*y+A(6)*y.^2+A(7)*x.^3+A(8)*x.^2.*y+A(9)*x.*y.^2+A(10)*y.^3; A=Z\image_b_luma_gain_reshape; image_b_luma_gain=A(1)+A(2)*x+A(3)*y+A(4)*x.^2+A(5)*x.*y+A(6)*y.^2+A(7)*x.^3+A(8)*x.^2.*y+A(9)*x.*y.^2+A(10)*y.^3; % surf(x,y,image_r_luma_gain) % hold on % surf(x,y,image_r_luma_gain_point) %% calulate lsc chroma gain for i = 1:side_num+1 for j = 1:side_num+1 image_r_chroma_gain(i,j) = image_r_luma_gain(i,j) - image_r_luma_gain_point(i,j); image_gr_chroma_gain(i,j) = image_gr_luma_gain(i,j) - image_gr_luma_gain_point(i,j); image_gb_chroma_gain(i,j) = image_gb_luma_gain(i,j) - image_gb_luma_gain_point(i,j); image_b_chroma_gain(i,j) = image_b_luma_gain(i,j) - image_b_luma_gain_point(i,j); end end %% caculate lsc result gain image_r_gain = image_r_luma_gain - image_r_chroma_gain; image_gr_gain = image_gr_luma_gain - image_gr_chroma_gain; image_gb_gain = image_gb_luma_gain - image_gb_chroma_gain; image_b_gain = image_b_luma_gain - image_b_chroma_gain; function image_gain_lut = lsc_data_gain_interpolation(image_gain, height, width, side_num) side_y_ori = floor(height/side_num); side_x_ori = floor(width/side_num); k = 0; l = 0; [gain_height, gain_width] = size(image_gain); for i = 1:gain_height-1 for j = 1:gain_width-1 data_gain_11 = image_gain(i, j); data_gain_12 = image_gain(i, j+1); data_gain_21 = image_gain(i+1, j); data_gain_22 = image_gain(i+1, j+1); if(j == gain_width-1 && ((j-1)*side_x + l) ~= width) side_x = width - (j-1)*side_x_ori; else side_x = side_x_ori; end if(i == gain_width-1 && ((i-1)*side_y + k) ~= width) side_y = height - (i-1)*side_y_ori; else side_y = side_y_ori; end for k = 1:side_y for l = 1:side_x label_y1 = 1; label_x1 = 1; label_y2 = side_y; label_x2 = side_x; image_gain_lut((i-1)*side_y_ori + k, (j-1)*side_x_ori + l) = ... data_gain_22/(label_x2-label_x1)/(label_y2-label_y1)* ... (l - label_x1) * (k - label_y1) + ... data_gain_21/(label_x2-label_x1)/(label_y2-label_y1)* ... (label_x2 - l) * (k - label_y1) + ... data_gain_12/(label_x2-label_x1)/(label_y2-label_y1)* ... (l - label_x1) * (label_y2 - k) + ... data_gain_11/(label_x2-label_x1)/(label_y2-label_y1)* ... (label_x2 - l) * (label_y2 - k); end end end end end
效果展示:
实验条件有限,图片有水波纹,仅用于理解算法
original image:
luma shading
chroma shading:
luma shading + chroma shading:
tuning
LSC的tuning一定要把校正图采集好,一般情况下raw图的G通道中心亮度在8bit的70%~80%之间,由于在不同色温情况下是经过插值的,因此需要校正多个光源,一般情况下TL84、D65、A光源下进行校正。将得到的LUT写入RAM中即可
注意:采集的raw图不要有filcker。
LSC强度一般是可调的,由于图像边角的增益会很大,因此在高倍gain下,可以把强度给降低,防止图像边角噪声压不住的情况。
由于各个平台不同,这里不做详细介绍,想到再补充。