Help increasing the speed of a piece of code
3 visualizzazioni (ultimi 30 giorni)
Mostra commenti meno recenti
Dear all,
Plain summary: I have a temperature matrix and I need to find the average temperature of the 8 pixels surrounding each pixel.
I need to subdivide my 134x134 matrices into 3x3 sub matrices. For each submatrix I need to compute the mean and the standard deviation, excluding the central pixel from the calculation. Then I need to find the values exceding the upper and lower boundaries, namely the mean + standard deviation and the mean - standard deviation, respectively. Then, I need to recalculate the mean excluding the values falling outside the lower and upper boundaries and assign the mean value to the central pixel location (in a new matrix).
The output will be a 134x134 matrix with every value (pixel) corresponding to the average of its surrounding pixels with the "outliers" excluded from the average.
I wrote a code (loop) to do the job, but it was taking ~20 seconds (link to old question and previous code: https://it.mathworks.com/matlabcentral/answers/1947148-help-improving-speed-of-this-code-loop?s_tid=srchtitle ). Someone helped me to improve the original code (see new code below), but it still takes ~2 seconds.
As I need to run it 10k+ times it will take ages to complete. I wonder if you could kindly suggest a faster code to achieve the same result?
Thanks a lot!
A=magic(134); % Assuming this is my 134x134 Temperature matrix
T_background = blockproc(A,[1,1],@BlkFun, 'BorderSize',[1,1],...
'TrimBorder',false, 'PadMethod',NaN);
%where BlkFun=
function v = BlkFun(s)
px = s.data;
if isequal(size(px),[3,3])
px(5) = [];
av = nanmean(px);
sd = nanstd(px);
ix = px>=(av-sd) & px<=(av+sd);
v = nanmean(px(ix));
else
v = NaN;
end
end
0 Commenti
Risposta accettata
Rik
il 15 Apr 2023
Modificato: Rik
il 15 Apr 2023
It sounds like a convolution would be perfect here.
kernel=ones(3,3);
kernel(2,2)=0;
kernel=kernel/sum(kernel(:));
That will find the average of the surrounding pixels.
Another strategy would be to create a 3D array with all shifts. Then you can use the functions already in your function. Instead of removing elements (or indexing) you should mark invalid values with NaN. Memory limits should not be an issue since it is only 134x134x8.
3 Commenti
Rik
il 15 Apr 2023
A=magic(4); % Assuming this is my 134x134 Temperature matrix
% pad the array with NaNs to make circshift easier
A_pad = [...
NaN(1,2+size(A,2));...
NaN(size(A,1),1) A NaN(size(A,1),1);...
NaN(1,2+size(A,2))];
A_3D = repmat(A_pad,[1 1 8]); % Extend to 3D
k = 0;
for row_shift = [-1 0 1]
for col_shift = [-1 0 1]
% skip the center pixel
if row_shift==0 && col_shift==0,continue,end
k = k+1;
A_3D(:,:,k) = circshift(A_pad,[row_shift col_shift]);
end
end
%A_3D([1 end],[1 end],:) = []; % remove NaN padding to get back to the original size
A_3D([1 end],:,:) = [];A_3D(:,[1 end],:) = []; % Matlab complains if you do this in one go
A_3D
Now you can use this array to calculate the SD and mean to your heart's content:
mean(A_3D,3,'omitnan')
Più risposte (1)
per isakson
il 15 Apr 2023
Modificato: per isakson
il 15 Apr 2023
"I need to do the other steps". Removing NaN before calculating the mean cuts the elapse time in half (on my PC with R2018b)
tic
A=magic(134); % Assuming this is my 134x134 Temperature matrix
T_background = blockproc(A,[1,1],@BlkFun, 'BorderSize',[1,1],...
'TrimBorder',false, 'PadMethod',NaN);
toc
function v = BlkFun(s)
px = s.data;
if isequal(size(px),[3,3])
px(isnan(px)|[0,0,0;0,1,0;0,0,0]) = [];
av = mean(px);
sd = std(px);
ix = px>=(av-sd) & px<=(av+sd);
v = mean(px(ix));
else
v = NaN;
end
end
and hopefully it returns the same result.
Vedere anche
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!