CODEWITHSUNDEEP
matlabimage-processingoptimizationconvolution
Alex Azazel
Alex Azazel

Reputation: 342

How to speed up 3 channel image convolution in MATLAB?

I'm using convolution for neural networks, currently it's being implemented like this

for f = 1:NumberOfKernels
    tempC = conv2(input(:,:,1),kernels(:,:,1,f),'same');
    tempM = conv2(input(:,:,2),kernels(:,:,2,f),'same');
    tempY = conv2(input(:,:,3),kernels(:,:,3,f),'same');
    preactivation(:,:,f) = tempC + tempM + tempY;
end

Can this be done in a single line without writing out conv2 for each color channel individually? Can this function from Image Processing Toolbox speed it up? Take into account that I have no GPU.

Upvotes: 0

Views: 455

Answers (1)

user664303
user664303

Reputation: 2063

You could do the following:

szk = size(kernels);
temp =  zeros(size(input)+szk(1:3)-1);
szk = ceil(szk(1:2) / 2);
temp(szk(1):szk(1)-1+size(input,1),szk(2):szk(2)-1+size(input,2),:) = input;
for f = 1:NumberOfKernels
   preactivation(:,:,f) = convn(temp,kernels(:,:,:,f),'valid');
end

However, I wouldn't expect it to be much faster. What would make things faster was if the kernels were separable.

Upvotes: 1

Related Questions