Fast Morphological image processing using cuda-7.5

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello all, i have implemented VHGW algorithm for faster morphological image processing(dilation and erosion).
I have implemented using matlab and by using parallel computing toolbox of matlab. BUt the problem is that by using MATLAB parallel computing toolbox, the gpu performance is slower than cpu version. This should be not the case.

I just want someone please implement this piece of code using cuda_7.5.
Below is i am writing matlab piece of code:-

clc;
clear vars;
close all;

img = rgb2gray(imread(‘D:\Desktop Works\TexturesCom_AnimalsVarious0015_1_S.jpg’));
%img = [114 16 251 37];
p = input(‘Enter the Structure element size(must be odd):’);

[rows, cols] = size(img);

%disp(mod(cols,p));

%img = gpuArray(img);
if(mod(cols, p) == 0)
left_pad = 0;
right_pad = 0;
else
pad = p - (mod(cols, p));
left_pad = pad/2;
right_pad = pad - left_pad;
end

left_pad = ceil(left_pad + (p - 1)/2);
right_pad = ceil(right_pad + (p - 1)/2);

ncols = cols +left_pad+right_pad;
%proc_mat = zeros([rows, cols]);

R = (zeros([1524, 1524]));
S = (zeros([1524, 1524]));

proc = (zeros([1524, 1524]));
result = (zeros([rows, ncols]));
output = (zeros([rows, cols]));

tic;

start = 1;
to_end = cols;
for i = 1: rows
for j = to_end:-1:1
proc(i,(j+left_pad)) = img(i,j);
end
end

for i = 1:rows
for j = start: left_pad
proc(i,j) = -128;
end
end

start = cols+left_pad+1;
to_end = start+right_pad-1;
for i = 1 : rows
for j = start:to_end
proc(i,j) = -128;
end
end

for i = 1 : rows
count = p;
for j = 1 : ncols - p + 1
count = count - 1;
maxNo = proc(i,j);
t = 0;
for k = 1 : count
t = t + 1;
num = proc(i, (j + t));
if(maxNo < num)
maxNo = num;
end
end
R(i,j) = maxNo;

    if(count == 0)
        count = p;
    end
end

end

for i = 1 : rows
count = 0;
for j = p:ncols
maxNo = proc(i,j);
t = 0;
for k = 1:count
%t = t - 1;
t = t + 1;
num = proc(i, (j - t));
if(maxNo < num)
maxNo = num;
end

    end
    S(i, j) = maxNo;
    count = count + 1;
    %t = t + 1;
    if count == p 
        count = 0;
        %t = 0;
    end
end

end

index = (p - 1)/2;
for i = 1: rows
for j = index+1: ncols - index
result(i, j) = max(R(i, (j - index)), S(i,(j + index)));
end
end

for i = 1:rows
for j = 1 : cols
output(i, j) = result(i, (j+left_pad));
end
end
val = toc;
imshow(mat2gray(output));
disp(num2str(val));

img1 = img’;
[rows, cols] = size(img1);

if(mod(cols, p) == 0)
left_pad = 0;
right_pad = 0;
else
pad = p - (mod(cols, p));
left_pad = pad/2;
right_pad = pad - left_pad;
end

left_pad = ceil(left_pad + (p - 1)/2);
right_pad = ceil(right_pad + (p - 1)/2);

ncols = cols +left_pad+right_pad;

R = (zeros([1524, 1524]));
S = (zeros([1524, 1524]));

proc = (zeros([1524, 1524]));
result = (zeros([rows, ncols]));
output = (zeros([rows, cols]));

tic;

start = 1;
to_end = cols;
for i = 1: rows
for j = to_end:-1:1
proc(i,(j+left_pad)) = img1(i,j);
end
end

for i = 1:rows
for j = start: left_pad
proc(i,j) = -128;
end
end

start = cols+left_pad+1;
to_end = start+right_pad-1;
for i = 1 : rows
for j = start:to_end
proc(i,j) = -128;
end
end

for i = 1 : rows
count = p;
for j = 1 : ncols - p + 1
count = count - 1;
maxNo = proc(i,j);
t = 0;
for k = 1 : count
t = t + 1;
num = proc(i, (j + t));
if(maxNo < num)
maxNo = num;
end
end
R(i,j) = maxNo;

    if(count == 0)
        count = p;
    end
end

end

for i = 1 : rows
count = 0;
for j = p:ncols
maxNo = proc(i,j);
t = 0;
for k = 1:count
%t = t - 1;
t = t + 1;
num = proc(i, (j - t));
if(maxNo < num)
maxNo = num;
end

    end
    S(i, j) = maxNo;
    count = count + 1;
    %t = t + 1;
    if count == p 
        count = 0;
        %t = 0;
    end
end

end

index = (p - 1)/2;
for i = 1: rows
for j = index+1: ncols - index
result(i, j) = max(R(i, (j - index)), S(i,(j + index)));
end
end

for i = 1:rows
for j = 1 : cols
output(i, j) = result(i, (j+left_pad));
end
end

output = output’;
val = toc;
disp(num2str(val));

2

There are a bunch of websites specialized in connecting freelancers with customers for small programming jobs (Google is your friend). This is not one of them.

What have you tried in terms of your own CUDA implementation, and where specifically did you get stuck?

3

  1. use code blocks rather than pasting your MATLAB code directly

  2. So you want somebody, for free, to do your work for you so you can take credit for work you did not do?

4

You could employ the NVIDIA NPP library (https://developer.nvidia.com/npp) or OpenCV or ArrayFire. Although, I think Matlab Parallel Toolbox uses ArrayFire under the hood so it won’t change your runtime. Note if you need only binary dilation/erosion (and not grayscale) of a mask, you can implement the dilation/erosion as an box filter convolution with an additional thresholding step as the end and take advantage of optimized box filter routines.

5

Well, strictly speaking one can post job offers here, but it’s in a sibling topic very much on the same forum.

https://devtalk.nvidia.com/default/board/54/gpu-computing-jobs/

6

As far as I can tell, the jobs sub-forum of this site is geared towards longer-term jobs and consulting opportunities, and isn’t a good venue for finding freelancers for small 1-2 day programming tasks, that is why I did not send the OP there.