Image processing - pixel comparison

Suppose there is a 5X5 filter window to traverse a image with zero padding.

My Question:

I'm trying to compare all the pixels in row-1 in 1 iteration. If pixels in 1st row are all similar, then RC for row-1 is 1.

similarly, for row-2, comparing pixels in 2nd row. If similar, then RC for that row-2 is 1.

Can anyone suggest some methodology/approach for this implement?

Even if not in one iteration, some other approach?

[Image Analyst]

What does "similar" mean to you? How similar? Perfect equality? Some value within a certain tolerance of each pixel? Did you try movmin() and movmax() to get the min and max within the window?

[Archit Sethi]

I would elucidiate my question & approach:-

Similarity of the pixels in row or column is done by comparing only their 4 most significant bits.

Step-1: algo compares the pixels in a row and column. If pixels in a row are similar(equal), the row comparison signal for that particular row is 1. Similarly do for column.

Then, if pixels in all rows are similar, R signal =1. similarly if pixels in all columns are similar, C signal = 1.

Step-2: Median is computed.

Row_Comparison = compare(MSB 4 bits of pixels in each row)

Column_Comparison = compare(MSB 4 bits of pixels in each column)

R_signal = (RC[0] & RC[1] & RC[2] & RC[3] & RC[4])

C_signal = (CC[0] & CC[1] & CC[2] & CC[3] & CC[4])

If not able to understand, I would share my script to see current approach.

[Archit Sethi]

Can anyone help me with adaptive median filter code with these steps:

window_size = 5;

adaptive_MF(window_size){

Row_Comparison = compare(MSB 4 bits of pixels in each row)

Column_Comparison = compare(MSB 4 bits of pixels in each column)

R_signal = (RC[0] & RC[1] & RC[2] & RC[3] & RC[4])

C_signal = (CC[0] & CC[1] & CC[2] & CC[3] & CC[4])

if (R_signal is 1 and C_signal is1)

Median = window(2,2)

elseif (R_signal is 1 or C_signal is 1)

Median = median(diagonal pixels of filter window 5x5)

else{

m1 = median(diagonal pixels of filter window 5x5)

m2 = median(Horizontal pixels from 2nd row of filter window 5x5)

m3 = median(Verticla pixels from 2nd column of filter window 5x5)

Median = median(m1,m2,m3)

}

window(2,2) = Median

}

I've tried many times this algo in MATLAB, but array index out of bounds error or outlined error pops ups? I've no clue how to tackle this problem with above pseudo code.

Also, i did msb = mod(floor(image/128),2); %for MSB extraction so that i can do pixel comparisons. But no luck.

Please help!

Answer 1

[Image Analyst]

For what it's worth, see my attached modified median filter demo.

[Archit Sethi]

Hi Image Analyst!

Yeah I had a look at ur code. But, I'm following the above mentioned algo for 2D adaptive median filter approach. Can you help with that code/Algo?

Please that would be very helpful!

Answer 2

[DGM]

The filter you describe doesn't make use of its height. If it's only evaluating each row within the window independently, it's a 1-D filter and only needs to be 1x5.

It's unclear what type of image you're working with. If we assume this is a binary image, then the answer is simple:

% read a test image and clean it up
inpict = im2double(imread('circles.png'));
inpict = inpict>0.1; % make it logical

anytrue = imdilate(inpict,ones(1,5));
anyfalse = imdilate(~inpict,ones(1,5));

outpict = xor(anytrue,anyfalse);

If you want to do this for a grayscale or color image, you're going to have to define some tolerance and test for equality within said tolerance. You may be able to use rangefilt() for grayscale inputs.

inpict = im2double(imread('cameraman.tif'));
inpict = medfilt2(inpict,[3 3]); % clean it up a bit

tol = 0.03;

rangemap = rangefilt(inpict,ones(1,5));
outpict = rangemap<tol;

For RGB inputs, you could use rangefilt() to generate range maps, test against the tolerance to get three logical maps, then logically combine them.

[Archit Sethi]

Its a 2D fitler. Using both Rows and columns.

im using lena.jpg 512 x 512.

Instead of converting the image to double, can we do on uint8()?But how?

[Archit Sethi]

Yes, the row & col comparison are being done in a particular window which moves on the image like any other filter methodology.

[DGM]

You can do it in uint8. You'd just set your threshold accordingly.

The filter you describe in your comment to Image Analyst is certainly a 5x5 filter, but it has two outputs and doesn't really match my example. I don't see an easy way to make use of existing filter tools like rangefilt() for this.

You may be able to use nlfilter() to do it, though in my experience, it's usually significantly faster to just write a sliding window filter in m-code (believe it or not).

[Archit Sethi]

For a 5x5 filter, can you explain or write some piece of script to know more about how i can code in above mentioned algorithm.

I'm struggling to write this for quite some time...but no success till now.

[DGM]

This example uses nlfilter():

inpict = rgb2gray(imread('sources/lena.tif'));

rowpict = nlfilter(inpict,[5 5],@rowfilt);
colpict = nlfilter(inpict,[5 5],@colfilt);

imshow(cat(1,rowpict,colpict))

function out = rowfilt(in)
    tol = 16;
    out = all(range(in,2)<tol);
end

function out = colfilt(in)
    tol = 16;
    out = all(range(in,1)<tol);
end

I'm not really sure what should be done with the two outputs at this point. Either way, that's how nlfilter() can be used.

This is the same thing using a simple sliding window written with loops:

inpict = rgb2gray(imread('sources/lena.tif'));
filtsize = [5 5];
tol = 16;

s0 = size(inpict);
padsize = floor(filtsize/2);
% nlfilter() uses zero-padding, but you could replicate, etc
inpict = padarray(inpict,padsize,0,'both');
s = size(inpict);

rowpict = false(s0);
colpict = false(s0);
osm = filtsize(1)-1;
osn = filtsize(2)-1;

for n = 1:(s(2)-2*padsize(2))
    for m = 1:(s(1)-2*padsize(1))
        sample = inpict(m:(m+osm),n:(n+osn));
        rowpict(m,n) = all(range(sample,2)<tol);
        colpict(m,n) = all(range(sample,1)<tol);
    end
end

imshow(cat(1,rowpict,colpict))

As much as everyone rags on writing things with loops, this is still ~3x as fast as using nlfilter()