How can I identify a particular shape at all degrees, while ignoring the rest?

Hello, I have an image of various lines and blobs, and I want my image to only show the lines. To accomplish this, I have utilized watershed, regionprops, ismember, and a few other methods in order to get to the image I am now. At this point, I can't reduce the area boundries, as I'll start losing lines before I lose the blobs, and solidity wouldn't accomplish anything as all of the shapes appear to have the same solidity (probably due to how the shapes were generated.) What options do I have left? Thanks in advance!

[Florian Morsch]

You could try to use SURF features, but if the object you want to identify is the long line-wise one then you will have a really hard time.

[Kimo Kalip]

Yeah, I have thousands of those longer lines at various angles (So unfortunately there is no uniform orientation). The idea is to reduce the image to the point where only those white lines remain, and all extraneous noise is weeded out - but I'm wondering if I've hit a wall. Thanks for the idea though!

Answer 1

[Image Analyst]

Try the ratio of the majoraxislength to minoraxisLength.

[Kimo Kalip]

Thanks for the input! If I'm following correctly, the idea is the lines I want to identify will always have a longer length than height (in one direction or another, so the actual orientation shouldn't matter if they're rotated around at various degrees right?). So would I then utilize regionprops somehow to weed out any blobs that don't fit that ratio?

[Kimo Kalip]

Following the general idea of the code provided at https://www.mathworks.com/matlabcentral/answers/356649-regionprops-major-minor-axis-determination-problem, I've gathered major and minor axis values:

• Major =
• 67.2744522341701
• 61.8447090104933
• 65.5798639882114
• 63.616962417864
• 18.7911211329771
• 62.6008185063681
• 60.1405039779664

Minor =

• 4.31004827260943
• 4.10319533153696
• 4.21845934806795
• 4.12938961158931
• 9.69429007605594
• 4.20106771691527
• 3.92452030731054

Granted, this is before the:

for k = 1:length(s)
      a(k,1) = s(k).MajorAxisLength/2;
      b(k,1) = s(k).MinorAxisLength/2;

I'm not sure why it was needed to divide by 2, but in any case, I think the 18.79:9.694 is probably a blob I'd like to get rid of, where as most of my other lines seem to have a larger disparity between the axis lengths. Where should I go from here?

[Image Analyst]

Try this (untested)

labeledImage = bwlabel(binaryImage);
props = regionprops(labeledImage, 'MajorAxisLength', 'MinorAxisLength');
aspectRatios = [props.MajorAxisLength] ./ [props.MinorAxisLength];
ratioThreshold = 8; % Whatever you want
stickLikeShapeIndexes = find(aspectRatios > ratioThreshold);
% Extract only those blobs shaped like sticks:
binaryImage = ismember(labeledImage, stickLikeShapeIndexes);
% Now call regionprops on the new binary image with only stick-like objects in it.
labeledImage = bwlabel(binaryImage);
props = regionprops(labeledImage, 'Area');

[Kimo Kalip]

Thank you sir, you are a hero! (Seriously, I've read like 30 of your posts over the last week trying to figure out how to do all this stuff). I've reduced my images down to just white lines, which is a miracle, but now I have to figure out how to automate the process. Is there mathematical reasoning behind the thresholds? (You often put histograms in your examples, but I don't understand how they relate as the red bar representing the threshold is often far removed from the hump.) In your latest response here, I image I could calculate the averages for the aspect ratios then set the threshold to be somewhat below that and it should work...

Anyway, your help has been invaluable, thank you so much!

[Image Analyst]

Picking thresholds is a bit of an art, and sometimes there just is no global threshold that will get everything accurately. You might want to look at my visual/interactive threshold picking app in my File Exchange.

I don't know anything about the image or the histogram you plotted, nor how the threshold was arrived at so I can't say whether it's any good or not. If you want help with that post the original gray scale image and say what features or objects you consider to be foreground and which you consider to be background.

[Kimo Kalip]

I definitely agree that it feels like an art. Luckily, a threshold of 80 seems to be working for me for the most part, but sometimes the lines are of various lengths, or what I need to detect are in fact dots, or they are different colors and shapes etc. I don't have permission to post the actual image online, but for lack of a better example I have included a picture of a donut. What I am trying to accomplish is identifying each of the lighter sprinkles on the donut (white sprinkles or frosting on the background should be considered 'noise'). Using the techniques you suggested in https://www.mathworks.com/matlabcentral/answers/166363-edge-detection-method-for-image-cropping, I've managed to weed out a lot of the background, but wondering if there is a relatively easier way of autocropping the image around the donut (When donuts can be various sizes, I can't simply hardcode a cropping, but wondering if I can just detect the edges of the circle before modifying the image further, or simply attempting to ignore the background as considered in the example). After a lot of messing with thresholds and whatnot, I'm happily convinced it is possible, but not sure if there is mathematical logic that I can program the computer to do most if not all of the adjustments necessary.

Anyway, I'll take a look at that thresholding image tutorial you linked, and I think I'm on the right track, so thanks again!

Also, if what I'm trying to accomplish sounds rather unrealistic, knowing now before I go deeper down the rabbit hole would be appreciated as well...

[Kimo Kalip]

As an aside, for one of my images I have:

rows = 4318

columns = 4318

numberOfColorBands = 1

And surprisingly for the most part, this chunk of code runs almost perfectly the same unchanged (Except for the fact that apparently I don't have licensing for the Saviztky-Golay filter, but it was optional anyway)

==========================

 % Let's get the vertical profile
 sides = [grayImage(:, 1:70), grayImage(:, 300:end)];
 verticalProfile = mean(sides, 2);

 % verticalProfile = int8(verticalProfile);
 % Let's smooth it with a Saviztky-Golay filter (optional)
 % verticalProfile = sgolayfilt(verticalProfile, 2, 45);

 subplot(3, 3, 3); 
 plot(verticalProfile, 'b-', 'LineWidth', 2);
 grid on;
 title('Vertical Profile of Background', 'FontSize', fontSize);

 % Make a background image that we can divide by.
 backgroundImage = repmat(verticalProfile, [1, columns]);
    subplot(3, 3, 4);
 imshow(backgroundImage, []);
 axis on;
 title('Background Image', 'FontSize', fontSize);

 % divide the image by the profile
 flattenedImage = uint8(255 * double(grayImage) ./ backgroundImage);
 subplot(3, 3, 5);
 imshow(flattenedImage, []);
 axis on;
 title('Background Corrected Image', 'FontSize', fontSize);

========================================

Will changing some of the dimensions here to fit my actual image get rid of some of that extraneous noise in the background?

[Image Analyst]

I don't know that you need autocropping. Why can't you just measure what's there in the original image? What does cropping get you? If you were picking up a lot of noise in the background, you can get rid of that with masking (not cropping) by finding the big donut and using that to exclude any junk not in that (i.e. in the background). Simply do

% Mask the image using bsxfun() function to multiply the mask by each channel individually.
maskedRgbImage = bsxfun(@times, rgbImage, cast(mask, 'like', rgbImage));

As far as finding light sprinkles you'd have to convert to a good color space for determining that, like lab or hsv, then do color segmentation and find sprinkles of the "light" colors. You may also have to do color calibration by comparing your image to a standard chart like the x-rite Color Checker Chart.

[Kimo Kalip]

aspectRatios = [props.MajorAxisLength] ./ [props.MinorAxisLength];
ratioThreshold = 5; % Whatever you want
ratioMaximum = 7;
stickLikeShapeIndexes = find(aspectRatios > ratioThreshold && aspectRatios < ratioMaximum);

Bit of a continuation of my previous question, lets say I wanted to impose a minimum and maximum limit. Ideally, I'd like to get the aspect ratios of each individual piece, find an average, and then filter out anything "too small" or "too large". The first question here is: this && doesn't work here, the other idea would be just to break it up into two parts, but is that the only way? Secondly, from my understanding aspectRatios is an array, correct? Is there a way for me to calculate an average from the numbers, throw out the extremes, then calculate an average again based on numbers without extremes?

Also, the masking seems to work, so thank you for the help!

[Kimo Kalip]

So I have successfully filtered out large unwanted aspect ratios from my image, but in this variant I have several that have an aspect ratio of around 1. The problem is, this little pixel I'm zoomed in here also has a aspect ratio of 1 (I think, having a hard time verifying, but I think its a reasonable assumption).

Is there a way to get bwconncomp() to have a minimum and maximum "object" size before you run it? The way I've been doing it so far is filtering out the image prior, but if there is a way to teach the program what is and isn't a reasonable and regular size, it may be easier.

[Image Analyst]

You can use bwareafilt() if you want instead.

Answer 2

[Image Analyst]

You might enjoy learning about Hu's moments: http://www.youtube.com/watch?v=Nc06tlZAv_Q

[Kimo Kalip]

Pretty close more or less to what I'm trying to achieve. Does he upload a reference photo for each logo at the start there though? In my case, the idea is that there are thousands of relatively similar shaped boxes rotated around a circle, the image is grayscale, and they usually the darkest thing on the image (Although not always, which is another thing I need to account for). What I'm getting at is: in many cases I won't have a reference to provide, other than the other objects in the image - does this still work? Also, the mms function he uses to determine the moments was something he wrote himself, and not provided by matlab? Thanks for the video though! End result was pretty cool