Matlab and Otsu Thresholding query
in [Matlab]
|
From: Kenneth Galea on 5 Feb 2010 18:40 "Kenneth Galea" <k.galea(a)hotmail.com> wrote in message <hki1bk$ee5$1(a)fred.mathworks.com>... > ImageAnalyst <imageanalyst(a)mailinator.com> wrote in message <43b5892c-04fa-4b61-b1cc-328deeb03719(a)t1g2000vbq.googlegroups.com>... > > Because I wanted to select objects that had high red values but low > > green and low blue values. Objects that have high red and low green > > and blue signal will appear as red in the image. Make sense? > > > > As a counter example if I wanted greenish things, I'd take reds in the > > range 0-redThresh, greens in range greenThresh-255, and blues in the > > range 0-blueThresh. > > > > If I wanted blue things I'd take pixels having higher blue values and > > lower red and green values. > > Does that make more sense now? > > Yes thanks a lot. Now I my problem is that when I subtract a white image from another white image some "salt and pepper nosie is seen". I used medfilt2 but its still not enough. After I used bwareaopen which didn't make any diffenerce even if I consider an area of 50 pixels...don't know why its not working!! Is there another method for this noise removal?? > > Thanks > Kenneth Hi sorry to send you again. I am having a bit of the same problem again with subtraction. I tried the background division as an alternative method and when I divided the test by background, the images are not even recognisable at that point. So I decided to keep with subtraction method. In this case I'm ensuring that illumination is always the same and from the same position (i'm not moving the camera) and when I compare a test image (containing an object) with the background image its perfect.....however when I compare a white image with another white image taken exactly from the same place with the same amount of light but taken a few seconds later there is a noticable difference since the result is quite a noisy image which sometimes can't be removed by filtering. Why am I still obtaining this difference....it's really frustrating !!!?? Maybe if i'll buy a camera with flash incorporated would it be better?.... since I'm really clueless at this stage:( Thanks Kenneth
From: ImageAnalyst on 5 Feb 2010 19:45 Kenneth You need to remove noise from your reference (background) image. Otherwise the noise essentially doubles up because you have noise from both images and they'll add in quadrature more or less. Imagine that you had an exposure that was hours long - you wouldn't have any video noise would you? This is what you want to achieve because this would represent the "true" illumination pattern. So how can you get that? Well you could average together a bunch of images to beat down the noise. Or you could try to run some noise reduction filters on the image, such as simple averaging or median filters, or more sophisticated methods like bilateral filters, sigma filters, etc. to the even more sophisticated methods like "non-local means," BM3D (http://www.cs.tut.fi/~foi/GCF-BM3D/), UINTA (http://www.cs.utah.edu/ ~suyash/pubs/uinta/), K-LLD (http://users.soe.ucsc.edu/~milanfar/ talks/), or many others. Of course more light is better but you have to have controlled light, which light from a flash IS NOT. If you do that, heck even if you don't, you might want to put a standard in your image, such as a gray scale step wedge running along one edge of your image. Then you can check that and apply the appropriate intensity correction to your test image to bring it back to the standard. Once you've removed the noise from your background, you'll have only the noise in your image to worry about. You can use the same kind of techniques there. It really depends on what analysis you want to do as far as how much effort you need to put into a noise reduction method. If the simplest works for you, then great. Otherwise try better ones. Why don't you post your "good, improved" intensity- corrected images and I'll see what I can do with them this weekend?
From: ImageAnalyst on 5 Feb 2010 19:53 Here's a big list of noise reduction methods, including some with MATLAB implementations: http://www.stanford.edu/~slansel/tutorial/software.htm
From: Kenneth Galea on 5 Feb 2010 20:33 Hi I wanted to post my final code first so that you can easily understand what I did: background = imread('C:\Users\Kenneth\Desktop\background.jpg'); subplot(3,3, 1); imshow(background); set(gcf, 'Position', get(0, 'ScreenSize')); % Maximize figure. title('Background Image'); testImage = imread('C:\Users\Kenneth\Desktop\test3.jpg'); subplot(3,3, 2); imshow(testImage); title('Test Image'); redPlane_test = double(testImage(:, :, 1)); greenPlane_test = double(testImage(:, :, 2)); bluePlane_test = double(testImage(:, :, 3)); redPlane_background = double(background(:, :, 1)); greenPlane_background = double(background(:, :, 2)); bluePlane_background = double(background(:, :, 3)); subtractedImageRed = redPlane_test - redPlane_background; subtractedImageGreen = greenPlane_test - greenPlane_background; subtractedImageBlue = bluePlane_test - bluePlane_background; subplot(3,3, 3); imshow(subtractedImageRed, []) title('Subtracted Red'); redThreshold = graythresh(subtractedImageRed) greenThreshold = graythresh(subtractedImageGreen) blueThreshold = graythresh(subtractedImageBlue) redThreshold = uint8(redThreshold * 255); greenThreshold = uint8(greenThreshold * 255); blueThreshold = uint8(blueThreshold * 255); binaryImageRed1 = (subtractedImageRed < -redThreshold) ; subplot(3,3, 4); imshow(binaryImageRed1, []) ; title('Dark things'); binaryImageRed2 =(subtractedImageRed > redThreshold); subplot(3,3, 5); imshow(binaryImageRed2, []); title('Bright things'); binaryImageRed = (subtractedImageRed < -redThreshold) |(subtractedImageRed > redThreshold); binaryImageGreen = (subtractedImageGreen < -greenThreshold) | (subtractedImageGreen > greenThreshold); binaryImageBlue = (subtractedImageBlue < -blueThreshold) | (subtractedImageBlue > blueThreshold); binaryImage = binaryImageRed & binaryImageGreen & binaryImageBlue ; binaryImage = ~binaryImage; subplot(3,3, 6); imshow(binaryImage) title('Both Dark and Bright things'); figure imshow(binaryImage) binaryImage = im2bw(binaryImage) filtered_binaryImage = medfilt2(binaryImage); figure imshow(filtered_binaryImage) Link: http://drop.io/thresholdbackground This link shows what I was trying to work on (N.B. these two images are 2 different background images taken exactly one after each other - everything the same though as I said when subtracting noise is observed as can be seen in the third image "result"). In my opinion, with objects thresholding was better when graythresh was used on a gray image instead of considering red,green and blue. However in the case when there are no objects using the latter method gave better results ...which still need to be improved as you can see in the link. Regards Kenneth
From: ImageAnalyst on 5 Feb 2010 22:44
Kenneth I'm not exactly sure how to respond. With objects in the image and a blank background you can use a threshold on monochrome images, or use color classification on color images. With color images, simple color classification can sometimes be done with thresholding the three color bands, which will essentially select out a rectangular chunk out of the 3D rgb color space. You pick a threshold such that it will select the objects and not select the background. If you can do this with an auto threshold, such as the Otsu method used by graythresh, then fine. However a simple auto method (like gray thresh) will force a threshold to be picked that divides the image into foreground and background. If you're looking at a picture that has no foreground because it's all background then it will basically split the background into two classes (light background and dark background), which is not what you want. That's why I said to use a fixed threshold. It will always pick out the proper things regardless of how large or small the foreground objects are. With an autothreshold method, the threshold that is picked depends on the area fraction of the objects in the foreground. If they are of substantial size with respect to the background it may pick a good threshold value but if the foreground is too small or nonexistent then it will try to pick a threshold based on just the background which will of course be unacceptable. In your latest case you had no foreground and so the subtraction is essentially all just noise. Then the autothreshold picked some gray level right in the midst of that noise and so it picks out the strongest noise as foreground objects. Obviously you don't want that. If you're going to have images that range anywhere from big objects to small objects to no objects, then you're going to have to have a threshold selection method that will give you the correct threshold. I've had cases like this and I can tell you that I use a fixed threshold. You pick a threshold and if there's something above it, it's an objects. If there's nothing above it there's o objects. It works for any and all object sizes, from an object that is 100% of the pixels all the way down to 0% of the pixels. It ALWAYS works, unlike an autothreshold method which only works well if the foreground and background can be split into two histogram humps (unless of course the auto method is smart enough to revert to a fixed threshold when it recognized that the test image histogram is essentially no different than the background histogram, but then we're back to the fixed threshold again.) So if you want to be completely general I'd stick with a fixed threshold. But if you're always going to have objects that take up a substantial portion of the image (like your scissors and pen image) AND you're going to correct for shading in the background illumination, then an auto threshold can work. But like I said, it won't work for the "no objects" case. |