File Exchange

version 1.4 (207 KB) by

A fast an well performing local image thresholding method.

4.85714
7 Ratings

Updated

The brief idea of the algorithm is that every image's pixel is set to black if its brightness is T percent lower than the average brightness of the surrounding pixels in the window of the specified size, otherwise it is set to white. The detail description of the algorithm is available at:
The advantage of this method is that the binarized images are subjectively almost as good as from Sauvola's method but the calculation is two times faster than Sauvola's method. Sauvola’s method calculates local mean and local variance, while Bradley’s method calculates just local mean. And because the variance can be calculated using following formula for variance:

Var(X) = E(X^2) – [E(X)]^2,

the calculation of variance reuses the result from the calculation of the local mean (E(X))^2 and just calculates E(X^2). And that takes the same amount of time as calculation of the local mean. Since calculation of local mean and variance is the most time consuming operation performed by these two methods, Bradley’s method is effectively two times faster than Sauvola’s method. A brief comparison of Bradley’s and Sauvola’s method is available at a blog:

The calculation of the local mean is performed with integral image method in constant time regardless of the kernel size.

Examples of real word applications are: bar-code scanners, license plate registration.

Max Fedorchuck

### Max Fedorchuck (view profile)

Thanks for sharing!

Tolentee

### Tolentee (view profile)

Does the code apply to 3D images??

Mandar

Gabor

### Gabor (view profile)

Thank you so much!
Great work. It's very appreciated.

Henry Kricancic

### Henry Kricancic (view profile)

Hi Jan, is this function based on this paper -- "Adaptive thresholding using the integral image" from 2007 by Bradley and Roth?

Cheers

Jan Motl

### Jan Motl (view profile)

Hi, Gunshi. You are right. I fixed the code. And thanks for letting me know!

gunshi gupta

gunshi gupta

### gunshi gupta (view profile)

Hi, this code is really useful, along with all the other binarisation methods!
I just had one doubt though, in the bradley default parameters, which is supposed to be 3 in length, the list is as follows:
optargs={[15 15] 10 0 'replicate'}
this seems like its of length four to me, am i missing something?
or is the T value supposed to be a list as such [10,0]?
Thanks.

Michael scheinfeild

### Michael scheinfeild (view profile)

excellent code help binarize image

haem

Jan Motl

### Jan Motl (view profile)

You can try to increase the size of the sliding window. Don't be afraid to increase the windows size to hundreds to hundreds pixels - the window should be bigger than the biggest symbol in the image.

Or you can try the post-processing step from Yanowitz and Bruckstein's binarization method: http://www.mathworks.com/matlabcentral/fileexchange/41786-remove-ghosts-from-binarized-images

Basically it looks at the edge of each foreground object. And if the average gradient along the edge is small, it concludes that it's false foreground object and deletes the object.

Max Neudecker

### Max Neudecker (view profile)

This method works quite well for binarizing my samples!
Note: I found that, depending on the grayvalue variations, a lot of background pixels are wrongly segmented as foreground. However, masking with another binary image, binarized e.g. with Otsu's threshold (function graythresh), remedied this.