The LBP tests the relation between pixel and it's neighbors, encoding this relation into a binary word. This allows detection of patterns/features, while being immune to contrast changes.
Current submission presents pixel-wise implementation, and filtering based implementation, achieving much shorter run-times. Both implementations achieve same results, while running at different passes. Do not use ‘pixelwiseLBP’ unless for educational or debugging purposes.
Current implementation is aligned with "Gray Scale and Rotation Invariant Texture Classification with Local Binary Patterns" from http://www.ee.oulu.fi/mvg/files/pdf/pdf_6.pdf.
This implementation supports multi-color inputs (RGB).
Rotation invariance can be same for all channels, or channel wise manner.
Good evening Amruta.
1) By using rotation invariant spatial LBP you can identify rotated textures as the same texture.
2) Both spatial and frequency domain can be used for texture classification. In case of LBP I'm not familiar with utilization of frequency domain.
3) Scale invariance can be achieved via pyramid of the image, or using changing neighborhood radius. Both will do the trick.
hi Good Afternoon,
Thanks for your valuable comment.
So,1) Can i do rotation invariant texture classification in transform domain?
2)Which domain is more suitable for texture classification ?
3)Can we do feature extraction from dataset(texture images) having different sizes?
Good night Amruta.
As far as I remember- rotation of an image in spatial domain, also rotates the frequency transform image. I assume here that when you say transform domain, you actually mean frequency transom domain.
Texture, in very simplified case, is composed of some repeating pattern, resulting in several spikes in freq. domain. When we perform rotation invariant analysis- we are able to identify textures, even if they are at different directions. You can try visualizing what each texture means in transform domain, by tacking Brodatz or alike images, applying FFT2 with fftshift, spectra=abs(ffthsift(fft2(img))); and presenting imshow( log10(spectra) , ).
You can also try my GUI for Visualization of Gabor filter bank- it is not the same, but related: http://www.mathworks.com/matlabcentral/fileexchange/30107-gabor-filetr-gui
Well I was wrong- the update after Michael bug fix included obsolete file copies, and thus the issue remained. I've resubmitted, hoping it will be fine now.
Thanks for finding the issue.
I'll briefly reply here. if you need additional explanations, please email me. I'm not familiar with the concept circular bi-linear interpolation. Regular bi-linear interpolation is basically about calculating value lying between nearest neighbors (interpolation ) values, via weighted average based on their distances to the desired value position (bi-linear). See First Order Hold as well.
In case of LBP, having a central point, and angle, and a radius, usually results in a point "falling" between pixels. This demands calculation of this value via interpolation. bi-linear is the easiest one. Bi-cubic for example is another, more expensive option.
Now again- in LBP image, each pixel value is calculated via sum of neighboring pixels (of when speaking of the whole image via it's shifted versions). As the shift is not always an integer value it demands interpolation. that's about all. See matlab imresize function, or any bi-linear based image zoom method for visualization.
Hope I've made things a bit more clear.
Hi Nikolay ..
I am very much new to LBP.
I have successfully calculated LBP.
Can you tell me how to calculate circular bi-linear interpolation. I am confused about radius of circular LBP and how circular neighbourhoods are calculated using bilinear interpolation.
Pls help me. I read Ojala's paper , but confused about uniform LBP and circular neighbourhood.
Using 32 neighbors to an image of [nxmxk] dimensions, results in storing an nxmx(kx32) matrix to the memory. Moreover, usage of 32 elements means storing LBP data in uint32, rather then uint8 (for up to 8 neighbors), which means *4 more memory. Altogether, the usage scenario you propose can indeed cause "out of memory" problem.
Currently I do not have a version with fever memory demands, so the only advice I can give you, it to devide the image to patches/windows, and applying the LBP to each such part in turn. You can then concatenate the sub-windows LBP into a single LBP. You can use my image concatenate function for this.
Regarding your grade (3/5)- I find it a bit insulting. Indeed, my implementation fails to deal with the hush scenario of 32 neighbors, especially if the image investigated is of large dimensions.
But you must understand, that some methods are limited, and you cannot apply them to any kind of data expecting it to work. perhaps there is an optimization, besides what I've proposed beforehand, but I have not figured one out so far. You are welcome to propose a better implementation, so I will be able to use and grade it.
If I set the number of neighbors to be 32（r=4）, the program failed with warning 'out of memory' because of the size of a variable was 2.^nNeigh. I wonder if it can be programmed in another way to avoid the memory problem because the variable is sparse.
Hi Nikolay, I am an Mtech student,doing project on LBP metohd.I want your help for interpolation method used in your LBP method.I want formulae of interpolation used in LBP method and therotical information about it.Which interpolation method is used in this given file?
Hi hamad mahmood.
I have not used LBP for face detection, but the topic is highly researched. Here is one of most referenced papers: http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=1717463&url=http%3A%2F%2Fieeexplore.ieee.org%2Fxpls%2Fabs_all.jsp%3Farnumber%3D1717463
I've also participated in a project where Vaioa Jones OpenCV LBP implementtation was used: http://docs.opencv.org/modules/contrib/doc/facerec/facerec_tutorial.html, http://docs.opencv.org/modules/contrib/doc/facerec/facerec_api.html#Ptr<FaceRecognizer> createLBPHFaceRecognizer(int radius, int neighbors, int grid_x, int grid_y, double threshold). You should give it a glance too.
Hi Nikolay,Thanks a lot. Would you be able to email me your nonparametric classifier codes described in your paper at http://www.ee.oulu.fi/mvg/files/pdf/pdf_6.pdf ? I need it for texture recognition and compare with my results. my email address is email@example.com Thanks a lot
Well, the use of the term bi-linear interpolation was a bit misleading. I'm actually doing a calculation we can refer to as bi-linear decomposition. That is, for a real number, I'm trying to find his nearest integer neighbors and their distances to him.
The bi-linear interpolation of those neighbors should result in the number we had in the first place. From my calculation the above statements is true for my formula.
Would you agree with me?
I have calculated the bilinear interpolation referring to expressions on the page (http://en.wikipedia.org/wiki/Bilinear_interpolation).
As you can see, the weight is calculated by the multiplication of two differentials. Would you agree with me?
Thanks for your comment- it seems you've gone intensively through my code.
The filter elements is composed of an average between "ceil" (larger neighbor) and "floor" (smaller neighbor) value of the x and y components, aimed to calculate weight of nearest neighbors from a fractional number. As you have noticed, it is indeed a linear interpolation, implying I could used build in functions, rather than implementing it myself.
Anyhow, I do not see a reason to use multiplication here, and unless I'm being mistaken here the proposed equation seems to be fine.
Best regards, Nikolay.
Thanks for the code.I'm comfused about an equation in generateRadialFilterLBP.m.That is ' radInterpFilt( rowsFloor(iP), colsFloor(iP), iP )= radInterpFilt( rowsFloor(iP), colsFloor(iP), iP )+rowsDistFloor(iP)+colsDistFloor(iP);'. If it is a bilinear interpolation process. I think it should be a multiplication.That is to say, 'radInterpFilt( rowsFloor(iP), colsFloor(iP), iP )= radInterpFilt( rowsFloor(iP), colsFloor(iP), iP )+rowsDistFloor(iP)*colsDistFloor(iP);'. Would you agree with me?
It seems the shift value with which the circshift is called is not an integer number. I do not see this happening when I'm running the demo. Let me email you so we can figure out the problem.
I'd suggest to send your questions to my email, rather then commenting here.
Regarding your question, you should use more then one histogram element. This is called "feature vector". In some cases you can reduce the feature vector length via PCA, and other methods. In case of LBP it was proven that some histogram elements carry more information than others, such as Rotation Invariant and others. I suggest reading relevant publications.
Ok. Thank You.
I have one more question, after we get the histogram feature vector how should I input that into my features matrix? i.e as it is or should I reduce LBP feature vector to a single feature? Please Reply. Thank You
Radial filter is uniform to all directions, and in most cases does the trick. Other shape filters can also be handy (especially if the problem in hand has other geometry shape preferences), but radial shape is the one commonly used in published papers.
Hope this will help a bit
The multiplication is needed for proper weighting of the radial filter elements. In many cases the filter element is between pixels, and therefore a an averaging ought to be done.
While calculating pixel-wise LBP, what is the need to multiply it with radial filter? Because when I read it they are only taking difference of neighbouring pixels from the center pixel. Please explain. Thank You.
Hello naif al harbi.
Well, you can have it, using the "Download submission" button above.
From the relatively low grade I can see you've found the button, but did not like the function... I'd be happy to hear your comments and improvement proposals.
Good evening Nicola Franzoso.
You'r comment is in place- thank you. I plan to add relevant functionality to the submission. The histogram is easy to achieve using "imhist" or "hist" with bins 1:(2^p-1) where p is the number of filter elements, applied to the LBP matrix converted to "single". What prevents me form posting this, it the Rotation Invariant (RI) case, where the histogram becomes very sparse. I failed to find an analytic way (an equation) to predict the indexes of obsolete bins. I can find the indexes of those bins for each case by applying the LBP minimization procedure, but this is inefficient and nasty. So i'm working on it. If you're not using RI-LBP, or you're fine with sparse histogram, please use my advice above.
Unfortunately this function does not compute the LBP correctly.
To fix it you need to change 'sign(currNieghDiff)' to 'currNieghDiff > 0'.
You should also change 'if nNeigh<=8' to 'if nNeigh<=9', as nNeigh = number of neighbours + 1.
Finally in the Primitive pixelwise solution you need to change
neighMat( ceil(nNeigh/2)+ 1 )=false;
i wanted to know how to implement the basic concept of LBP operator with uniform patterns
i have implemented the basic concept of LBP operator in matlab
now i am using the concept of uniform pattern
the problem is that i am not able to understand the basic concept of LBP with uniform patterns
How exactly it has to be done
is there any documentation available which can help me
Hello Fa Fa.
As far as I'm concerned, this comments are here for feedback, errors reporting , ratings, thanks, and code related question. Your question (and similar ones) should be be addressed via other media (directly via email, etc..).
Now, to your issue:
I don't quiet understand the question/. As far as I can see you have an interesting project ahead of you. If by "help" you mean "will I do the project for you", the answer is definitely NO.
If you wish to get some assistance with the topics of your project, that's better, but again I'm not the address in this case. I'd help you if were carrying this project under my supervision, but you aren't. I would help you nevertheless, if I was proficient in the above topics, and had some spare time, but unfortunately this is also net the case.
To conclude, it seems I can not assist you.
Many thanks goes to Chris Forne for his sharp eye. Bugs fixed, and some modification introduced.
28 Aug 2012
A Helix/Snail indexing function was аdded to scan neighborhood pixels as spiral.
09 Jan 2014
- Support Circular filter, as in M. Pietikäinen papers.
- Works on multiple color channels (gray-scale, RGB, La*b and even spectral images).
- Support Rotation Invariant mode.
- Two LBP version- pixels wise, and efficient.
16 Jan 2014
Changed filter direction (to CCW), starting point (3 o'clock instead of 12), support pixels interpolation.
22 May 2014
The function “roundnS” was added to the submission aimed to replace the “roundn” function that is apparently part of the Matlab Mapping Toolbox, which may be missing from some of the users trying to run this code.
05 Sep 2014
Resolved the bug reported by Michael.
06 Jan 2015
Removed duplicated outdated files. thanks to Olga for finding this.
10 Jan 2015
Resolved a few bugs related to circishift (found by Michael and Olga) and sort. Added if changing function behavior based on Matlab version.