How to use local binary pattern to extract features?

I am new to MATLAB. Kindly help me how to use local binary pattern to extract features from images?

Answer 1

[Image Analyst]

See my attached demo to compute the LBP. It doesn't extract any features from it - it merely computes it. Somehow you'd then have to analyze the LBP to get some feature that you're interested in.

[Usman]

Thanks. any idea how to extract features?

[Image Analyst]

Thresholding, etc. It really depends on what features you want to find. There is no one universal set of "features" that applies to all images ever generated in the history of the universe. What do you want to find and measure in your image?

[Usman]

like uncovered lips and covered lips

[Image Analyst]

I don't know. I probably would not use LBP for that.

[Usman]

so what method you actually use, can you help with that?

[Image Analyst]

I don't know what the current best approach is. And it might depend on your field of view and how much contrast the lips have to the face. I refer you to the latest literature on it because I am not a researcher in that field of image processing: http://www.visionbib.com/bibliography/contentspeople.html#Face%20Recognition,%20Detection,%20Tracking,%20Gesture%20Recognition,%20Fingerprints,%20Biometrics

[Stefan Karlsson]

Last I checked, LBPs where the best performing low-level features for facial, and facial parts tracking... as far as the Viola-jones cascaded ada-boost framework goes. I am not very updated though. The LBP version found in OpenCV out-performs the previous version in both efficiency and accuracy. The old version, however, was just Haar-like features, so this is not suprsing. For state of the art today, I guess deep-learning should, or will soon dominate here as well. That approach leaves feature selection in the hands of the framework.

When i used LBPs in OpenCV, the documentation mentioned their rotation invariant properties, but the actual implementation was not rotation invariant. Perhaps this has changed too now. Rotation invariance for anything with facial tracking is usually important. Scale invariance is built-into the VJ classifier, so you dont need that in the features.... good luck!