How to tune Local Binary Pattern?

Question

I am using LBP on grayscale images for facial expression recognition task. How can I figure out the best parameters (number of neighbour set points, and radius of circle) for LBP which will be used as the input of my CNN?

p.s. Using skimage's LBP implementation

Answer 1

You should probably use a CNN directly for this task, but in general, for any learning task, you should use cross-validation to estimate task performance. scikit-learn has excellent documentation on model selection, cross-validation, and parameter search:

• Cross-validation module
• Model selection tutorial
• Grid search and other strategies

More advanced model selection includes Bayesian parameter search, for which the modules hyperopt and Optuna might be useful.