Hi fqhuy,
  Even though I know it's neither the appropriate nor the polite way to ask this question here, it may be the fastest way to contact with you.

  I'm a machine learning researcher in a online advertising company. Recently, I want to decorate some elements on a background image to generate a banner template this is very similar to your remarkable work entitled "Patternista: Learning Element Style Compatibility and Spatial Composition for Ring-based Layout Decoration".

  I have read your work, but due to my knowledge limitation of CV, I'm confused by the generatiing procedure of texture-based features:
    1. You generate MR8 features of each sliding window and calculate the maximum kurtosis value ,
then pick the maximum of them?
    2. I'm totally confused by the generation of lambda.

  What's more, do you have the implementation of these algorithms proposed in this work? Can you share it with me if you have?
  I would greatly appreciate your help with those questions.
Thank you very much!