Texture Recognition and Synthesis A Nonparametric MultiScale Statistical

Texture Recognition and Synthesis A Non-parametric Multi-Scale Statistical Model by De Bonet & Viola Artificial Intelligence Lab MIT Presentation by Pooja

Main Goal 1. Train on example images 2. Recognize novel images 3. Generate new images How?

Markov Random Fields (MRFs) • Based on simple, local interactions • Success in restoration • Weak generative properties – Inability to capture long range interactions

Wavelet Transform • Effective for modeling natural images • Measures the underlying causes of images – Assumption: causes are statistically independent • Coefficients are uncorrelated

Multi-scale Wavelet Techniques • Iterative convolution of bank of filters – Pyramid of low frequency downsampled images • Images are a linear transform of statistically independent causes

Texture Synthesis • Bergen and Heeger – Inverse wavelet transform

Other synthesis failures Find me!

Other synthesis failures (contd) Find me!

Other synthesis failures (contd) Hehehe, find me this time!!

Think! Why/When does synthesis fail? What does it tell us about requirements in a successful synthesis technique?

Objective of Synthesis • Different from the original • Generated by the same underlying stochastic process

Back to Texture Recognition Gaurav’ll do all the synthesis explaining (Thank God!)

Wavelet coefficients not independent • Long edges? • Parent vector of a pixel defined as

Probabilistic model Generation of nearby pixels strongly dependent

Conditional Distributions Estimated as a ratio of Parzen window density estimators :

Cross Entropy (Motivation) • • • Biased coin: p(h) = 0. 75 1 st output: #h = 75, #t = 25 2 nd output: #h = 100, #t = 0 Which is more likely? Which is more typical? Concept of cross entropy (Kullback-Liebler divergence)

Cross Entropy • Viewed as the difference between two expected log likelihoods • Replace integral with monte-carlo sampling • Lowest cross entropy vs. false positives (or negatives)?

Results • Standardized tests – “easy” data sets – Brodatz texture test suite – 100% correct classification • Natural textures – 20 types of natural texture – 87% correct classification (humans: 93%)

Pros & Cons • Pros – Pyramidal dependency – Concept of likely vs. typical – False positives vs. overall low cross entropy • Cons – – – Estimation of conditional distributions? The R(. ) function? Works well on simple texture sets! So? Only 20 natural textures? ? Errors in the paper

And finally…. n n A not so boring slide! Do you think their method would work on this texture? ?