Outline • Texture modeling - continued – Julesz ensemble FRAME Model – review • FRAME model – Filtering, random field, and maximum entropy – A well-defined mathematical model for textures by combining filtering and random field models – Maximum entropy is used when constructing the probability distribution on the image space – Minimum entropy is used when selecting filters from a large bank of filters – Together this is called min-max entropy principle 5/29/2016 Visual Perception Modeling 2 FRAME Model – review • Maximum Entropy Distribution – Given the expectations of some functions, the maximum entropy solution for p(x) is – where 1 N p( x; ) exp n n ( x) Z () n 1 N (1 , 2 ,, N ) and Z( ) exp - n n ( x)dx n 1 5/29/2016 Visual Perception Modeling 3 FRAME Model – review • Maximum Entropy – continued – (1 , 2 ,, N ) are determined by the constraints – Gradient ascend to maximize log p ( x; ) dn d log p( x; ) E p ( x; ) [ n ( x)] n dt dn n 1, 2, , N 5/29/2016 Visual Perception Modeling 4 Julesz Ensemble • The original texture modeling question – What features and statistics are characteristics of a texture pattern, so that texture pairs that share the same features and statistics cannot be told apart by pre-attentive human visual perception? --- Julesz, 1962 5/29/2016 Visual Perception Modeling 5 Summary of Existing Texture Features 5/29/2016 Visual Perception Modeling 6 Existing Feature Statistics 5/29/2016 Visual Perception Modeling 7 Most General Feature Statistics 5/29/2016 Visual Perception Modeling 8 Julesz Ensemble – cont. • Definition – Given a set of normalized statistics on lattice h {h( ) :, 1,2, K } a Julesz ensemble W(h) is the limit of the following set as Z2 and H {h} under some boundary conditions W ( H ) {I : h( I ) H} 5/29/2016 Visual Perception Modeling 9 Julesz Ensemble – cont. • Feature selection – A feature can be selected from a large set of features through information gain, or the decrease in entropy 5/29/2016 Visual Perception Modeling 10 Julesz Ensemble – cont. 5/29/2016 Visual Perception Modeling 11 Julesz Ensemble – cont. • Sampling the Julesz ensemble – In the Julesz ensemble, a texture type is defined as all the images sharing the observed statistics and features • It is an inverse problem in order to generate texture images or verify the statistics – The problem is again the dimensionality • If the image size is 256x256 and each pixel can have 8 values, there are 865536 different images – Markov chain Monte-Carlo algorithms 5/29/2016 Visual Perception Modeling 12 Julesz Ensemble – cont. • Given observed feature statistics {H()obs}, we associate an energy with any image I as K ( ) Ε (I) |H I( ) ( z ) H obs ( z) | p 1 z • Then the corresponding Gibbs distribution is q (I) 1 E (I) exp( ) ZT T – The q(I) can be sampled using a Gibbs sampler or other Markov chain Monte-Carlo algorithms 5/29/2016 Visual Perception Modeling 13 Image Synthesis Algorithm • Compute {Hobs} from an observed texture image • Initialize Isyn as any image, and T as T0 • Repeat Randomly pick a pixel v in Isyn Calculate the conditional probability q(Isyn(v)| Isyn(-v)) Choose new Isyn(v) under q(Isyn(v)| Isyn(-v)) Reduce T gradually • Until E(I) < e 5/29/2016 Visual Perception Modeling 14 A Texture Synthesis Example Observed image 5/29/2016 Initial synthesized image Visual Perception Modeling 15 A Texture Synthesis Example Temperature 5/29/2016 Image patch Energy Visual Perception Modeling Conditional probability 16 A Texture Synthesis Example - continued Average spectral histogram error 5/29/2016 Visual Perception Modeling 17 Texture Synthesis Examples - continued Observed image 5/29/2016 Synthesized image Visual Perception Modeling 18 Texture Synthesis Examples - continued Observed image 5/29/2016 Synthesized image Visual Perception Modeling 19 Texture Synthesis Examples - continued Mud image Synthesized image 5/29/2016 Visual Perception Modeling 20 Texture Synthesis Examples - continued Observed image Synthesized image 5/29/2016 Visual Perception Modeling 21 Texture Synthesis Examples - continued Observed image 5/29/2016 Synthesized image Visual Perception Modeling 22 Texture Synthesis Examples - continued Synthesized image Original cheetah skin patch 5/29/2016 Visual Perception Modeling 23 Texture Synthesis Examples - continued Observed image 5/29/2016 Synthesized image Visual Perception Modeling 24 Texture Synthesis Examples - continued Observed image 5/29/2016 Synthesized image Visual Perception Modeling 25 Texture Synthesis Examples - continued Observed image Synthesized image 5/29/2016 Visual Perception Modeling 26 An Synthesis Example for Fun 5/29/2016 Visual Perception Modeling 27 Comparison with Texture Synthesis Method - continued • An example from Heeger and Bergen’s algorithm Cross image 5/29/2016 Heeger and Bergen’s Visual Perception Modeling Our result 28 Julesz Ensemble – cont. • Remarks – The results shown here are based on histograms of filter responses – However, the Julesz ensemble applies to any features/statistics of your choice – You can also define Julesz ensemble for images other than textures 5/29/2016 Visual Perception Modeling 29 Julesz Ensemble – cont. • Applications – This essentially provides a framework to systematically verify the sufficiency of chosen features/statistics • Normally, features/statistics are evaluated empirically. In other words, features are evaluated on a limited number of images 5/29/2016 Visual Perception Modeling 30