Information Theory and Aesthetics. Application to the Evolution of Van Gogh’s Artwork Pembroke College, Oxford Jaume Rigau, Miquel Feixas and Mateu Sbert Graphics and Imaging Laboratory, University of Girona, Spain gilab.udg.edu Imaging GILab’s logo Laboratory Graphics & March 2015 Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Preliminary Painter Imaging GILab’s logo I Paintings by Vincent van Gogh (1853-1890), public domain I Full set of 800 paintings I Digital photographic reproductions by David Brooks, The Vincent van c 1996-2008 David Brooks Gogh Gallery Laboratory Graphics & Self-Portrait with Bandaged Ear and Pipe, 1889 Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Preliminary Imaging GILab’s logo Preliminary Informational Aesthetic Measures Previous Work Basic Tools Three Aesthetic Measures Informational Analysis Van Gogh Periods Color Mutual Information Definition Van Gogh Periods Entropy-Rate Definition Van Gogh Periods Conclusions Acknowledgments More Info... Pink Peach Tree in Blossom, 1888 Laboratory Graphics & Outline Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Aesthetic Measures Previous Work Birkhoff’s Aesthetic Measure The origins George David Birkhoff, 1884-1944 Photograph by Bachrach c 2008 National Academy of Sciences Definition I [Birkhoff 33] Ratio order and complexity I O C Impossibility of comparing objects of different classes I The aesthetic experience depends on each observer Imaging GILab’s logo Laboratory Graphics & M= Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Aesthetic Measures Previous Work Bense’s Model The birth of Informational Aesthetics Max Bense, 1910-1990 c 2006 Universität Stuttgart Artistic process [Bense 69] I A repertoire of elements is transmitted to the final product I The creative process is a selective process I Order is produced from disorder 1. Initial repertoire: range of colors 2. Palette: selected repertoire Imaging GILab’s logo Laboratory Graphics & 3. Spatial distribution: palette → physical support Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Aesthetic Measures Previous Work Bense’s Model The birth of Informational Aesthetics Max Bense, 1910-1990 c 2006 Universität Stuttgart Artist representation to code Imaging GILab’s logo Rep1 Laboratory Graphics & Description Product Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Aesthetic Measures Previous Work Bense’s Model The birth of Informational Aesthetics Max Bense, 1910-1990 c 2006 Universität Stuttgart Artist representation to code Product recognition to decode Observer Rep1 ∩ Rep2 Description Rep1 Rep2 Evaluation GILab’s logo Laboratory Graphics & Art is communication! Imaging Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Aesthetic Measures Basic Tools Complexity and Order How to measure them? Conceptualizing Birkhoff’s aesthetic measure following Bense’s perspective I Heterogeneity of the palette: Shannon entropy I Information transfer: Mutual information I Descriptive complexity: Kolmogorov complexity Imaging GILab’s logo Laboratory Graphics & Informational aesthetics framework Complexity and order interpretations Normalized measures of M Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Aesthetic Measures Basic Tools Shannon Entropy The basis of Information Theory Claude Elwood Shannon, 1916-2001 c 2006-2008 Alcatel-Lucent Information I Discrete stochastic variable: X = {x1 , . . . , xn } I Probability distribution: p(X ) = {p(x1 ), . . . , p(xn )} I Surprise or information: S(xi ) = − log p(xi ) Discrete Shannon entropy [Shannon 48] H(X ) = EX (S(X )) = − X pi log pi bits GILab’s logo Laboratory Graphics & x∈X Imaging Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Aesthetic Measures Basic Tools Information Channel A discrete memoryless channel Loss Imaging GILab’s logo Sender In Out Receiver Laboratory Graphics & X Noise Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Y Informational Aesthetic Measures Basic Tools Information Channel A discrete memoryless channel Loss X Sender Noise In Out p(X ) Receiver p(Y |X ) p(Y ) X →Y Information transfer between X and Y Mutual Information [Shannon 48] I (X , Y ) = XX p(x, y ) log Imaging GILab’s logo Laboratory Graphics & x∈X y ∈Y p(x, y ) bits p(x)p(y ) Dependence or correlation between X and Y Information Theory and Aesthetics Y Info-Metrics Institute 2015 Spring Workshop Informational Aesthetic Measures Basic Tools Information Channel A discrete memoryless channel Loss X Sender Noise In p(X ) Out Receiver p(Y |X ) X →Y Entropy of X conditional on Y Discrete conditional Shannon entropy H(X |Y ) = − XX [Shannon 48] p(x, y ) log p(x|y ) bits Imaging GILab’s logo Laboratory Graphics & y ∈Y x∈X Remaining uncertainty of X known Y Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Y p(Y ) Informational Aesthetic Measures Basic Tools Mutual Information Venn diagram Entropy relation I (X , Y ) = H(X ) − H(X |Y ) = H(Y ) − H(Y |X ) H(X ) = I (X , Y ) + H(X |Y ) H(X ) H(Y ) Imaging GILab’s logo Laboratory Graphics & H(X |Y ) I (X , Y ) H(Y |X ) Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Aesthetic Measures Basic Tools Kolmogorov Complexity A descriptive complexity Andrey Nikolayevich Kolmogorov, 1903-1987 c 1966 Novosti Press Agency K [Kolmogorov, Chaitin, and Solomonoff 60’s] Length of the shortest binary program to compute x: K (x) = minU(p)=x {`(p)} Imaging GILab’s logo Laboratory Graphics & Algorithmic information / randomness Length of the ultimate compressed version of x Non-computable ⇒ approximated by standard real-world compressors Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Aesthetic Measures Three Aesthetic Measures Image Data Color histograms Image I of N pixels Histogram H = f (I, bins) Palette Represented by Tristimulus sRGB color system Alphabet Xrgb = {0, . . . , 255}3 Luminance ITU-R Recommendation BT.709 Alphabet X` = [0, 255] HSV sRGB representation = (hue, saturation, value) Alphabet XHSV = {0, . . . , 360} × {0, . . . , 1}2 Imaging GILab’s logo Laboratory Graphics & Color (C, C ) = [(Xrgb , Xrgb )|(X` , X` )|(XHSV , XHSV )] Region (R, R) Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Aesthetic Measures Three Aesthetic Measures Bense’s Perspective Based on Shannon entropy Redundancy Maximum entropy − Entropy of the palette Hmax − H(C ) C = Xrgb ⇒ n = |Xrgb | = 2563 ⇒ Hmax = log n = 24 C = X` ⇒ n = |X` | = 256 ⇒ Hmax = log n = 8 I Order: redundancy I Complexity: maximum entropy I Interpretation: reduction of uncertainty due to the choice of a given palette Imaging GILab’s logo MB = Hmax − H(C ) Hmax Laboratory Graphics & Aesthetic measure Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Aesthetic Measures Three Aesthetic Measures Kolmogorov’s Perspective Based on algorithmic information Compression ratio I Quantify the reduction in data size produced by a data compression algorithm I More structure or regular patterns ⇒ order↑ ⇒ compression↑ N = number of pixels Imaging GILab’s logo I Order: algorithmic reduction of the data size I Complexity: data size using a constant code I Interpretation: degree of order without any a priori knowledge on the palette MK = NHmax − K (I) NHmax Laboratory Graphics & Aesthetic measure Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Aesthetic Measures Three Aesthetic Measures Channel Perspective Self-Portrait in Front of the Easel, 1888 Wheat Field with Crows, 1890 The creative channel Imaging GILab’s logo Laboratory Graphics & The creative process described by Max Bense (1969) can be understood as the realization of an information channel between the palette and a set of regions of the image. Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Aesthetic Measures Three Aesthetic Measures Channel Perspective p(C ) p(R) C R p(R|C ) Histogram Image Regions Imaging GILab’s logo I C is the input r.v. associated with the set of bins C of H. I R is the output r.v. associated with the set of regions R of I. I p(C ) is the probability of bin c ∈ C. I p(R) is the normalized area of region r ∈ R. I p(R|C ) is the transition probability between c and r . Laboratory Graphics & Image information channel Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Aesthetic Measures Three Aesthetic Measures Channel Perspective p(C ) p(R) C R p(R|C ) Histogram Image Regions Mutual information Shared information or correlation between C and R: I (C , R) = XX Imaging GILab’s logo p(c, r ) p(c)p(r ) Laboratory Graphics & c∈C r ∈R p(c, r ) log Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Aesthetic Measures Three Aesthetic Measures Channel Perspective Looking for the compositional structure Image decomposition [Rigau et al. 05] I A partitioning algorithm builds a region-tree guided by I (C , R) I Discrete information channel C → R I Acquisition of information increases the mutual information and decreases uncertainty: H(C ) = I (C , R)↑ + H(C |R)↓ The tree captures the structure and hierarchy of the image Imaging GILab’s logo Laboratory Graphics & The I (C , R)↑ quantifies the capacity of an image to be ordered or the feasibility of decomposing it by an observer Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Aesthetic Measures Three Aesthetic Measures Channel Perspective Looking for the compositional structure Aesthetic measure I Order: degree of I captured by a decomposition I Complexity: initial uncertainty (entropy) I Ratio of I : I Image complexity: Ms (n) = Ms-1 I (C , R) H(C ) I (C , R) H(C ) =n Imaging GILab’s logo Laboratory Graphics & The number of regions n for a ratio of I is a measure of image complexity Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Aesthetic Measures Three Aesthetic Measures A Simple Composition Imaging GILab’s logo Laboratory Graphics & Potato Planting, 1884 Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Aesthetic Measures Three Aesthetic Measures A Simple Composition Potato Planting, 1884 Imaging GILab’s logo Laboratory Graphics & -1 = 29 regions Ms,0.25 Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Aesthetic Measures Three Aesthetic Measures A Complex Composition Imaging GILab’s logo Laboratory Graphics & Irises, 1889 Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Aesthetic Measures Three Aesthetic Measures A Complex Composition Imaging GILab’s logo Laboratory Graphics & Irises, 1889 -1 Ms,0.25 = 3, 378 regions Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Analysis Van Gogh Periods Average per Period Imaging GILab’s logo # 26 172 209 181 137 75 800 MB 0.422 0.486 0.384 0.351 0.342 0.334 0.388 MK 0.769 0.794 0.712 0.688 0.665 0.659 0.713 Ms-1 (0.25) 1,019 1,145 1,689 1,748 2,331 2,081 1,710 Laboratory Graphics & Period Early Nuenen Paris Arles St-Rémy Auvers Global Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Analysis Van Gogh Periods Average per Period Period Early Nuenen Paris Arles St-Rémy Auvers Global Imaging GILab’s logo MB 0.422 0.486 0.384 0.351 0.342 0.334 0.388 MK 0.769 0.794 0.712 0.688 0.665 0.659 0.713 Ms-1 (0.25) 1,019 1,145 1,689 1,748 2,331 2,081 1,710 Redundancy decrease → Increase color diversity (from Nuenen) Laboratory Graphics & I # 26 172 209 181 137 75 800 Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Analysis Van Gogh Periods Average per Period Imaging GILab’s logo # 26 172 209 181 137 75 800 MB 0.422 0.486 0.384 0.351 0.342 0.334 0.388 MK 0.769 0.794 0.712 0.688 0.665 0.659 0.713 Ms-1 (0.25) 1,019 1,145 1,689 1,748 2,331 2,081 1,710 I Redundancy decrease → Increase color diversity (from Nuenen) I Order decrease → complexity (from Nuenen) Laboratory Graphics & Period Early Nuenen Paris Arles St-Rémy Auvers Global Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Analysis Van Gogh Periods Average per Period Imaging GILab’s logo # 26 172 209 181 137 75 800 MB 0.422 0.486 0.384 0.351 0.342 0.334 0.388 MK 0.769 0.794 0.712 0.688 0.665 0.659 0.713 Ms-1 (0.25) 1,019 1,145 1,689 1,748 2,331 2,081 1,710 I Redundancy decrease → Increase color diversity (from Nuenen) I Order decrease → complexity (from Nuenen) I Compositional complexity increase (before Auvers) Laboratory Graphics & Period Early Nuenen Paris Arles St-Rémy Auvers Global Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Analysis Van Gogh Periods Average per Period Imaging GILab’s logo # 26 172 209 181 137 75 800 MB 0.422 0.486 0.384 0.351 0.342 0.334 0.388 MK 0.769 0.794 0.712 0.688 0.665 0.659 0.713 Ms-1 (0.25) 1,019 1,145 1,689 1,748 2,331 2,081 1,710 I Redundancy decrease → Increase color diversity (from Nuenen) I Order decrease → complexity (from Nuenen) I Compositional complexity increase (before Auvers) Laboratory Graphics & Period Early Nuenen Paris Arles St-Rémy Auvers Global Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Analysis Van Gogh Periods Earliest Paintings Fisherman’s Wife on the Beach, 1882 GILab’s logo Laboratory Graphics & -1 (MB ,MK ,Ms,0.25 )=(0.418, 0.759, 1264) (0.422, 0.769, 1019) Imaging Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Analysis Van Gogh Periods Earliest Paintings Fisherman’s Wife on the Beach, 1882 MB ↑ Limited palette MK ↑ High degree of order Ms-1 ↓ Basic compositions GILab’s logo Laboratory Graphics & -1 (MB ,MK ,Ms,0.25 )=(0.418, 0.759, 1264) (0.422, 0.769, 1019) Imaging Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Analysis Van Gogh Periods Nuenen/Antwerp Shepherd with a Flock of Sheep, 1884 GILab’s logo Laboratory Graphics & -1 (MB ,MK ,Ms,0.25 )=(0.463, 0.739, 875) (0.486, 0.794, 1145) Imaging Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Analysis Van Gogh Periods Nuenen/Antwerp Shepherd with a Flock of Sheep, 1884 MB ↑ Dark palette and dull in tones MK ' Few colors and easy structure Ms-1 ↓ Simple compositions GILab’s logo Laboratory Graphics & -1 (MB ,MK ,Ms,0.25 )=(0.463, 0.739, 875) (0.486, 0.794, 1145) Imaging Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Analysis Van Gogh Periods Paris The Seine with the Pont de la Grande Jette, 1887 GILab’s logo Laboratory Graphics & -1 (MB ,MK ,Ms,0.25 )=(0.385, 0.718, 1396) (0.384, 0.712, 1689) Imaging Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Analysis Van Gogh Periods Paris The Seine with the Pont de la Grande Jette, 1887 MB ↓ From dark-hued to bright colors MK ↓ More details, elements, and colors Ms-1 ↑ Complex compositions GILab’s logo Laboratory Graphics & -1 (MB ,MK ,Ms,0.25 )=(0.385, 0.718, 1396) (0.384, 0.712, 1689) Imaging Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Analysis Van Gogh Periods Arles Sunset: Wheat Fields Near Arles, 1888 GILab’s logo Laboratory Graphics & -1 (MB ,MK ,Ms,0.25 )=(0.345, 0.697, 1648) (0.351, 0.688, 1748) Imaging Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Analysis Van Gogh Periods Arles Sunset: Wheat Fields Near Arles, 1888 MB ↓ Vibrant and saturated colors MK ↓ Exploration of complementary color contrasts Ms-1 ↑ Complex compositions GILab’s logo Laboratory Graphics & -1 (MB ,MK ,Ms,0.25 )=(0.345, 0.697, 1648) (0.351, 0.688, 1748) Imaging Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Analysis Van Gogh Periods Saint-Rémy Olive Grove: Pale Blue Sky, 1889 GILab’s logo Laboratory Graphics & -1 (MB ,MK ,Ms,0.25 )=(0.339, 0.593, 2456) (0.342, 0.665, 2331) Imaging Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Analysis Van Gogh Periods Saint-Rémy Olive Grove: Pale Blue Sky, 1889 MB ↓ Heterogeneous and bright palette MK ↑ Landscapes and swirls Ms-1 ↑ High compositional complexity GILab’s logo Laboratory Graphics & -1 (MB ,MK ,Ms,0.25 )=(0.339, 0.593, 2456) (0.342, 0.665, 2331) Imaging Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Analysis Van Gogh Periods Auvers-sur-Oise Daubigny’s Garden, 1890 GILab’s logo Laboratory Graphics & -1 (MB ,MK ,Ms,0.25 )=(0.315, 0.714, 2375) (0.334, 0.659, 2081) Imaging Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Informational Analysis Van Gogh Periods Auvers-sur-Oise Daubigny’s Garden, 1890 MB ↓ Palette highly contrasted MK ↓ Luminance used in a complex way Ms-1 ' High compositional complexity GILab’s logo Laboratory Graphics & -1 (MB ,MK ,Ms,0.25 )=(0.315, 0.714, 2375) (0.334, 0.659, 2081) Imaging Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Periods 25% Measure Plots Van Gogh Periods Página 1 de 1 Informational Analysis MB (reversed) 25% MK (reversed) 25% Página 1 de 1 Página 1 de 1 -1 Ms,0.15 25% Página 1 de 1 Página 1 de 1 25% Página 1 de 1 -1 Ms,0.20 Imaging GILab’s logo Laboratory Graphics & 25% -1 Ms,0.25 Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Periods 25% Measure Plots Van Gogh Periods Página 1 de 1 Informational Analysis MB (reversed) 25% MK (reversed) Página 1 de 1 25% Página 1 de 1 Origins 1 2 25% Página 1 de 1 Maturity 4 5 6 Página 1 de 1 -1 Ms,0.15 3 Transition 25% Página 1 de 1 -1 Ms,0.20 Imaging GILab’s logo Laboratory Graphics & 25% -1 Ms,0.25 Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Color Mutual Information Definition Color Mutual Information Color information distribution in the creative channel p(C ) p(R) C R p(R|C ) Histogram Image Regions Mutual Information I (C , R) = X p(c) c∈C X r ∈R p(r |c) log p(r |c) p(r ) Imaging GILab’s logo Laboratory Graphics & How is the color information distributed? Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Color Mutual Information Definition Color Mutual Information Color information distribution in the creative channel p(C ) p(R) C R p(R|C ) Histogram Image Regions Mutual Information I (C , R) = X p(c) c∈C X r ∈R p(r |c) log p(r |c) p(r ) Imaging GILab’s logo Laboratory Graphics & How is the color information distributed? Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Color Mutual Information Definition Color Mutual Information Color information distribution in the creative channel p(C ) p(R) C R p(R|C ) Histogram Image Regions Color Mutual Information I (C , R) = X p(c)I (c, R) c∈C I (c, R) = X p(r |c) log r ∈R p(r |c) p(r ) Imaging GILab’s logo Laboratory Graphics & Information or saliency associated with each color Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Color Mutual Information Definition Color Mutual Information Color information distribution in the creative channel p(C ) p(R) C R p(R|C ) Histogram Image Regions Kullback-Leibler distance I (c, R) = KL(p(R|c) k p(R)) KL(p k q) = X p(x) log x∈X Imaging GILab’s logo Laboratory Graphics & I (c, R)↑ High correlation between a color and a region I (c, R)↓ Color distributed uniformly in the image Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop p(x) q(x) Color Mutual Information Van Gogh Periods Earliest Paintings Two Women in the Woods, 1882 Imaging GILab’s logo Laboratory Graphics & -1 (MB ,MK ,Ms,0.25 ,I0.25 )=(0.310, 0.650, 2020, 1.910) Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Color Mutual Information Van Gogh Periods Nuenen/Antwerp Imaging GILab’s logo 1417, 1.291) Laboratory Graphics & The Potato Eaters, 1885 -1 (MB ,MK ,Ms,0.25 ,I0.25 )=(0.575, 0.850, Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Color Mutual Information Van Gogh Periods Paris Imaging GILab’s logo 1.854) Laboratory Graphics & Self-Portrait with Straw Hat, 1887 -1 (MB ,MK ,Ms,0.25 ,I0.25 )=(0.295, 0.726, 1272, Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Color Mutual Information Van Gogh Periods Arles Imaging GILab’s logo 1.905) Laboratory Graphics & Vase with Fifteen Sunflowers, 1888 -1 (MB ,MK ,Ms,0.25 ,I0.25 )=(0.349, 0.581, 2736, Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Color Mutual Information Van Gogh Periods Saint-Rémy Imaging GILab’s logo 1758, 1.937) Laboratory Graphics & Starry Night, 1889 -1 (MB ,MK ,Ms,0.25 ,I0.25 )=(0.322, 0.594, Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Color Mutual Information Van Gogh Periods Auvers-sur-Oise Imaging GILab’s logo 1.966) Laboratory Graphics & Thatched Cottages at Cordeville, 1890 -1 (MB ,MK ,Ms,0.25 ,I0.25 )=(0.312, 0.592, 2095, Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Entropy-Rate Definition Block Entropy Information in streams of data L-Block I Chain of random variables: . . . X-2 X-1 X0 X1 X2 . . . I Block of L consecutive random variables: X L = X1 . . . XL I L-Block probability: p(x L ) L-Block Entropy H(X L ) = − X p(x L ) log p(x L ) Imaging GILab’s logo Laboratory Graphics & x L ∈X L Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Entropy-Rate Definition Block Entropy Information in streams of data L-Block I Chain of random variables: . . . X-2 X-1 X0 X1 X2 . . . I Block of L consecutive random variables: X L = X1 . . . XL I L-Block probability: p(x L ) L-Block Entropy H(X L ) = − X p(x L ) log p(x L ) x L ∈X L L-Block Symbol Entropy Imaging GILab’s logo Laboratory Graphics & hx (L) = H(XL |X L−1 ) ≡ H(X L ) − H(X L−1 ) Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Entropy-Rate Definition Randomness Based on the Shannon entropy Entropy-Rate I Average amount of information (irreducible randomness) per symbol I Uncertainty associated with a given symbol (knowing the preceding symbols) I Optimal algorithmic compression Imaging GILab’s logo H(X L ) L ≡ limL→∞ hx (L) Laboratory Graphics & hx = limL→∞ Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Entropy-Rate Definition Randomness Based on the Shannon entropy Entropy-Rate I Average amount of information (irreducible randomness) per symbol I Uncertainty associated with a given symbol (knowing the preceding symbols) I Optimal algorithmic compression hx = limL→∞ H(X L ) L ≡ limL→∞ hx (L) Imaging GILab’s logo I Average uncertainty surrounding a pixel I Degree of randomness (disorder ' compression difficulty) I Empirical positive correlation with contrast Laboratory Graphics & Aesthetic measure Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Entropy-Rate Van Gogh Periods Entropy-Rate per Period Imaging GILab’s logo x hH x 5.251 4.899 5.564 5.855 5.859 5.925 5.561 hSx s(x) 0.724 0.810 0.709 0.598 0.652 0.531 0.804 x 6.551 6.219 6.802 7.034 6.972 7.116 6.780 I Entropy-rate ≡ contrast I Early → Nuenen: hx ↓ I Nuenen → Auvers: hx ↑ I Optimal compression hVx s(x) 0.443 0.602 0.469 0.305 0.322 0.309 0.502 x 6.802 6.215 6.972 7.272 7.427 7.471 6.996 s(x) 0.452 0.776 0.405 0.286 0.248 0.305 0.547 Laboratory Graphics & Period Name Early Nuenen Paris Arles St-Rémy Auvers Global Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Entropy-Rate Van Gogh Periods Entropy-Rate Example Imaging GILab’s logo Sheaves of Wheat, 1890 I x x C hHSV = (5.718, 7.696, 9.279) and hHSV = (4.156, 7.215, 5.000) B I x hHSV ↑≡ high contrast (i.e., spots in foliage) I x hHSV ↓≡ low contrast (i.e., sheaves and field) Laboratory Graphics & The Grove, 1890 Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Conclusions Summary I Informational aesthetic measures I I I I Characterization of a painting style Study the evolution of a painter Visualize the most informative or salient colors and elements. Future work I I Test with other artists Study of perceptual measures Main contribution Imaging GILab’s logo Laboratory Graphics & Proposal of a set of tools to help to study artist’s work and discover relevant characteristics of a painting (or painter’s style) which could go unnoticed by the observer. Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop Acknowledgments Imaging GILab’s logo I you for your attention I Info-Metrics Institute and PIIP organizers for their invitation I David Brooks, The Vincent van Gogh Gallery, www.vggallery.com I Grant TIN2013-47276-C6-1-R from Spanish Government I 2014 SGR1232 from Catalan Government Laboratory Graphics & Thanks to The Cafe Terrace on the Place du Forum at Night, 1888 Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop More Info... Imaging GILab’s logo Laboratory Graphics & Additional info to be found in... Information Theory and Aesthetics Info-Metrics Institute 2015 Spring Workshop