Neural art.io (2)

11
neuralArt.io Large Scale Classification of Fine Art Paintings Mike Osorio Galvanize DSI

Transcript of Neural art.io (2)

Page 1: Neural art.io (2)

neuralArt.ioLarge Scale Classification

of Fine Art Paintings

Mike OsorioGalvanize DSI

Page 2: Neural art.io (2)

Motivation

● Build an Intelligent Machine to learn and judge complex visual concepts and stylistic elements of art.

● Develop a ML model that is capable of classifying a painting’s genre and its artist.

Page 3: Neural art.io (2)

Data Acquisition● Scraped 2000+ high resolution fine art painting images from a Picasa Web Album.● Chose Three Artists● Built a Genre Classifier to determine whether or not a painting is a portrait● Built an Artist Classifier to determine if painted by Van Gogh, Joseph Mallord Turner, or Cezanne

Joseph Mallord Turner

Van Gogh Cezanne

Page 4: Neural art.io (2)

Modeling Process

Joseph Mallord TurnerFishermen in a squall, c 1802

Patching,Dominant Colors,

Grayscale,Denoising,

Edge Detection

Patching

SVC, KNN, Random Forest,

Gradient Boosting

Neural Networks

Page 5: Neural art.io (2)

Feature Engineering● Resize our images to 480 x 480 pixels and extract 30 random 80 X 96 patches of the painting.● Determine Dominant Colors of the patches by performing k means clustering on the image data.● Perform the following image transformations on the patch images:

Grayscale + Total Variation Denoising

Canny Edge DetectionRaw Pixels

Page 6: Neural art.io (2)

Patch Extraction

Joseph Mallord TurnerView the High Street, c 1834

Page 7: Neural art.io (2)

Modeling Process

Joseph Mallord TurnerFishermen in a squall, c 1802

Patching,Dominant Colors,

Grayscale,Denoising,

Edge Detection

Patching

SVC, KNN, Random Forest,

Gradient Boosting

Neural Networks

68 % Accuracy

75 % Accuracy

Page 8: Neural art.io (2)

MVP System for Classification of Art Painting

Raw

Vis

ual F

eatu

res

Imag

e

Res

izin

g

Pat

ch

Ext

ract

ion

GenreClassifier

(Neural Network)

ArtistClassifier

(Neural Network)

Genre:Portrait

Artist:Van Gogh

CezanneHarlequin, c 1888-1890

Page 9: Neural art.io (2)

Feature Learning

xn

xn-1

x1

x2

Van Gogh

Turner

...

...

...

...

Input Layer

Hidden Layer 1

(512 nodes)

Hidden Layer 2

(512 nodes)

HiddenLayer 3

(512 nodes)

Output Layer

Cezanne

Page 11: Neural art.io (2)

Further Work

● Train a large, deep convolutional neural network similar to Alex Krizhevksy’s Alexnet Model.o Expected to give more accurate classifications and allow us to classify

more artists, genres, and themes.o With a more robust classifier, we will have the capability to scale the

app to a phone app (Shazam of Art).

● Experiment with other type of image transformation technique such as wavelets.