[Google] Google Arts and Culture

Ref: https://artsandculture.google.com/

 

 

• Using Machine Learning for Identification of Art Paintings

Alexander Blessing

Published 2010

 特征点 + using a multi-class SVM with state-of-the-art features.

 

• Learning to automatically recognize artists by their artworks Master thesis

August 2018

The train set contains 17.026 images, the validation set contains 2.436 images, and the test set contains 4.850 images.

As follows from Table 3, the less layers that are fixed, the higher the test performance.

 

• Comparison of Machine Learning Techniques for Artist Identification

Ref: https://www.kaggle.com/c/painter-by-numbers

Narrowed down to 15 artists with at least 450 paintings each for a total of 7462 paintings

 

• Art Painting Identification using Convolutional Neural Network

And then, these downloaded 100 images are randomly distorted by various operations, such as projection, rotation, scaling, etc., to simulate situation when the regions of the interest of the art paintings would be captured from videos or photographs.

Each image has 300 randomly distorted versions, thus, for 100 art painting images we generated 30000 distorted images.

We use 25000 of the total distorted images for training and reserve 5000 for testing purposes. All images are resized to width=256 pixels and height=256 pixels. In the following, we will describe about what kind of image distortions and corresponding distortion strength that we used to generate distorted images.

使用了形变。

错误率对比：2% of cnn, and other 16.8%左右