Classifying Poses throughout Art History with Transfer Learning

Allam, Sarah-Emanuela (2022) Classifying Poses throughout Art History with Transfer Learning. Bachelor's Thesis, Artificial Intelligence.

Preview	Text Bachelor_Thesis_s3747328 (2).pdf Download (6MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (120kB)

Abstract

With the growing museum digitization of artworks and open application programming interfaces (APIs), automatized tools are needed for organizing fine-art databases. Convolutional Neural Networks (CNNs) are used in image classification tasks as a supervised learning technique. The rise of Deep-CNNs, more robust to the complex features of Fine Art images, has been facilitated by Transfer Learning techniques. This thesis attempts to solve an art historical pose classification problem on six different pose classes from a self-made database consisting of 1000 artworks across eight different museum APIs. The classification is done with three CNN architectures (ResNet50, Xception and VGG16) with pre-trained weights from the ImageNet database. Due to data scarcity and class imbalances of the initial poses database, several data augmentation techniques were applied. It was determined that the model could distinguish between the six classes, with the most promising one being the Fine-tuned and pre-trained Xception architecture with SMOTE oversampling of the minority classes and imgaug techniques on the training dataset, which obtained a mean test accuracy of 0.85 and an F1-score of 0.83. A qualitative analysis of the feature maps suggests what formal elements are learned by the CNN in the well-performing classes compared to the under-performing ones.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Sabatelli, M.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	09 Mar 2022 13:44
Last Modified:	09 Mar 2022 13:44
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/26673

Actions (login required)

View Item