Comparison of Machine Learning techniques for painting classification.
Open Access
- Author:
- Rao, Abhishek Krishna
- Graduate Program:
- Electrical Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- July 01, 2015
- Committee Members:
- David Miller, Thesis Advisor/Co-Advisor
- Keywords:
- Machine Learning
Deep Learning
Art
Image classification
Classifiers - Abstract:
- The aim of this thesis is to classify paintings style using machine learning techniques. The data set consists of paintings from WikiArt.org website and the classification metric we use is F1 score. The techniques include standard as well as a new one. First we compare different machine learning techniques where each labeled painting belongs to a particular art style like Abstract, Realism etc. We compare the performance of deep neural networks with traditional machine learning techniques that would require hand crafted feature extraction. The techniques we compare include Histogram with Linear Classifier, SIFT with Gaussian Mixture Model and features from Pre trained Convolutional Neural Networks (CNN) with Linear Classifier. The last method gives the best classification result. Then various experiments are performed on the pre trained CNN to see if they can be improved. We explore the dependency of resolution on the performance of pre trained CNN. We see if the classification performance can be improved using data augmentation. Our experiments indicate no change in results with data augmentation. Finally inspired by the outstanding results from pre trained CNN a new technique called Layered Neural Network (LNN) is created and explored. This is a transfer learning technique that creates features learned from simpler tasks. The working of this classifier is illustrated with some toy data sets. It is initially trained on CalTech 101 dataset, and some other classification tasks. Then it is used to classify the paintings data set. However the performance result on paintings is similar to pre trained CNNs. The LNN shows similarity between tasks and which learnt tasks are more useful.