Objective: For this project I attempted to use convoluted neural networks and transfer learning to train a model to be able to classify a paintings art style. I used a dataset of paintings found on Kaggle. The details can be found at the following link.
https://www.kaggle.com/c/painter-by-numbers/data
Process:
The initial data exploration revealed that not all art styles had enough examples. So I picked 5 art styles that had atleast 1000 samples and that were visually quite different from each other.
Engineered a CNN on top of pretrained model VGG19 at the backend to train on over 3000 painting images labeled by their art style.
Conducted hyper parameter tuning using Talos.
Final model: VGG19 + 4 Dense Layers (Relu) + final layer (Softmax). Achieved test accuracy of 68%.
Project Partner Yi Shuen Lim