The aim of this project is to come up with a texture classification scheme which utilizes correlations of feature activations of a pretrained CNN, as described in Neural Style Transfer paper of Gatys et. al.[1].
You can find the detailed explanations in the project report
This method is tested on two texture datasets, Textures under varying Illumination, Pose and Scale (KTH-TIPS) [2] and Describable Textures Dataset (DTD). You need to download at least one of them in order to test the algorithm.
When the download completes, extract the contents to your desired location and point data_dir variable in model_training.py to the extracted folder. You also need to set n_class variable according to the number of classes of the dataset. Then run model_training.py to perform training.
You need to have PyTorch installed and Nvidia CUDA Toolkit enabled in order to run this project.
[1] L. Gatys, A. S. Ecker, and M. Bethge. A Neural Algorithm of Artistic Style. arXiv:1508.06576[cs.CV]. August 2015