Attempt 1
We were able to find a dataset composed of close up images of different types of fabrics. We decided to train a deep learning network to classify the type of fabric given such an input. As a starting point, we decided we want to classify between 3 different types – cotton, polyester, and wool. This is because these materials have the most images in the dataset and so we would get better results compared to other fabrics when training a neural net. We created a 80-10-10 split for each of the three materials for training, validation, and test set. We also converted all the images to grayscale because we believe colour is not a strong or useful feature for determining material. A VGG-16 model architecture was used because in the paper discussing this dataset and results, a VGG-M model was used and this is a larger version of that architecture. Our testing set achieved an overall accuracy of 76%. Class-wise, this was 76% recall for cotton, 89% for polyester, and 39% for wool. Wool had considerably less images compared to the other classes and for our next iteration we will attempt to account for this class imbalance.
Identifybre 