A dataset of mineral images and labels for machine learning purposes. The dataset is created from this website http://www.minerals.net/ . The dataset consists of around 4,000 images of mineral specimens with labels. There are numerous labels included such as mineral name, crystal system, chemical groups, rock types and fracture. A complete list of labels is found in mineral.py. To generate this dataset run mineral_net.py and this will scrape all the images and mineral labels as well as generate tfrecords for them. The images are saved in data and the tfrecords are saved in tfrecords. When generating records there is a test train split of 80 percent. Here are a few images for show.
I have also included a few scripts to train and test neural networks in predicting labels from images. This proved an extremely difficult task due to the small dataset size. I made a few attempts to overcome this by using a network pretrained on Cifar however this did not help enough. In my opinion this problem is definitely solvable but it will require a larger training set and better transfer learning techniques. I believe that using a well proven network such as Inception 3 or 4 that is trained on Imagenet would be a good step at getting transfer learning working. I also think that the dataset could be expanded and refined heavily by scraping images from Ebay or other cites that sell minerals. Right now I don't really have the time to tackle these problems though.