Classification of single-crystal “diffuse” images

In the x-ray diffraction experiment, we collected about 8 million diffraction images in total, where less then 0.3% (20 thousand) were “good” images that could be used for further scientific study. Even among those “good” images, about 95% was “non-diffuse” image (1b, 1d, 1f), and only 5% (1 thousand) was “diffuse” image (1a, 1c, 1e) that we decided to extract out using machine learning methods in this report. To achieve this goal, we built two models — the diffusion map method and the convolutional neural network (CNN) based on the famous VGG16 pipeline [5][6]. For training and testing, we only used a small portion of those 20 thousand “good” images due to the difficulty to label all of them within limited time. After several failed tried, we finally got a 91.3% classification accuracy for the diffusion map method and a 93.6% accuracy for the CNN model.