Automated characterization of coal image samples using pretrained deep convolutional neural networks and transfer learning: A comparative Study

Quality of coal are usually determined by manually examining samples under an optical microscope, and then conducting complex chemical tests for further grading. However, such methods are expensive, time–consuming, and dependent on the competency and expertise of the concerned petrologist. Preliminary quality assessment of coal can be performed with the help of image analysis and artificial intelligence–based techniques identifying the presence of coal maceral constituents. The usage of image processing–based methods for intelligent coal gradation is simpler, faster, and less expensive, and it eliminates the subjectivity of petrologists. However, the performance of traditional image processing–based coal gradation methods, on the other hand, is dependent on the accuracy of segmentation and feature extraction methods. Deep learning models, which do not require image segmentation or manual feature engineering, are an alternative solution to the above classification problem. Pretrained deep neural networks eliminates the time–consuming process of determining the best combination of layer architecture, which would be required by standard Convolutional Neural Network models. As a result, the proposed research presents an automated coal gradation system that utilizes pretrained deep convolutional neural networks to classify coal microscopic images. Six different types of pretrained neural networks viz. VGG19, VGG16, GoogLeNet, AlexNet, ResNet50, ResNet18 are simulated over 1700 Australian coal optical microscopic images of three grades. For the considered coal data set, all of the pretrained models achieve classification accuracy of over 80%.