Intelligent identification of coal macerals using improved semi-supervised semantic segmentation methods

Recently, the demand for automatic coal maceral identification has gradually received much attention, and hence deep learning has been applied to the identification of coal macerals. However, a large number of labels are necessary for supervised learning, which imposes challenges for automatic coal maceral identification. In this study, the methods for identifying coal macerals were fully reviewed. Considering the limited data and the complexity of annotation, a semi-supervised semantic segmentation model combined with conditional random fields (CRF) algorithm was suggested for pixel-level identification of coal macerals. Initially, a new dataset of coal macerals was established. The dataset contains many different coal maceral images collected from the USA and China, as well as the corresponding labeled images. Then the model was trained through adversarial loss, and the prediction results were evaluated through pixel accuracy (PA) and intersection over union (IoU). The results are compared with other three existing unsupervised image segmentation methods. The semi-supervised model achieved, on average, PA and IoU of 84 % and 74 %, respectively. The results show that semi-supervised semantic segmentation can achieve high-precision identification of coal macerals. The CRF algorithm is then employed on the predictions of the model, and the accuracies for the three coal maceral groups achieved 81 %, 84 %, and 88 %, respectively. Finally, the application results of the model on the testing dataset are discussed to compare the differences between artificial intelligence and manual identification. This study demonstrates that semi-supervised semantic segmentation combined with CRF algorithm can be successfully applied to automatic coal maceral identification.