A CNN-based regression framework for estimating coal ash content on microscopic images

Coal ash content is an important criterion for evaluating coal quality. In recent years, the online ash measurement approach based on a convolutional neural network (CNN) has gotten a lot of attention. However, learning continuous targets from a small and unbalanced dataset is one of the biggest challenges for ash content estimation using CNN. In this paper, a CNN-based regression framework was proposed for rapidly estimating the ash content of coal. Firstly, data synthesis was performed to augment the limited dataset, and label distribution smoothing (LDS) was employed to alleviate the imbalance in datasets. Secondly, separable convolution (SC) and attention modules were introduced into multi-branch (MB) blocks of the backbone. SC was applied to fuse both spatial and channel-wise information, and attention modules were used to enhance feature extraction capability. Finally, as a final estimation value, the regression head outputted a float in the range [0, 100]. The results showed that the proposed approach achieved 0.31% error on the 1,145 test images, where 81.76% had a margin of error less than 0.5% and 96.25% less than 1.0%. Furthermore, the prediction error analysis revealed that the accuracy of the predictions was highly related to the homogeneity of the materials. The visualization results demonstrated that the proposed regression framework could merge multi-scale information and that the synthetic dataset was viable.