An Effective Lightweight Measurement Model to Enable Coal Dust Size Distribution Analysis

This article presents an image segmentation model based on U-Net combined with visual sensing. It is aimed at studying the inherent mechanism of coal dust imagery characteristics and is employed to carry out semantic segmentation of particle images. For the feature maps with more redundant information generated by the conventional method, the parameter quantity of the proposed model can be lessened by Ghost-bottleneck, which substitutes for the original convolution in U-Net. The Ghost-SE bottleneck is introduced in the feature backbone to capture the attention of key feature channels, which is able to improve the learning ability for particle characteristics. Thus, particles with different shapes and sizes are credibly discriminated. Finally, the proposed model is evaluated with the coal dust particle proportion in six dust-producing scenes. The experimental results show that the performance metrics accuracy, recall rate, and
$F1$
scores are increased to 0.8732, 0.8434, and 0.8580, respectively, and the mean error
$\varepsilon $
with
$R < 75~\mu \text{m}$
is 3.526% compared with the reference standard. This indicates that introducing modules in the backbone is indeed able to speed up the convergence of the network and improve the accuracy of coal dust particle segmentation. It can be used as an effective auxiliary measurement of coal dust detection for coal mine safety video monitoring.