An Efficient Segmentation Model With Multipath Attention Mechanism Enabling Particle Size Characterization of Coal Dust

Characterizing the coal particle size distribution is an effective means to reduce potential safety hazards in smart mine construction for disaster prevention and security management. Aiming at the existing approach, pixelwise manual labeling on low-contrast particle images that have irregular contours is challenging and time consuming. Thus, a novel deep learning model based on BlendMask instance segmentation is proposed for investigating the detail characteristic of complex coal particles. First, a nonparametric attention mechanism is merged into the backbone network to enhance the detail feature attention of coal dust particles. Second, a bottom-up path is added within the feature fusion structure to reduce the loss of semantic feature information for small target particles. Meanwhile, a compact dual-attention structure is built in the bottom module to improve the edge feature discrimination of coal particles and enhance the segmentation accuracy of particle adhesion regions. Finally, depthwise separable convolutions are applied to balance the computational burden brought by the improved modules. Experimental results show that the proposed model performs better than other segmentation models in terms of pixel accuracy, intersection ratio, precision, and recall for different particle sizes. In the particle size characterization evaluation, the average relative error of the particle size distribution is less than 5% for three particle size ranges, which achieves desirable tradeoffs between computational speed and segmentation accuracy.