A VGGNet-like approach for classifying and segmenting coal dust particles with overlapping regions

The Automatic extraction of coal particles characteristics has great importance in smart mine construction for security planning and disaster prevention. Traditional approaches, including visual interpretation, which requires manually outlining textures to describe particles, provide unclear texture characteristics due to the complexity of low-contrast grey particle imagery. Thus, a novel feature learning approach with a simplified VGGNet-like network is investigated to learn the characteristic details of complex spatial particle sample image sets. The sample data information includes 2000 particle images of four scenes acquired from a coal preparation plant. First, the particle overlapping regions are located by the detected feature points. Then particle swarms are separated with positioning-labels by their discriminative characteristics. Afterwards, feature classification by fully connected layers and image segmentation with up-sampling module introduced are realized based on improved VGGNet. Furthermore, particle sizes are evaluated over the hybrid particle distribution by characteristic learning. The experimental results demonstrate that, under the proper conditions, improved discrimination performance can be achieved by the proposed approach compared with that of other state-of-the-art approaches. The extraction performance can indeed be an effective reference to determine the particle size distribution (i.e., the granulometric analysis) of the sampled particulate. (c) 2021 Elsevier B.V. All rights reserved.

Automated summary

The study explores a novel feature learning approach using a simplified VGGNet-like network to learn the characteristic details of complex spatial particle sample images, achieving automatic extraction and classification of particle images, as well as evaluation of particle sizes.