RDT-FragNet: A DCN-Transformer network for intelligent rock fragment recognition and particle size distribution acquisition

Abstract

Accurately and promptly identifying rock fragments and particle size distribution after blasting is crucial for rock transportation and aggregate control in hydraulic and hydropower engineering. Manual screening and traditional edge detection methods suffer from subjectivity and inefficiency, resulting in considerable processing time. Images of rock fragments post-blasting, captured in open-air conditions, present challenges due to overlapping fragments, complicating intelligent recognition. To address this, an instance segmentation model, RDT-FragNet, is designed for rock fragment segmentation. RDT-FragNet is a hybrid model that integrates the Deformable Convolutional Network (DCN) and the Transformer Attention Mechanism (TAM). The DCN-Transformer structure adaptively preserves global and local features, enhancing the segmentation and recognition of rock fragment edges. Comparative analyses and rigorous ablation studies demonstrate RDT-FragNet’s competitive advantages. RDT-FragNet outperforms other advanced models in both quantitative metrics and visual results. The visualization results and the characteristic and maximum particle size of rock fragments closely match the actual situation. The robustness and applicability of the RDT-FragNet model are validated using images from two additional engineering projects. This research introduces an intelligent, efficient, and objective method for rock fragment analysis in open-air settings.