Abstract: Identifying blasting dust in open-pit mines is challenging because the dust exhibits abrupt generation, rapid diffusion,non-rigid morphology, and translucency, and the task is further complicated by intense light, smoke plumes, and complex backgrounds,making it difficult for traditional visual detection methods to achieve stable,real-time dust identification. To address these issues,this study proposes an intelligent identification method based on an improved Faster R -CNN. First, keyframe extraction and image preprocessing are performed on field videos,followed by the construction of a motion saliency map via the frame difference method to enhance foreground-background feature separability. Subsequently,a Convolutional Block Attention Module (CBAM) is embedded into the ResNet-50 backbone and integrated with a Feature Pyramid Network (FPN) to construct an improred multl-scale detection fremewore and achieve robust localization of dust targets at different scales. Experiments on a custom blasting dust dataset demonstrate that the proposed method achieves a recall of 86. 49%,a precision of 88. 33%,an F1-score of 0. 87,and an average precision (AP) of 0. 89,outperforming ResNet-50 + FPN,standard ResNet-50, and YOLOv3. Ablation studies confirm that the introduction of FPN and CBAM improves AP by 6% and 3%, respectively. Furthermore,cross- domain generalization tests at the Heidaigou Open-Pit Coal Mine yield an AP of 0. 80 and a recall of 78. 51%,representing improvements of approximately 8. 4% and 7. 9%, respectively, over the original Faster R-CNN, demonstrating the method's cross-scenario adaptability under different mining areas,viewpoints,and background conditions. This research provides effective technical support for the automated identification,process monitoring,and safety early warning of mine blasting dust.