Abstract: Flotation efficiency is closely related to the interaction mechanism between bubbles and particles within the slurry. Existing visual feature analysis of the flotation process mostly focuses on the froth layer, which is susceptible to environmental interference and pays insufficient attention to the cross-scale interaction between bubbles and particles inside the slurry during actual industrial flotation. To address these issues, a deep learning-based cross-scale recognition system for bubbles and particles in flotation slurry was developed. The system acquires slurry images via a slurry image acquisition device and constructs a standard dataset for training and evaluation. Based on the Mask R-CNN model and Feature Pyramid Network (FPN), the system achieves instance segmentation and multi-scale feature fusion, and the COCO evaluation metrics are employed to quantitatively evaluate the segmentation performance. Feature parameters of bubbles and particles are automatically extracted from the segmentation results, and detection errors are quantitatively evaluated by comparing the results with manual annotations. Industrial verification was conducted in the flotation cells of a coal flotation workshop in Tangshan, Hebei. The results indicate that the system can stably capture clear slurry images and accurately identify large-scale bubbles and small-scale mineral particles. The overall average precision (AP) for instance segmentation reached 75.1%, the average precision for large objects（AP_L） reached 78.8%, the relative error for bubble equivalent diameter measurement was 0.299%, and the relative error for the median particle size was 2.680%. The system achieves precise cross-scale segmentation and feature quantification of bubbles and particles within the slurry, providing reliable data support for the online monitoring and intelligent optimization of the flotation process, which demonstrates promising prospects for industrial application.