Feature-Marker Recognition and Dosage Effect Analysis for Coal-Gangue-Backfill Separation

To address the difficulty of separating coal, gangue, and backfill caused by crushed backfill mixing into coal streams during re-mining of room-and-pillar residual areas, this study proposes a feature-marker-enhanced visual recognition method for backfill materials. Three markers, namely red ceramic particles, fluorescent particles, and glass microbeads, were premixed into backfill materials at dosages of 5%, 10%, and 15%. After specimen preparation and crushing, images were collected using a dynamic belt-conveying platform to construct a 7708-image dataset containing four target categories: coal, gangue, ordinary backfill, and marked backfill. A lightweight object detection model integrating RepGhost, imMSAM, and imShapeIoU was used to evaluate backfill recognition under different marker schemes. The results show that the model achieved mAP@0.5, mAP@0.5:0.95, Precision, and Recall values of 0.939, 0.826, 0.923, and 0.941, respectively, on the overall test set. Compared with the unmarked group, the Recall and mAP@0.5 of backfill targets generally increased after marker addition, indicating that the markers enhanced the visual differences between backfill and coal/gangue and reduced missed detections. Under the experimental conditions, the fluorescent particle group with a 15% dosage showed better overall recognition performance, with Precision, Recall, and mAP@0.5 values of 0.997, 1.000, and 0.995, respectively, while the glass microbead group with a 5% dosage performed better in mAP@0.5:0.95. The results provide a reference for marker parameter selection and intelligent recognition in coal-gangue-backfill separation.