In response to the challenges posed by degraded visual features of coal gangue under complex underground mining environments—which often result in low detection accuracy and deviations in centroid positioning, thereby severely constraining the acquisition of precise meso-scale morphological parameters required for trajectory prediction of non-spherical particles in pneumatic separation—this study introduces a novel method integrating adaptive image enhancement and edge extraction for robust gangue recognition and measurement. A specialized detection network, ZCGNet, is developed by combining the C2fcoal backbone, RFANet feature fusion module, ASFFv10Detect detection head, and the adaptive loss function ADIoU, thereby significantly boosting the overall detection performanceunder harsh visual conditions. Furthermore, an Adaptive Continuous Edge (ACE) framework is proposed to achieve high-precision segmentation and centroid localisation of gangue objects. Experimental evaluations on a proprietary dataset show that ZCGNet achieves a Mean Average Precision at Intersection over Union (IoU) threshold (mAP0.5) of 94.6%, a precision rate of 93.2%, and a processing speed of 67.1 FPS, fully meeting the demands of real-time coal gangue detection. Compared with established models such as YOLOv5n, YOLOv7-tiny, YOLOv8n, CenterNet, and Faster Regions with Convolutional Neural Network (R-CNN), ZCGNet consistently delivers superior detection performance. In terms of edge extraction, ACE surpasses classical algorithms including Canny, Sobel, Prewitt, and Kirsch, providing more complete contour extraction even under low-contrast conditions. It effectively eliminates contour discontinuities and achieves sub-pixel level accuracy in gangue localisation. This study provides critical parametric support for optimizing particle trajectory prediction and precise air ejection in gas-solid two-phase flow separation processes.