STRATUM-Seg: An Instance Segmentation Network Exploring Symmetry and Asymmetry in Feature Representation for Coal–Gangue Sorting

Accurate coal–gangue instance segmentation remains challenging because visually similar materials, irregular scale variation, and densely adjacent boundaries must be handled simultaneously. To address these coupled problems, this paper proposes STRATUM-Seg, a compact network based on the nano instance-segmentation variant of You Only Look Once version 11 (YOLO11n-seg), which coordinates parameter-sharing symmetry with heterogeneous and direction-selective feature processing. In the backbone, the Shared Dilated Pyramid Module applies one kernel across multiple dilation rates to extract multi-receptive-field texture features with limited parameter redundancy. In the neck, the Multi-Kernel Focus Fusion Neck aligns three pyramid levels, re-injects backbone features, and performs two-stage aggregation using heterogeneous depth-wise kernels. In the prediction head, the Difference-Enhanced Convolutional Head combines a shared trunk with direction-selective difference convolutions and inference-time re-parameterization to improve boundary representation. Experiments on the Wangjialing subset of the Dataset for Coal, Gangue, and Foreign Objects (DsCGF) show that STRATUM-Seg increases mask mean average precision (mAP) over intersection-over-union (IoU) thresholds from 0.50 to 0.95 from 0.615 to 0.668 over three independent runs while maintaining a compact deployment-form model and real-time 32-bit floating-point (FP32) end-to-end inference on the evaluated NVIDIA RTX 4090 platform.