<p>The real-time detection of foreign objects (e.g., anchor rods and industrial waste) on the conveyor belts in underground coal mines is challenging. The coal occlusion and blurry object boundaries, together with the limited feature extraction and insufficient real-time speed, hinder the practical deployment. In this paper, an extended reparameterization feature calibration method (ECNet) is proposed to address these issues. First, the extended re-parameterized feature extraction residual module (ERF-RepC3) was designed in the backbone network. It combines the local and semantic residuals with re-parameterization to expand the receptive field and improve the feature extraction efficiency to address the mismatch between the receptive field and semantic level. Moreover, this paper presents a high and low frequency attention intra-scale feature interaction (HLAIFI). By decoupling self-attention heads to focus on the specific frequencies, it eliminates the redundant computations and enhances the efficiency of foreign object detection. In addition, we propose a spatial calibration feature fusion network (SCFFN) that calibrates the multi-scale features via foreground-aware spatial alignment. It also incorporates the dynamic interpolation for the adaptive feature fusion, improving the robustness in the densely occluded scenes. Finally, the performance of the model is validated in CUMT-BeIT, an open dataset of coal mine underground foreign objects. The experimental results demonstrate that ECNet achieves a 3.4% improvement in accuracy over baseline models while maintaining comparable parameter counts and reducing computational demands, with detection speeds reaching 84.74 FPS. Compared to current mainstream object detection algorithms (such as DEIM-D-FINE-L, YOLOv13-l, and RT-DETR-v2), ECNet demonstrates superior performance with at least a 0.3% improvement in accuracy, providing robust safeguards for coal mine safety.</p>