<p>Human pose estimation in coal mining environments faces challenges from pervasive coal dust interference and non-uniform illumination conditions, which create composite noise patterns that degrade pose estimation accuracy. Although existing methods have improved robustness through deep learning architectures, their substantial computational demands hinder practical deployment in mining scenarios. To address these challenges, we propose SLIC-POSE, a novel Self-Learning Image-Conditioned human pose estimation network specifically designed for coal mining environments. First, we design a Differentiable Self-learning Image Restoration Module (DSIRM) that operates as a learnable preprocessing stage, dynamically rectifying photometric and structural degradation without requiring paired training data. DSIRM employs a Context-Aware Parameter Learner and a Differentiable Restoration Layer to enhance images. It sequentially performs self-adaptive de-dusting, self-calibrating illumination, and feature re-projection. These processes are optimized end-to-end using feedback from downstream tasks. Second, we present a Channel Importance-aware Knowledge Distillation (CIKD) framework that systematically evaluates channel significance through group-wise L1-based metrics to guide selective knowledge transfer. By constructing adaptive knowledge filters that emphasize high-value channels while attenuating low-contribution features, CIKD achieves efficient model compression without sacrificing essential representation capabilities. Experimental results demonstrate that SLIC-POSE delivers robust performance across both YOLO and OKS metrics. Furthermore, we demonstrate the practical effectiveness of SLIC-POSE via FlexZone, a real-time hazardous area intrusion warning system. This system has been deployed across six diverse mining scenarios and achieves 67 FPS on edge devices, enabling proactive safety management in coal mining.</p>