Airborne dust plumes are difficult to localize in underground mine-face video because the scene is dark, illumination moves with machinery, and dust is confused with lamp bloom, reflective steel, and water-spray aerosol. We present DustVeil, a label-free two-stage system for image-space plume localization. Its software teacher combines background-referenced veiling (C1), local texture decay (C2), and absolute dark-channel response (C3) with a probabilistic soft-OR, then applies glare and chroma gates. The teacher returns a dimensionless response map in [0, 1], a binary plume mask and the corresponding image-area ratio; it does not estimate dust concentration, particle-size distribution, respirable exposure, or hazard categories. Teacher outputs from 342 frames in 114 clips/24 sessions supervise a 0.47 M parameter TinyU-Net. Evaluation uses a 144-image synthetic calibration set and a 72-frame real test set drawn from 72 clips in 18 sessions, with all roles separated at clip and session levels. Thresholds are selected only on synthetic masks and frozen before real scoring. After replacing per-image score normalization with fixed baseline-normal calibration and using reference implementations of the anomaly methods, DustVeil obtains IoU/F1 of 0.366/0.500 and the lowest clean-frame false-positive area (3.7% versus 13.9–59.9%). A separate water-spray set quantifies visual specificity. TinyU-Net runs at 610 FPS for network-only inference and 233 FPS aggregate in the measured six-stream decode-to-mask pipeline; optical flow is excluded from these figures. The validated scope is six fixed visible-light RGB cameras with camera-specific unlabelled calibration at one site, rather than concentration monitoring or camera-disjoint deployment.

We study lightweight approximation of a fixed engineering enhancement operator while limiting changes to usable image regions. Training only on corrected images leaves the pipeline’s intermediate intervention decisions unused. We propose MinePD-Net, which jointly distills corrected RGB outputs and native glare and dark-noise masks from a fixed, non-neural teacher. Predicted masks gate a bounded residual over an input identity path, limiting correction amplitude where intervention scores are low. A region-balanced objective assigns teacher reconstruction and raw-image identity to disjoint teacher-effect and clean-core regions, while retaining the trade-off between native-mask fidelity and transition reconstruction. This design couples decision transfer with correction control in a compact depthwise-separable encoder–decoder. Evaluation on UCM-WFE uses 4888 paired 2688 × 1520 frames from four fixed cameras at a single underground working face, source-clip-separated testing, repeated training, and leave-one-camera-out validation. Relative to a backbone-matched direct regressor, MinePD-Net improves reference peak signal-to-noise ratio by 0.33 dB and reduces clean-core identity error and flow-aligned temporal error by 36.1% and 14.8%, respectively. At native resolution, the 25,600-parameter model reaches 26.68 dB reference PSNR and 0.9234 SSIM. Network-only FP16 inference delivers 26.2 fps on an RTX 4070 Ti (batch size one), while the measured single-stream gateway-to-detection path delivers 17.90 fps with 28.40% dropped frames at a 25 fps input rate. A frozen equipment detector evaluated against human annotations improves from 0.446 to 0.472 mean average precision across intersection-over-union thresholds of 0.50–0.95, with a paired gain of +0.026 (95% CI: +0.009 to +0.043). Reference and mask scores quantify fidelity to the teacher’s outputs and decisions, not recovery of independently known clean scenes. Input preservation and human-annotated detection assess complementary benefits.