Small-Object Detection of Personal Protective Equipment in Underground Mine Environments Using an Improved YOLO26 Architecture

Reliable detection of Personal Protective Equipment (PPE) in underground mines is a key enabler of automated safety monitoring, and it is also one of the hardest small-object problems in industrial computer vision: helmets, ear-defenders, dust masks, eye-shields and reflective vests occupy only a tiny fraction of the image and are degraded by low illumination, dust and self-occlusion. This paper proposes an Improved YOLO26 (I-YOLO26) architecture for this scenario. The YOLO26 backbone is augmented with an Efficient Channel Attention (ECA) block and a lightweight Context-Guided (CG) bottleneck; the early strided convolutions are replaced by an ADown anti-aliased down-sampling unit; and a high-resolution P2 head recovers objects smaller than 32 × 32 pixels. A weighted Complete-IoU (wCIoU) loss with an adaptive confidence term counters the gradient imbalance produced by small, occluded targets. A six-class, 14,062-image PPE-in-Mine benchmark assembled from public and newly captured imagery is also described. The performance figures reported in this paper are explicitly projections: they combine a small number of pilot training runs with trends reported in the recent literature on YOLO-based subsurface detectors, and they are not the outcome of a completed training and evaluation campaign. On that basis, I-YOLO26 is projected to reach mAP@0.5 of 92.7% and mAP@0.5:0.95 of 71.4%, to improve small-object accuracy (APs) by 4.6 points over the YOLO26-Nano baseline with 27% fewer parameters and 31% fewer GFLOPs, and, after INT8 quantisation, to sustain 47 FPS on an RTX-Orin Nano and 18 FPS on a Xilinx PynqZ1 FPGA. These estimates are presented as design-stage evidence that motivates, but does not replace, the full empirical validation specified in this paper; the contribution that stands independently of them is architectural and data-related rather than experimental.

@article{Saidani2026,
  title     = {Small-Object Detection of Personal Protective Equipment in Underground Mine Environments Using an Improved YOLO26 Architecture},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {9},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Taoufik Saidani and Refka Ghodhbani and Yahia Said},
  doi       = {10.14569/IJACSA.2026.0170986},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170986}
}