Abstract: Aiming at the problems of low accuracy and high model complexity of small target detection algorithms for mine personnel safety helmets in complex mine environments, an improved detection algorithm based on YOLO11n is designed. First, a cross-stage partial-partial multi-scale feature aggregation (CSP-PMSFA) module is designed. This module uses 3×3, 5×5, and 7×7 convolution kernels of different sizes to extract features in parallel, and realizes effective fusion of multi-scale features through a partial selective connection mechanism, enhancing the ability to capture detailed features of safety helmets at multiple scales. Second, a small object enhance pyramid (SOEP) module is designed in the Neck layer. Combining the space-to-depth convolution (SPDConv) spatial feature extraction technology with cross-stage partial multi-branch convolution (CSP-OmniKernel), it optimizes the feature transmission path of the feature pyramid network, effectively reducing information loss of small targets during feature fusion, and enhancing the detection effect of long-distance and low-resolution targets. Meanwhile, the detection layer structure was further adjusted by adding a high-resolution P₂ small-target detection layer and removing the redundant P₅ large-target detection layer, concentrating the computing resources of model more on small-target detection tasks. Finally, a lightweight shared convolution detection head (LSCD) is introduced. Group normalization is used instead of batch normalization to improve the stability of small-batch training, and computing overhead is reduced by sharing convolution parameters. The results show that the average precision mean (mAP@0.5) of the improved model on the self-built dataset reaches 83.6%, which is 4.4% higher than that of the original YOLO11n, with the parameter count reduced to 2.474×10⁶ and the detection speed reaching 201.2 frames per second (fps). Compared with detection algorithms such as YOLOv5n, YOLOv8n, and YOLOv10n, the improved model performs better in accuracy and lightweight design, while balancing real-time performance, achieving the best comprehensive detection performance. The performance of the improved model is verified using the public dataset DsLMF+, and the average precision mean (mAP@0.5) of the model reaches 96.2%, which is higher than algorithms such as YOLOv5n, YOLOv8n, and YOLO11n, further verifying its good generalization ability and providing effective support for the deployment of underground edge devices.