A detection method for wearing safety helmets in underground coal mines that integrates wavelet convolution and collaborative multi-scale perception

To address challenges such as low detection accuracy and the susceptibility of small targets to feature loss caused by uneven illumination, pervasive dust, and multiple occlusions in the unstructured environments of underground coal mines, this paper proposes an object detection method that integrates frequency-spatial collaborative perception with a lightweight reconstruction mechanism. Based on the YOLOv11 architecture, the proposed method first incorporates a C3k2_WTConv backbone module utilizing wavelet convolution. By employing multi-resolution analysis, this module decouples illumination noise from texture features, thereby enhancing feature extraction capabilities under low signal-to-noise ratio (SNR) conditions. Subsequently, a lightweight feature fusion network, designated as SlimNeck, is constructed by introducing GSConv and VoVGSCSP modules. This design maintains cross-scale feature interaction while reducing computational redundancy. Furthermore, a Detect_Efficient detection head with shared parameters is employed to optimize inference efficiency. Additionally, transfer learning and Exponential Moving Average (EMA) strategies are adopted to overcome the issue of sample scarcity. Experimental results demonstrate that the proposed method achieves an mAP@0.5 of 97.5% on the DsLMF+Helmet dataset, with an inference time of only 10.2 ms per frame. This effectively balances the requirements for high-precision perception and real-time computation.