Due to the complex underground environment, low lighting conditions, and the small size of hard hats, the detection results are not ideal. To address low-quality images in complex environments, an improved YOLOv7 for hard hat detection in low-quality images from underground coal mines was proposed. Firstly, addressing the limitation that image features were susceptible to noise interference under low-light conditions, a multi-scale MELAN module was introduced. By constructing a multi-scale attention mechanism, broader contextual information was captured, thereby enhancing feature extraction and effectively suppressing noise interference. Secondly, the OD-SMP module was constructed using soft pooling and full-dimensional dynamic convolution in the backbone network, which reduced information diffusion in feature mappings, retained more contextual information, and enhanced the detection capability for small targets. Finally, to address the varying quality of detection samples caused by the complex backgrounds and environments with different lighting and distances in underground coal mines, Wise-IoU was used as the loss function. Experimental results show that the average precision of the improved model is 94.9%, which is 13.5% higher than the original YOLOv7 model, demonstrating better detection performance.