Key Symbol Detection in Coal Mine Engineering Drawings Based on the Fusion of Convolution and Self-Attention Mechanisms

To address the challenges posed by the complex structure, large volume of information, slow manual interpretation, and low accuracy of coal mine safety monitoring system drawings, this paper proposes a key symbol detection algorithm for coal mine engineering drawings based on the fusion of convolution and self-attention mechanisms to assist workers in reviewing the drawings. First, to improve detection accuracy in areas with dense symbols, we introduce a fusion module, ACmix, which combines convolution and self-attention mechanisms, enhancing the model's performance while maintaining computational efficiency. Next, the RNE_SCConv module is designed to reduce redundant computations in the feature extraction process, both in spatial and channel dimensions, improving feature extraction efficiency and further enhancing detection performance. Finally, to address the issue of bounding box regression accuracy, the Inner-IoU loss function is integrated into the original CIoU loss function to optimize its performance, thereby improving the accuracy of bounding box regression. A dataset was constructed using drawings from several coal mine safety monitoring systems in a specific province, with separate training and test sets, to evaluate the performance of the proposed algorithm. Experimental results show that, on this self-constructed dataset, the improved algorithm achieves an accuracy of 94.8%, recall of 93.7%, and mAP@0.5 of 94.2%, outperforming mainstream methods by 6.5%, 0.2%, and 3.0%, respectively. These results demonstrate a significant improvement in overall performance, underscoring the algorithm's enhanced practical value.