A Lightweight Recognition Model for Tunneling Machine Components

To address the challenge of balancing accuracy and real-time performance in target recognition for Mixed Reality (MR)-assisted maintenance, this paper proposes a lightweight recognition model based on YOLOv8. Firstly, a heterogeneous convolution module is introduced to replace the original C2f module, reducing the number of floating-point operations (FLOPs) and improving recognition speed. Secondly, linear deformable convolution modules are used to replace some standard convolution modules in the original backbone network, further enhancing the model's real-time capability. Thirdly, the neck network is reconstructed using a Reparameterized Generalized Feature Pyramid Network (RepGFPN) to strengthen feature interactions and improve the model's feature fusion ability. Finally, the original loss function is improved by adopting the Shape-IoU loss function, which places greater emphasis on target shape, leading to more accurate bounding box regression. Experimental results on a self-built dataset of tunneling machine components demonstrate that the improved model achieves a 0.4% increase in precision, a 30% reduction in parameter count, and a 13.9% increase in frame rate. This ensures accuracy while significantly improving recognition speed, meeting the application requirements for MR-assisted maintenance systems in mining equipment.