Research on the Real-Time Scoring of Chest Ring Target Based on Transfer Learning and Improved Lightweight Neural Network

To achieve real-time scoring of chest ring targets during outdoor shooting training while addressing the issues of limited memory and low performance in portable devices, this paper proposes a real-time scoring method for chest ring targets based on an improved lightweight neural network architecture. To enable efficient and high-precision real-time bullet hole detection, the paper lightweight processes the YOLOv10n end-to-end regression framework. First, the Repvit-MobileNet block (reparameterized depth wise separable mobile network module) is introduced into its backbone network to reduce model size and computational complexity while improving detection speed. Then, grouped shuffle convolution (GSConv) is used to reconstruct the standard down sampling convolutions in the PANFPN structure, promoting feature information extraction and fusion at bullet hole edges, reducing model computational complexity, and maintaining detection accuracy. Additionally, to enhance the model’s ability to extract local feature information, the efficient channel attention (ECA) mechanism is integrated to replace the PSA attention mechanism in the backbone network, strengthening the edge features of bullet holes on chest ring targets and improving the recall rate for small target detection. Finally, transfer learning is incorporated to accelerate model convergence. The lightweight neural network model is deployed on portable mobile devices, and a real-time scoring system for chest ring targets is developed using MFC and C++. Experiments show that this method exhibits good adaptability and robustness in different environments. Compared with the original YOLOv10n model, the improved model achieves a 12.1% increase in mAP50, a 10% increase in mAP50-95, an inference speed of 25.43 FPS, and a model weight of 5.149 MB, meeting the real-time scoring requirements for chest ring targets in outdoor shooting training.