Research on Network Pruning based Target Detection Model for CenterFusion

Deep learning has achieved remarkable success in the field of object detection. However, efficiently deploying these models in embedded systems still poses challenges related to model size and computational complexity. This paper focuses on optimizing the CenterFusion model through the application of the weight-based L1 norm pruning method. We adopt a targeted network pruning approach, specifically addressing the Centernet component within CenterFusion, to reduce the model’s parameter count and complexity, making it more suitable for resource-constrained environments in embedded systems. The innovation of this research lies in the inaugural application of the weight-based L1 norm pruning method to the CenterFusion model. Additionally, it incorporates two iterations of pruning based on weight changes, allowing for further compression of the network model. This provides a practical solution for the efficient deployment of the model on embedded platforms. Testing before and after pruning shows a reduction in the model size by 31.86% and 52.94%, with an overall accuracy loss of only 3.18%. The pruned model significantly reduces volume while maintaining object detection performance. This suggests that the pruned network achieves a lightweight object detection model, demonstrating the feasibility and effectiveness of deploying advanced object detection technology in embedded systems.