Pruning of the Object Detection Model Based on Multi-level Feature De-correlation

In response to the lightweight requirements of object detection models, this paper introduces a pruning method based on multi-level feature de-correlation aimed at reducing both the parameters and computational cost of these models. This approach provides a technical solution for the practical deployment of object detection models on resource-constrained devices. Three core innovations are designed with consideration for the complex structure of object detection models, the high similarity among learned features, and their sensitivity to pruning: (1) Incorporating feature de-correlation constraints during model training to diminish correlations between features, thereby enhancing the model's efficiency in learning; (2) Developing a dynamic evaluation criterion for channel importance based on the L1 norm of feature activation to identify and eliminate redundant structures within the model; (3) Proposing a modular pruning framework that complements mainstream channel selection algorithms from a technological perspective. On the PASCAL VOC benchmark dataset, this method has demonstrated its effectiveness by compressing 40% of both parameters and computational costs in the YOLOv5 model while incurring only a 1.15% reduction in mean Average Precision (mAP). Additionally, it achieved a 20% reduction in parameters and computational costs for the YOLOv8n model at an expense of merely 0.4% mAP loss. This methodology is particularly suitable for lightweight implementations before actual deployment of object detection models.