Binding Pruning Method via Channel Slimming for Efficient YOLOv5 Object Detection

Current YOLO structural pruning methods exhibit insufficient redundancy reduction in Shortcut structures. Specifically, when dealing with the Shortcut structure in C3 modules of YOLOv5 networks, conventional pruning approaches encounter challenges of feature map tensor mismatch or architectural damage due to residual connections. To address this critical issue, we propose a new pruning strategy, the Binding Pruning Method (BPM), which implements binding constraints on convolutional kernels that require feature map addition during scaling factor ranking and channel pruning, effectively providing an available method for this technical bottleneck. Experimental results demonstrate that our method achieves remarkable compression while maintaining model performance: in high-compression mode, it reduces parameters by 66.7% and computational costs by 48.2% while limiting accuracy degradation to within 5.8%; in balanced mode, it achieves 50.5% parameter reduction and 32.1% computation reduction with merely 1.9% accuracy loss. This work not only demonstrates improvements in YOLOv5 model compression, contributes a new performance benchmark for efficient model compression, but also overcomes fundamental limitations in existing pruning methodologies, providing a reference solution for lightweight deployment in edge object detection scenarios.