Model Pruning and Deployment Optimization for Ship Detection

For unmanned aerial vehicles (UAVs) or satellites, realizing real-time detection of intelligent maritime targets based on edge devices is an urgent need for high-efficiency information acquisition. However, current artificial intelligence detection models have large number of model parameters and computational requirements, making it difficult to achieve real-time processing on edge devices. To address this problem, this paper proposes a human-in-the-loop(HIL) model pruning method, which improves the sparse training effect of the model by introducing Gaussian penalty terms, and combines automatic and manual method to set pruning threshold to meet the goal of model lightweighting. Based on the pruned and quantized model, the size of the model is reduced by 87.94% compared to the original model (YOLOv5), with only a 1.48% decrease in accuracy. After deployment optimization on GPUs, the real-time performance is improved by 2.28 times.