To enhance the performance of aerial object detection, numerous deep neural networks with larger scales and more complex architectures have been successively proposed. However, this progress has also led to a substantial increase in redundant parameters, computational load, and resource consumption, posing significant challenges for efficient deployment on resource-constrained UAV platforms. To address this issue, network pruning has been widely adopted and has emerged as a research hotspot. Nevertheless, most existing pruning methods primarily focus on reducing the number of parameters and computational complexity. While these approaches can indirectly reduce the model’s resource consumption to some extent, they often fail to meet the strict energy constraints of UAV platforms and struggle to maintain model accuracy under limited power budgets. To overcome these limitations, this paper proposes an energy constrained and structure sensitivity guided pruning algorithm (ECSSG) for aerial object detection networks. The algorithm performs precise decoupled modeling and quantitative evaluation of network energy consumption, using energy as the direct optimization target. It further integrates structural sensitivity to guide the pruning process and incorporates an energy-aware mechanism to determine the pruning ratio for each network layer adaptively. The resulting pruned models not only meet the deployment requirements of various devices with different energy constraints but also maintain high detection accuracy. Extensive experiments conducted on three public aerial datasets—VisDrone, SIMD, and CARPK—using various detection networks demonstrate that the proposed method achieves higher accuracy under lower energy budgets, striking an optimal balance between energy efficiency and model performance.
Sheng et al. (Tue,) studied this question.