An automatic pruning method for SAR target detection based on multitask reinforcement learning

In recent years, research on synthetic aperture radar (SAR) target detection based on deep learning methods has made substantial progress in model accuracy. However, there are still limitations related to model size and computational complexity, which makes it challenging to deploy the model on edge devices, limiting its practical application. To this end, this paper proposes an automatic pruning method for SAR target detection models based on multitask reinforcement learning called SARRLP. First, an automatic grouping strategy (GSP) for a deep learning SAR target-detection model network structure is designed. The network structure of the model is coupled with groups, and the same group has a unified pruning layout. The coupled groups are sorted according to the maximum entropy principle to achieve end-to-end pruning of the SAR target detection model. Second, an automatic pruning strategy search method based on multitask reinforcement learning (RLAS) is designed. This method automatically searches for the best pruning strategy for each network structure, and an entropy-based multitask reward function MOR is constructed to guide RLAS in searching for the optimal pruning strategy considering the edge accuracy, sparsity, and model feature channels entropy. Experimental results based on SSDD and HRSID show that the SARRLP method can achieve automatic optimal pruning detection models. With YOLOv5s as the baseline and a pruning rate of 75%, the accuracy of the SSDD and HRSID datasets is reduced by 0/0.4, and the inference speed increases by 184.4/106.2 images per second.