Automatic design of efficient SAR ship detection networks for large-scale scenes via neural architecture search

Traditional deep learning-based Synthetic Aperture Radar (SAR) ship detection networks rely on expert manual design, making it difficult to adapt to the diversity of targets and environments. Neural Architecture Search (NAS) can automatically learn SAR target characteristics and improve model robustness, but existing methods suffer from issues such as large search spaces, low efficiency, and high computational costs, limiting their practical application in SAR ship detection. To this end, we propose an efficient automatic network design method for SAR ship detection based on NAS. We first construct a high-capacity SuperNet based on weight sharing, in which multiple convolutional kernels are merged into a unified superkernel to accelerate the architecture search process. Then, we employ the Gumbel-Softmax distribution to dynamically adjust the weights of different paths, integrate multi-branch convolutional outputs, enhance the model's generalization capability. Finally, we unify the architecture search spaces of the Backbone and Neck into a joint optimization framework to more effectively fuse and strengthen hierarchical feature representations. On the SSDD and HRSID datasets, the proposed method achieves search times of 1.981 h and 11.449 h, respectively, with mAP50 detection accuracies of 98.2% and 92.3%, respectively. Demonstrate that our approach can efficiently perform model search for different datasets while achieving high detection accuracy.