DRFS-Net: dynamic receptive field selection network with hierarchical attention for remote sensing object detection

This study addresses the challenge of ship detection in optical remote sensing images by proposing a Dynamic Receptive Field Selector Network (DRFS-Net). The method effectively resolves issues of feature coupling and interference inherent in traditional shared feature maps by introducing an innovative task-parameter decoupling mechanism for classification and regression tasks. Specifically, DRFS-Net contains two core modules: a Parallel Three-branch Precision Extraction Module (PTPEM) and a Lightweight Attentional Dynamic Weight-allocation Module (LADWM). The PTPEM employs three parallel convolutional branches with distinct receptive fields to extract features at varying scales: local details, intermediate contextual information, and large-scale global semantics, respectively. These branches are designed to precisely adapt to targets of different sizes. The LADWM introduces a lightweight attentional mechanism to dynamically allocate weights during feature fusion, thereby enhancing focus on critical target information and suppressing less discriminative features. Extensive experiments were conducted to validate the effectiveness of DRFS-Net, utilizing mainstream datasets including HRSC2016 and UCAS-AOD. The results from ablation and comparative studies demonstrate that DRFS-Net significantly improves detection accuracy, achieving a mean average precision (mAP) of 91.54% on the HRSC2016 dataset. This performance surpasses existing methods, including both traditional approaches (e.g., RRPN, R2CNN) and more recent axis learning techniques. Overall, the experimental results indicate that DRFSNet obtains superior performance in complex remote sensing scenarios by leveraging task-level feature decoupling and parameter-level dynamic allocation mechanisms. Consequently, it offers an efficient and accurate solution for target detection in optical remote sensing images.