Abstract: To address the limitations of restricted receptive fields, inefficient multi-scale feature fusion, and high computational complexity in existing synthetic aperture radar(SAR)ship detection models, a channel-aggregated feature pyramid network(CA-FPN)was proposed. At the feature pyramid architecture level, a cross-level dynamically weighted feature fusion mechanism was introduced. This mechanism adaptively calibrated multi-resolution features using learnable channel attention weights, which overcame the fixed sampling constraints of the traditional asymptotic feature pyramid network(AFPN)and significantly enhanced multi-scale target representation. In designing the feature extraction unit, a parallel multi-scale dilated convolution block(PDBlock)was developed. By integrating the squeeze-and-excitation(SE)channel attention mechanism with dilated convolution techniques, and employing a feature channel aggregation gating mechanism, the PDBlock effectively mitigated semantic conflicts during multi-scale feature fusion. Experimental evaluations on the standard SSDD and LS-SSDD datasets demonstrated that, compared to the baseline AFPN model, CA-FPN maintained detection accuracy while reducing model parameters from 1.93×10⁶ to 1.17×10⁶(a 39% reduction)and computational complexity(in G_FLOPs)from 4.24 to 3.19(a 24% reduction). The mean average precision increased by 2.8% on SSDD and 3.5% on LS-SSDD. CA-FPN was more effective and better adapted to the requirements of SAR ship target detection tasks.

Key words: synthetic aperture radar, ship detection, multi-scale, cross-level feature fusion, attention mechanism, dilated convolution