LMFAN: Lightweight multi-scale feature aggregation network with channel pruning and knowledge distillation for ship detection in remote sensing images

Abstract

Ship detection in remote sensing images, especially in synthetic aperture radar (SAR) images holds significant application value in maritime traffic monitoring, military reconnaissance, and related domains. To address the accuracy degradation caused by complex maritime background interference, multi-scale target coexistence, and weak feature representation of small targets in SAR images, we propose a lightweight multi-scale feature aggregation network (LMFAN) based on YOLO11. Firstly, a gram polynomial-driven dynamic convolution module that employs differentiable Gram-Kolmogorov basis functions is designed to expand the receptive field of convolutional kernels, enhancing coarse-grained feature representation. Secondly, we employ an enhanced spatial-orientation pyramid utilizing an Omni-Kernel to fuse global, large-local, and local detailed features, significantly improving feature responses for small targets. Thirdly, we design a multi-scale detail-sharing decoder incorporating detail-enhanced and shared convolution to preserve contextual information while reducing computational overhead. Moreover, a dynamic channel pruning strategy with channel-wise cascaded optimization and a channel-wise knowledge distillation with the Kullback–Leibler divergence loss function is introduced to resolves the feature coupling issue in dense target scenarios. Experimental results demonstrate that compared to baseline YOLO11n model, LMFAN achieves mAP$@$0.5:0.95 improvements of 3.6%, 4.9%, and 1.5% on the SSDD, HRSID, and SAR-Ship-Dataset, respectively. Additional validation confirms the model’s superior generalization capability. Compared with state-of-the-art methods, the LMFAN can achieve better accuracy and fewer parameters.