With the increasing number of the ocean ships, the demand for synthetic aperture radar (SAR) image ship classification has been much increased. With the development of deep learning, many neural network-based ship classification methods have been presented. However, these networks show unsatisfactory performance on low-quality SAR ship datasets. In this paper, we propose a SAR ship classification method based on dual Metaformer-like networks and learnable decision fusion, which we call LDF-D-MLCNNs. First, we design a Metaformer-like convolutional block to improve learning performance. Secondly, we implement two networks with different kernel sizes and propose the learnable decision fusion module to obtain the final prediction. Kernels of different sizes exhibit diverse extraction capabilities. Experimental results show that the accuracy of the proposed method outperforms many existing SAR ship classification networks.
Target detection in synthetic aperture radar (SAR) images is constrained by three primary challenges. First, speckle noise overlaps heavily with the high-frequency features of target edges in the frequency domain, so standard convolutions cannot suppress noise without sacrificing edge texture. Second, the scattering signature of a SAR target varies markedly with viewing angle, and a fixed-parameter convolution kernel cannot accommodate this spatial non-stationarity. Third, deep and shallow levels of the feature pyramid differ in semantics and resolution, and a naive element-wise sum either introduces noise interference or loses small-target signals. We propose the Frequency–Spatial Detection Network (FSDNet), whose core FSDBlock cascades three operators to address these failure modes in turn. Wavelet Convolution (WTConv) projects features into Haar sub-bands and applies independent low- and high-frequency kernels prior to inverse-DWT reconstruction, suppressing noise while preserving edges. Receptive-Field Attention Convolution (RFAConv) generates location-conditional kernels and so adapts to non-stationary scattering. Spatial Context Self-Attention (SCSA) aggregates discrete scattering points into coherent target representations via long-range grouped attention. At the fusion stage, CGAFusion replaces FPN element-wise addition with a channel–spatial–pixel triple-attention soft switch that mitigates deep–shallow semantic mismatch. On HRSID, FSDNet attains mAP50 % and mAP50:95 %. On SSDD, it attains mAP50 % and mAP50:95 %. Both sets of results consistently surpass the baseline methods. Against the strongest YOLO baseline (YOLOv11n), FSDNet improves HRSID mAP50 by percentage points (pp) and mAP50:95 by pp, and SSDD mAP50 by pp and mAP50:95 by pp; against the capacity-fair YOLOv11s reference (∼51% more parameters), FSDNet still leads on mAP50, mAP50:95, recall, and F1. Ablation studies and power-spectral-density analyses corroborate the contribution of each module and confirm WTConv’s role in preserving high-frequency target features.