Due to the unique imaging mechanism and electromagnetic scattering characteristics of synthetic aperture radar (SAR), the images it acquires are often contain strong speckle noise and complex background interference, which severely limit object detection accuracy and robustness. To mitigate noise and address the limitations of insufficient multi-scale feature modeling and limited frequency information usage in existing methods, this paper proposes a SAR object detection network combining frequency dynamic convolution with spatial-frequency feature enhancement. A dynamic frequency-domain convolution module is first introduced, which adaptively modulates high and low frequency components via learnable Fourier coefficients and grouped reconstructed convolution kernels, combined with a feature restoration module, enriching the diversity of the kernel’s frequency responses and enhancing the capability of feature representation. A spatial-frequency feature enhancement module is then constructed using joint frequency and spatial self-attention with a spatial-frequency fusion strategy, enhancing the features of targets. Experiments on MSAR and SARDet-100K datasets demonstrate that, compared with DenoDet, FCOS, PVT-T, Faster R-CNN, and YOLOF, the proposed method achieves significant improvements across multiple evaluation metrics and shows higher detection accuracy, robustness, and generalization, offering a new approach for SAR image object detection.