Traditional surface defect detection methods directly process features in the spatial domain, resulting in confusion between defect and background features, loss of directional information, and difficulty in simultaneously capturing local details and global texture features, thereby causing false and missed detections. To address these issues, we propose DPFD-DETR: Directional Prior guided Frequency Decomposition with Axial-Spatial Context Alignment for surface defect detection. Specifically, we designed an Adaptive Wavelet-Statistical Enhancer (AWS) that reduces feature redundancy through wavelet transform-based frequency domain decoupling while enhancing low-contrast feature differences using statistical information; proposed an Adaptive Bias Token module (ABT) that performs dynamic position encoding based on the fusion of local details and global semantics, adaptively adjusting spatial relationships and building long-range dependency modeling; and constructed a Prior-Guided Bidirectional Multi-Scale Extraction Module (PBM), utilizing directional priors to adaptively guide orthogonal decomposition of spatial information, enhancing multi-scale directional feature capture capabilities, and optimizing the trade-off between depth and width. Experimental results demonstrate that DPFD-DETR achieves improvements of \(4.3\%\) and \(2.8\%\) on map \(_{50}\) and map \(_{50-95}\) metrics respectively compared to baseline models.