<p>Automated detection of prohibited items using X-ray security images remains challenging due to severe object overlap, low contrast, and complex structures. Existing methods often exhibit limited capability in capturing complementary frequency and spatial characteristics that are essential for reliable recognition in such conditions. This paper presents a frequency–spatial guided deformable fusion network that enhances multi-scale feature representations for X-ray prohibited item detection. The proposed framework incorporates three core components. The first is a frequency–spatial guided deformable convolution, which integrates fixed and learnable Gabor filters with frequency-aware channel encoding and depthwise deformable convolution to improve the extraction of structural and textural details. The second is a balanced channel bottleneck that produces a softmax channel weighting map to adaptively emphasize discriminative feature channels. The third is a tri-path frequency–spatial modulation module that combines enhanced and residual features through multiplicative interaction, thereby improving robustness in scenes containing densely overlapping items. These modules are embedded within a deformable depthwise separable feature pyramid to achieve efficient multi-scale aggregation. Extensive experiments on widely used X-ray benchmarks, PIDray, OPIXray, and HiXray, demonstrate that the proposed model achieves superior and real-time performance.</p>