<p>Accurate and fully automated segmentation of contraband in X-ray security imagery is crucial for maintaining public safety. However, existing methods remain insufficient in addressing challenges such as boundary blurring caused by stacking and occlusion, extreme variations in object scales, and complex background interference. To overcome these limitations, we propose a Detail-perceptive and Multi-scale Collaborative Network (DMC-Net) that integrates Transformer and CNN architectures. First, we introduce a Boundary Detail Perception (BDP) module that incorporates gradient-aware convolution operators to preserve high-frequency edge information, effectively mitigating boundary ambiguity. Second, we design a Multi-scale Collaboration (MSC) module that employs parallel atrous convolutions with diverse dilation rates, a dedicated branch for small objects, and attention mechanisms to collaboratively enhance multi-scale feature representation. Finally, we incorporate a Global–Local Spatial Aggregation (GLSA) module in the decoder to improve segmentation accuracy under complex backgrounds by combining global contextual cues with local spatial information and using high-level semantics to refine fine-grained details. Experiments on the PIDray dataset demonstrate that DMC-Net achieves mIoU scores of 78.97%, 59.24%, and 74.30% on the Easy, Hidden, and Hard subsets, respectively, validating its effectiveness and robustness in challenging security screening scenarios.</p>