To address the issues of low resolution, blurred edges of hazardous items, and ineffective segmentation in terahertz imaging, this paper proposes a novel network architecture combining adversarial generative networks and multi-head attention mechanisms for intelligent segmentation of terahertz security inspection images. The architecture optimizes the generator through deep discriminator feature maps to produce more realistic generated images, while the introduction of multi-head attention mechanisms enhances the model’s ability to recognize features of hazardous items. Extensive experimental results demonstrate that, compared to traditional convolutional neural networks, the proposed adversarial generative network exhibits better generalization capabilities at the same depth, and the incorporation of multi-head attention mechanisms strengthens the model’s learning of hazardous item features. This approach shows good performance even when dealing with unknown categories of hazardous items, improving IOU overlap metrics by 9.6% over ResNet-50, 21.3% over ResNet-18, and 12.3% over U-Net. This research supports more accurate and efficient processing of terahertz security inspection images, facilitating the further application of terahertz intelligent inspection systems.