Concealed Hazardous Object Instance Segmentation for Terahertz Security Inspection Images With Channel-Spatialized Transformer

The terahertz (THz) wave can penetrate many nonmetallic materials, such as paper, plastic, and clothing, making it highly valuable in the field of security checks. However, the low contrast between target objects and the background in THz security inspection images poses challenges for object semantic segmentation tasks. Although various solutions are proposed to address these issues, achieving a balance among low localization and segmentation performance, training data, and model complexity remains a problem. To tackle this problem, we propose an instance segmentation method called YOLO-based instance segmentation method (IYOLO) to locate and segment concealed hazardous objects in THz security inspection images. A channel-spatialized Transformer (CST) module is developed to strengthen backbone feature extraction capabilities by combining global feature learning with optimized channel and spatial features. Moreover, an enhanced upsample block (EUB) is designed to integrate channel operations, encoder-decoder structure, and feature enhancement to improve feature maps. In addition, a shuffle convention (Conv) layer (SCLayer) is proposed in the segmentation head to optimize model parameters and complexity while enhancing the perception and accuracy of the model. Many experiments are conducted on public datasets. The results demonstrate that IYOLO outperforms other state-of-the-art (SOTA) methods in the concealed hazardous object instance segmentation task of THz security inspection images. Especially, in complex scenes, IYOLO shows excellent segmentation ability and effectively avoids false positives and false negatives. Ablation experiments confirm the effectiveness of the proposed module, which improves detection performance while maintaining moderate model complexity. Moreover, compared with other methods, the IYOLO model exhibits superior size and computational complexity, making it more practical for real-world security inspection applications. The code of the proposed method is available at https://github.com/xiaokai11/IYOLO.