STRay: A Model for Prohibited Item Detection in Security Check Images

Addressing issues such as mutual occlusion of items and small scale of prohibited items in X-ray security inspection image detection, we propose an improved X-ray contraband detection model based on YOLOv7 named STRay. Firstly, in the backbone network, the model employs Swin Transformer, applying a sliding window multi-head self-attention mechanism to suppress background interference, enabling the network to focus more on contraband items and reducing the false negative rate. Secondly, conventional convolutions in E-ELAN are replaced with deformable dilated convolutions, adjusting the convolutional kernel's shape by learning sampling offsets to better match the contours of contraband items and effectively address mutual occlusion issues. Lastly, the detection head in the head section is replaced with an Efficient decoupled detection head, decoupling separate feature channels for localization and classification tasks, thereby enhancing the classification and localization capabilities for small-scale contraband items. The proposed model is tested on large datasets SIXray, OPIXray, and PIDray, achieving mAPs of 95.3%, 88.8%, and 83.1% respectively, effectively improving contraband detection capabilities while maintaining fast detection speeds. Compared to current mainstream models, it demonstrates certain advancements, providing excellent technical support for ensuring public safety.