In recent years, as the transportation system becomes more and more developed and people travel more and more frequently, it has become an important means to check people′s packages through X-ray security screening machines to prevent incidents endangering public safety. At present, dangerous goods detection of X-ray images in many places is still carried out manually by security inspectors, which has problems of heavy workload and low efficiency. Therefore, it is necessary to use the existing target detection technology to detect dangerous goods automatically. However, the background of dangerous goods in X-ray images is often complicated, which leads to the low accuracy of the automatic detection of dangerous goods. Therefore, this paper proposes an improved att_decouple_YOLOv5s model. By introducing the Convolutional Block Attention Module (CBAM) in the feature fusion part of YOLOv5s model, the relevant feature information is strengthened and the background information is suppressed; The detection head of the model is decoupled by the parallel branch method to solve the conflict problem caused by the coupling of classification and positioning tasks. Finally, this paper tested 12 kinds of dangerous goods in X-ray images in the public data set pidray. The experimental results showed that the average accuracy of the detection performance index of the proposed model reached 88.1% when the IoU (Intersection over Union) threshold was 0.5, which was 2.8% higher than that of the YOLOv5s model; When the IoU threshold was between 0.5 and 0.95, the average precision reached 76.6%, which was 4.3% higher than that of YOLOv5s. The effectiveness of the improved algorithm is fully verified. [ABSTRACT FROM AUTHOR]