Aiming at the problems that the data set used by the existing recognition algorithms is too simple, the recognition accuracy of dangerous goods in security inspection images in real scenes is low, and it is easy to lead to false detection and missed detection, we propose a class-balanced hierarchical refinement algorithm based on penetration hypothesis, which combines multi-layer channel attention mechanism and space attention mechanism. First, based on the hierarchical modeling of security image, channel attention mechanism is added to the feature map to give different weight to different channel features. Then, spatial attention mechanism is added to give different weight to the unique color features of security image in space. Finally, the residual network is used to add double attention mechanism to different layers of security image for ablation experiment. The experimental results show that after adding double attention mechanism to the fixed two layers at the same time, the network can significantly improve the identification accuracy of dangerous goods in security inspection, and verify the effectiveness and robustness of the multi-layer attention mechanism algorithm.