Contraband Detection In Security Inspection System Based On Att-RetinaNet

Aiming at the problem of poor detection accuracy caused by contraband small targets,multiple targets,and mutual occlusion between targets in the security inspection system,this paper proposes a target detection model Att-RetinaNet based on the improved RetinaNet model for secondary positioning.First,based on the RetinaNet model,the secondary fusion of local features and global features is achieved through the Feature Pyramid Networks(FPN).Secondly,after the features are merged,a self-attention module is added to realize information interaction between channels and spaces,so that the model can screen out important information features,and improve the ability of extracting effective information from the network and anti-interference ability.Finally,in order to alleviate the problem that the hyperparameters are difficult to determine,and filter out the repeated detection frames in the prediction process,and reduce the missed detection rate,this paper uses the Soft-NMS algorithm.The data set comes from images of contraband collected manually in real-time at Hankou Metro Station in Wuhan.The evaluation indicators use accuracy(Map)indicators and detection rate(Fps)indicators.In order to show that the model Att-RetinaNet proposed in this paper has better detection capabilities,17 sets of comparison models are selected.The empirical analysis results show that the Att-RetinaNet model proposed in this paper has a good effect in terms of accuracy and detection speed.