X-ray Security Inspection Of Prohibited Items Detection Based On Deep Learning

In order to ensure public safety and prevent potential hazards caused by inflammable and explosive contraband,security inspection equipment is widely used in airports,railway stations,subways,and other public transportation places.At present,in security inspection work,the security inspection mode of’ manual-based,machine-assisted is usually adopted,which leads to the inefficiency of the security inspection work and the inability to meet the needs of passengers’ rapid passage.In order to improve the efficiency of safety inspection work,it is necessary to carry out the automatic detection of contraband in luggage.The existing contraband detection models still have limitations,such as low accuracy.Because of the characteristics of X-ray security images,our research applies deep learning methods to X-ray image contraband detection tasks.The main work and achievements are as follows:(1)A contraband detection model based on residual dilated convolution and Transformer is proposed.Based on the self-attention feature extraction encoder of the moving window Transformer,this model uses the dilated residual module based on dilated convolution and residual connection to expand the receptive field of the feature extraction encoder so that the encoder can capture the multi-scale X-ray security image feature map with context information.The experimental results show that this method enhances the model’s ability to extract multi-scale features and improves the average accuracy of prohibited items while increasing the amount of calculation.(2)A contraband detection model based on a cross-stage attention mechanism is proposed.The model obtains a feature map from the feature extraction network and fuses it through bottom-up and top-down paths to obtain a multi-scale feature pyramid.After that,the feature pyramid adaptively generates the weights of different positions and channels by combining the space and channel attention modules and using them to weigh the image features of different levels of the feature pyramid.The experimental results show that the cross-level pyramid combining attention mechanism and feature pyramid can effectively utilize multi-level features,suppress irrelevant information brought by complex backgrounds,and effectively improve the detection performance of the model.(3)A contraband detection model based on class balance perception loss is proposed.Based on the above two models,the model applies the class balance perception loss function to the candidate box generator and detector head stages.Through additional adjustment factors,the weight of foreground samples in the generation of candidate boxes is increased.The weight of contraband with a small number of samples in the detector head stage is increased so that the model pays more attention to the contraband target rather than the background,and increases the attention to the contraband category with a small number of samples.Experiments show that the model can solve the problem of poor average accuracy of contraband categories with a relatively small number of samples due to the large gap in the number of samples between different categories of contraband,to improve the average accuracy of the model to detect contraband,and the method has sure accuracy and robustness.