Research On Dangerous Goods Detection Algorithm Of Security Inspection Machine Based On Incremental Learning

Security inspection system plays a role in ensuring human’s safety in crowded places.At present,most security inspection systems have been intelligent,but with the upgrade of terrorists in terrorist incidents,many reconstructed heterogeneous dangerous goods have appeared.The existing dangerous goods knowledge prior model cannot be effectively covered,and it is necessary to re-train the detection model through the expanded data set to realize the identification of the dangerous goods.In order to solve the problem,an incremental learning target detection algorithm that can make full use of the original model and improve training efficiency is studied.Based on the basic principle of the traditional target detection algorithm,the incremental learning target detection algorithm based on Fast rcnn is studied in order to avoid the gradient disappearing during the training process.The feature extraction network is replaced with a 50-layer residual network;at the moment,in order to enable the network structure to achieve learning of new categories,the classification neurons and the regression neurons that are corresponding to the number of new categories are added to the fully connected portion which are used to the output of the new classes;and then the loss function is designed for the training process of the network so that the network can avoid the disaster forgetting while learning the new class goal.Further,for the problem that the Fast rcnn network needs to generate a separate area,the Faster rcnn network is adopted and its feature extraction network is replaced.The incremental learning algorithm of RPN regional proposal network under the target detection framework is focused on;finally,the incremental learning idea is applied to the classification and border regression network in the Faster rcnn framework,the algorithm improves the detection accuracy and training speed.The algorithms in this paper are all conducted on public driverless data set and security dataset to verify the feasibility and effectiveness.The experimental results show that the algorithm obtains more than 85% detection confidence in both old and new classes,which means that the algorithm can effectively use the network parameters in the old class,and can also learn the new class objectives,and then improve the problem of low efficiency of the training model.