| In recent years,with the rapid development of terahertz imaging systems and technologies,terahertz has been considered the most competitive next-generation human security technology.Terahertz can penetrate clothing and discover hidden objects carried by the human body.Compared with metal detectors,metal can only be detected.The terahertz security system can detect various items including metal,ceramic,plastic,powder,etc.Rays,which have low photon energy and no ionizing radiation to the human body,are increasingly concerned with research based on terahertz imaging security systems.In this paper,based on the characteristics of terahertz image dimming and low visual recognition rate,combined with target detection in deep learning,a network structure based on VGG16 and improved terahertz dangerous goods detection based on Faster RCNN is designed.Firstly,in view of the fact that the sample size of the actual terahertz data set is small,through the operations of rotation,translation and brightness adjustment,the number of data sets is increased without increasing the number of samples,and the robustness of the model is improved.The classification accuracy rate increased by 2.37%.In addition,a pre-training model is obtained by using the public data set pascal voc training,then the convolution layer of the pre-training model is fixed,and the learning rate of the softmax layer and the fully connected layer are changed respectively,and the parameters of the pre-training model are used instead of the parameters of the random initialization.Different migration learning strategies to help the training of the new model,after several experimental comparisons,found that the highest classification accuracy of the model increased by about 2.3%.After data enhancement and the use of pre-trained models,the classification accuracy of the model increased by approximately 3.2%.Finally,aiming at the problem of positive and negative sample imbalance in the model,referring to the idea of ??bootstrap method in traditional machine learning,the difficulty layer oftraining is selected by improving the Loss layer,and the gradient of the remaining candidate frames is set to 0.To solve the problem of sample imbalance,compared with setting the positive and negative ratio random sampling,the scheme of modifying the Loss layer not only reduces the setting of the hyperparameter,but also improves the model by about 1.8% mAP(mean average precision)... |
