Research On Foreign Object Detection And Recognition Technology Of Unbalanced Small Sample Based On Deep Learning

In recent years,terrorist attacks occur frequently at home and abroad.The traditional security detection equipment can only detect metal objects.The detection and identification of hidden dangerous objects is usually done manually,which is not only time-consuming and laborious,but also has a high rate of missed detection.Hidden dangerous objects based on terahertz images can penetrate clothing,find hidden metal and non-metallic dangerous objects,and have low photon energy,which is safe,fast and reliable.In this paper,the machine learning technology is introduced into the detection and recognition of terahertz image hidden danger.For the unbalanced and small sample problems of terahertz image dataset,the unbalanced small sample foreign object detection and recognition method based on deep learning is studied.As follows:1.Aiming at the non-equilibrium problem of terahertz image dataset,a human body foreign object detection and recognition method based on ensemble deep network is proposed.Firstly,the traditional data deformation method is used to initially enhance the overall data and the minority data,and then the deep convolutional auto-encoder residual network is proposed to reinforce the data,which alleviates the sample imbalance to a certain extent.Then,the ensemble deep neural network model is designed,which integrate three pre-training models to extract features automatically.The adaptive sample weighted integrated training network is used to improve the generalization ability of the classifier.The feature extraction results are voted,and the voting result is sent to the detection network.Performing a bounding box regression process to obtain a final detection result.During the training,the penalty weight of the minority class is increased.Finally,it can be verified by multiple sets of comparison experiments that the proposed method can perform accurate foreign object detection under unbalanced data conditions,and the mean Average Precision is improved by 13%-20% compared with the comparison method.2.Aiming at the problem of few samples of terahertz image markers,a method for detecting and recognition foreign objects in deep siamese network based on large margin is proposed.A siamese network with two convolution sub-networks is constructed.The number of training samples is amplified by inputting pairs of samples of the same class and different classes,the large margin loss is defined,and discriminative features with distance-preserving properties are learned.Considering the sample distribution,the correlation between different dimensions and the difference of scales,the Markov distance is used to measure the similarity between samples,the weights are shared during pre-training and the contrast loss function is used,and the large margin loss function is used when training the classification network.The results of multiple sets of comparison experiments show that the method can accurately detect and identify small sample data,reaching 93.89% on Average Accuracy.3.Aiming at the over-fitting problem caused by the terahertz image data marker samples,based on the large margin deep siamese network,a deep pseudo-siamese network foreign object detection and recognition method based on Convolutional Neural Network(CNN)and Long Short-Term Memory(LSTM)Networks is proposed.The constructed network has two branch networks.The one is the CNN branch network,which is effective to learn the potential characteristics between the continuous and discontinuous data of the sample.The other is the LSTM branch network,which is used to input the information of different dimensions into a sequence.In the network,I hope to learn the potential relationship between different dimensions and improve the generalization ability of the model.During the pre-training,the weights of the two branch networks are not shared,and the sample classification boundaries of different classes are kept away from the samples through training.Through multiple sets of comparative experiments,the method has improved the overall classification accuracy,reaching 95.58% on Average Accuracy.