| In recent years,there is a growing concern about public security problems caused by terrorist attacks,and public places such as airports and train stations are major locations for terrorists to commit crimes.A common criminal tactic of the perpetrators is to fill the battery part of the cell phone with explosives and install fuses to form a cell phone bomb,disguised as a normal cell phone to carry around and detonate in public transportation.Therefore,it is of vital importance to accurately classify normal phones and explosive phones and take timely explosion-proof measures in the security screening.Note that explosive mobile phone classification is a typical class imbalance task.Although the number of explosive mobile phones is far smaller than the number of normal mobile phones in daily life,one missed phone can cause a huge loss of life and property.The main research of this thesis is to classify the X-ray images formed by normal mobile phones and explosive mobile phones under the security screening machine,using the explosive mobile phones X-ray image dataset EMXray.Deep learning is an effective method to solve the problem of image classification in recent years.Since the development of deep learning,many neural network models have been proposed for image classification tasks.The method of deep learning is used in this thesis,and five common neural networks are used as baseline models,and two methods are designed to improve the classification accuracy of neural networks and resolve the class imbalance problem.First,in the face of the class imbalance,this thesis introduces an equilibrium coefficient called sample cost for each sample and designs its dynamic update rule.The sample cost represents the importance of each sample to the classification task.This thesis reconstructs a sample-oriented loss function based on sample cost that multiplies the sample cost with the cross-entropy loss so that the loss of different samples affects the overall gradient value to different degrees.This method can increase the importance of positive and difficult samples in the training process,avoiding the majority of easily classified samples from overwhelming the classifier.In this thesis,the sample costbased loss function is applied to the EMXray dataset and the comparison experiments with other loss functions are performed,the loss function proposed in this thesis has a significant improvement on the explosive phone classification task.In order to accurately capture important features in the face of class imbalance and similar images of positive and negative samples,this thesis propose a hybrid attention mechanism,named position information attention module.Since the explosives and fuses in explosive mobile phones have obvious position information,the position information attention module segments the feature map based on absolute position information,forming the spatial weights of each patch and the channel weights inside each patch,which are added to the network training process.The fine-grained segment gives the convolutional neural network the ability to highlight detailed image information and the attention to position features,it makes the network focus the attention on the explosives and fuses.The position information attention module is a lightweight module that can be plug-and-play into any network.This thesis inserts the module into five baseline networks,performed ablation experiments,and conducted comparison experiments and GRAD-CAM visualization experiments with the other three popular attention mechanisms.The position information attention module achieves the best improvement in classifying explosive phones on the EMXray dataset.In order to verify the general applicability of the position information attention module,this thesis performs comparison experiments between the module and the other three attention mechanisms on the CIFAR-10 and CIFAR-100 dataset,the position information attention module also achieves relatively outstanding improvement. |
PDF Full Text Request