Research On Model Lightweight Method For X-ray Image Contraband Detection

With the rapid development of society,it’s more and more convenient for people to travel.As an important part of ensuring travelling safety,it has gotten a common method to use X-ray machines for contraband detection.At present,the manual checking method can no longer meet the needs of the rapid growth in peak passenger traffic,at the same time,some automatic detection methods can’t be applied due to the large amounts of calculation,high model complexity,and low detection efficiency.This thesis focuses on the lightweight methods,in order to train an X-ray image contraband detection model that meets the needs of both accuracy and speed,which can improve the efficiency of contraband detection.On the one hand,deep neural network generally has a complex network structure now,on the other hand,most X-ray machines are limited by cost and can’t provide GPU resources with high calculation capability,thus causing the models fail to be applied.For this problem,this thesis proposes a model lightweight method for X-ray contraband detection based on knowledge distillation.This method can greatly reduce the size of the model and improve the inference speed with a small accuracy loss.Firstly,aiming at the problem that only part of the teacher network information is used in previous knowledge distillation methods,this thesis proposes a feature reuse mechanism based on the cascade structure,which can reasonably use the information of the middle layers and connect the features between adjacent layers to improve the information utilization.Secondly,this thesis proposes a simplified feature fusion strategy,which combines the middle-layer features and the deepest features at the cost of minimal computation.Finally,this thesis designs a loss function to focus on the output logits of the teacher network to accelerate the convergence process.For it is difficult to train a large teacher network because of poor GPU sometimes,this thesis proposes a contraband detection method based on self distillation without training a complex network in advance.This method can get better accuracy by training two simple student networks compared to directly train a small network.Firstly,dividing the output features of the first student network into high-order and low-order parts,using the final results to distill the low-order and the high-order,and adding the stochastic depth to the second student network.These two networks can be improved alternatively by sharing information.Secondly,this thesis proposes a method to calculate the similarity of feature structures,and it can improve utilization of the middle layers,so as to increase the accuracy with little calculation cost.Finally,this thesis designs two different loss functions for the two student networks respectively,so that both networks can achieve better results.In addition,this thesis builds a dataset containing thousands of X-ray images,sufficient experiments on this dataset show that the proposed methods perform well,which produce a smaller model with higher efficiency.