Research On An Efficient X-Ray Image Prohibited Item Detection Method Based On Improved Convolution Neural Network

The use of X-ray security machines to locate and identify contraband in luggage items is a major measure to ensure public safety,and more accurate X-ray image contraband detection methods can effectively improve the efficiency of security work.Convolutional Neural Network(CNN)based target detection methods in X-ray image contraband detection has achieved certain research results,but it mainly uses the convolutional neural network designed by human hand,relying heavily on the empirical knowledge of researchers,and the final detection effect is not ideal.In addition,there are problems in identifying and locating contraband in X-ray images with complex backgrounds,and in ignoring the directional characteristics of the items.This paper investigates these problems and proposes a contraband detection method to address the difficulties and challenges in X-ray contraband detection.The main research work and conclusions of this paper include the following two parts:(1)To address the problems of low automation of convolutional neural network design and unsatisfactory accuracy and speed of contraband detection in the context of complex X-ray images,this paper proposes an automatic search design target detection model backbone network(Backbone)component based on Neural Network Architecture Search(NAS)This paper proposes an automatic search design for the backbone network(Backbone)component of the X-ray image contraband detection method: Layer-by-Layer Progressive Neural Network Architecture Search(LLP-NAS).First,the method defines the search space based on the Rep VGG residual network structure,which consists of 10 residual branches.Second,a layerby-layer progressive performance evaluation strategy is used to compute the Batch Normalization(BN)metric to search for the optimal side branch structure for each layer structure with different feature extraction capabilities.Then,a new Backbone component is constructed by searching layer by layer.Finally,the Backbone component of the original target detection model is replaced by the Backbone component obtained from the search to form a new target detection model.The method was experimentally validated on three contraband detection datasets Hi Xray,OPIXray,and PIDray,and all the indicators were due to other mainstream target detection methods,achieving 83.4%,87.2%,and 70.4% detection accuracy,respectively.(2)In order to further solve the problems of the method in(1),this paper proposes a method based on the improved YOLOv7 X-ray image rotation target detection for X-ray image contraband detection in the identification and positioning difficulties and ignoring the problem of the directional characteristics of the items.First,the model’s ability to extract deep important features is improved by fusing efficient attention mechanism modules in the backbone network of YOLOv7;then,the feature fusion path of the extended efficient long-range attention mechanism in YOLOv7 is improved by adding jump connections and 1×1 convolutional architecture between the attention mechanism modules to enable the network to extract richer item features;finally,for the X-ray image The problem of arbitrary directions of contraband placement is addressed by using a dense coded label representation to discretize the angles in order to improve the accuracy of contraband localization.The improved YOLOv7 network was compared with mainstream target detection algorithms such as FCOS,Re Det and SCRDet on Hi Xray,OPIXray and PIDray contraband detection datasets,and the method achieved 91.2%,92.6% and 66.4% detection accuracy,respectively.The experimental results show that the proposed method in this paper can effectively improve the accuracy and speed of contraband detection in complex X-ray image background,which provides a good technical support for ensuring public safety.