Detection Of Dangerous Objects From X-ray Inspection Images Based On Deep Learning:Algorithm Development And Its Application

The popularity of national comprehensive three-dimensional public transport facilities and the prosperity of express logistics and transportation industry have brought great convenience to people’s life,but they also have big security risk.Non-contact security detection is a key measure to ensure public security.The traditional security inspection method relies heavily on manual detection with high labor costs,which is easily affected by environmental and subjective mental factors,resulting in hidden safety hazards of missed detection and false positives.Combined with artificial intelligence technology represented by deep learning,research and development of algorithms and systems that can automatically detect dangerous goods in X-ray security inspection images can effectively improve the security and reliability of security inspections,and provide a solid guarantee for the construction of a security system for public services,to help the construction of smart transportation.This paper is devoted to the algorithm research and system implementation of Xray security inspection image dangerous object detection,mainly from the two different dimensions of improving accuracy and light weight.The intelligent security inspection system deploys the detection network into the security inspection equipment,and realizes the real-time detection and data management of the equipment terminal.The main contents of this article include the following points:1.A symmetrical triangle detection algorithm is proposed: according to the characteristics of X-ray security inspection images,this paper discusses the advantages and applicability of anchor-free mechanism in security inspection tasks from theoretical analysis and comparative experiments,and proposes a symmetrical triangle detection network based on anchor-free.At the same time,a symmetrical triangular pyramid module is proposed,which effectively integrates the semantic and spatial location information of different scale feature maps,and adopts hierarchical independent prediction to alleviate the overlapping problem and improve the ability of multi-scale detection of the network.The experimental results of DE dataset and public dataset collected in this paper show that this algorithm can effectively improve the detection accuracy.2.A lightweight target detection algorithm light R-CNN is proposed: this algorithm introduces a residual branch feature extraction module in the backbone network to expand the receptive field and retain more shallow features,which enhances the feature extraction capability of the backbone network.At the same time,the proposed multi scale feature fusion model combined with depth-wise separable convolution fully integrates feature map information of different scales with only a small computational cost.Finally,the detection head structure is compressed and optimized,and a split-channel self-attention module is proposed to optimize the feature distribution and improve the sensitivity of the network to foreground positions.The algorithm achieves a balance between accuracy and network complexity,and is suitable for deployment and application in resource-constrained security equipment.3.An intelligent detection system for dangerous object in X-ray security inspection images is implemented: the system is computerized based on the lightweight target detection network light R-CNN.The functions of image acquisition,transmission,preprocessing,intelligent detection of dangerous goods,cloud data storage,image annotation,and statistical analysis and query are realized through the development of three modules: acquisition terminal program,intelligent detection system program and remote data management program,and the security inspection image and data management can be detected in real time at the security inspection equipment terminal.