Research On Subway X-ray Image Detection Algorithm Based On Convolutional Neural Network

More and more cities have built new subways,which means there will be more security check facilities.Subway security check is related to passengers’ personal safety,which puts forward higher requirements for subway security check.This requires more intelligent security algorithms to assist the check.The method based on convolutional neural network is the most potential intelligent subway security check algorithm at present.It can use the convolutional neural network to mine the feature information from a large number of training data,and detect or segment various objects in the image through feature extraction and feature combination.It has the advantages of high speed and high precision.This thesis mainly studies the algorithm of object detection and instance segmentation in subway X-ray image based on convolutional neural network.Object detection and instance segmentation algorithms based on convolutional neural networks can be divided into two different algorithms according to their structures,one-stage algorithm and two-stage algorithm.Two-stage algorithm needs to generate candidate boxes,and then detect or segment them on the candidate boxes,while the one-stage algorithm directly detects or segment them on the feature map.The difference in structure leads to the high speed of the one-stage algorithm but slightly lower precision.Since the security check algorithm has a high requirement for real-time performance,this thesis further improves the accuracy of the object detection and the instance segmentation algorithm based on the one-stage algorithm.All data used in this thesis are collected from security check facilities.In reality,there are fewer prohibited items in security check scene,which leads to the problem of unbalanced data collection category.In order to reduce the model error caused by data imbalance and make the detection model more robust,this thesis uses CT-WGAN algorithm to generate security check images on the foreground image data set of prohibited items.The experimental results show that the image generation algorithm has excellent performance and can improve the accuracy of detection model.Based on the new research idea,this thesis first introduced a object detection algorithm ARF-YOLO.Attention mechanism module was added to the residual structure to improve the feature extraction ability.Secondly,the receptive field module was introduced to increase the receptive field of the model.Moreover,Focal Loss function was added to the loss function to improve the optimization direction of the model.The experimental results on the collected data set show that the improved network can improve the accuracy rate on the basis of adding a few parameters,and can deal with a variety of complex environments.In order to obtain a more accurate profile,a segmentation algorithm,HSA-Yolact,was proposed.The Hierarchical-Split block with stronger feature extraction ability was added to the network,and the multi-level feature map was adaptive fused to enhance the feature expression ability.After the experiment on the collection data set,the network can be qualified for the security scene which requires high real-time.