Research And Application Of Dangerous Goods Detection Algorithm In X-ray Images Based On Deep Learning

At present,security inspection work mainly relies on security staff to identify whether there are dangerous goods or not in X-ray images,but this method exist high labor costs,low detection efficiency,and excessive reliance on security staff can easily lead to problems such as misdetection.In order to solve this problem,this paper proposes a YOLOv3-HZ dangerous goods detection algorithm based on YOLOv3 algorithm,and designs and implements a system for dangerous goods detection.In the era of rapid development of artificial intelligence technology,using deep learning algorithms to assist security staff to determining whether there are dangerous goods or not in X-ray images,helping to save labor costs and improve detection efficiency,this research has important theoretical significance and application value.The main research work and research results obtained in this paper are as follows:1.In order to improve the detection accuracy of the YOLOv3 algorithm,the following improvements are mades to the algorithm:(1)this paper will optimizing the YOLOv3 network structure,add some residual unit in darknet53 to improve the expressive ability of the model.In order to extract the spatial characteristics of different receptive fields and to obtain richer local feature information,this paper introducing the SPP-Net structure into the YOLOv3.It is founds that the number of convolutional layers feature maps decreases too fast,and there is a possibility of losing information,so some convolutional layers with a convolution kernel of 1x1 are added.(2)There are many overlaps of target objects in the data set,so this paper will improve the border selection algorithm,to reduce the mistaken deletion of target objects.(3)In order to make the network better and make more reasonable predictions,K-means clustering method is use to perform cluster analysis based on the data set in this article to obtain more appropriate priori anchors for the data set in this paper,thereby improved the convergence speed during model training and improve the detection accuracy.(4)In order to enhance the robustness of the model,data enhancement methods such as rotation and translation of the original image ware used to increase the diversity of the data set,thereby improving the detection accuracy.(5)This article will use a multi-scale training method to train its own data set,so that the network can fully learn the characteristics of images at different scales,so that it can better complete the detection task at different resolutions.2.In order to meet the needs of real-time model detection and deployment of mobile or embedded devices,the following methods ware used to improve the YOLOv3 model:(1)Learning from the linear bottlenecks in Mobilenetv2,and design a lightweight network structure suitable for dangerous goods detection,which greatly reduces network parameters and calculations.Speed up the detection speed of the model.(2)In order to better achieve the compression of the model,YOLO v3-HZ will introduce depth wise separable convolution to replace part of the standard convolution.The amount of network parameters and calculations of the depthwise separable convolution is much lower than standard convolution.This improvement greatly reduces the complexity of the YOLOv3-HZ model.(3)Based on the improved YOLOv3-HZ dangerous goods detection algorithm,the dangerous goods detection system have designed and implemented.The main functional modules of the system can divided into front-end display and server-side.The main functions of front-end display include uploading pictures and displaying detection results.The main functions of the server include model management,common user management,and image detection modules.