Research On X-ray Contraband Identification Technology Based On Convolution Neural Network

With the country’s increasing emphasis on the field of public safety,X-ray security inspection equipment is widely used in various transportation hubs.However,in the current security inspection process,the identification process of contraband still needs to rely on manual identification by security personnel.Often in the complex scenes such as train stations and subway stations,false detections and omissions occur.This article addresses the problem by introducing the concept of deep learning.Through the improved YOLO v3 model,a set of systems that can automatically mark and identify contrabands has been successfully constructed.While improving the accuracy of contraband inspections,it has also improved the overall security inspection.This paper first proposes a new image enhancement idea based on the characteristics of the X-ray security inspection image and the images required by the convolutional neural network model,referring to medical X-ray processing technology,and adopting multiple complementary image enhancement methods.First use the Laplacian and Sobel operators to obtain the detail map and gradient map,and then use bilateral filtering on the gradient map and use it as a template to multiply the concept of the image mask by the detail map.USM sharpening obtains the final sharpened image,uses image fusion and addition of the original image,and finally uses CLAHE to increase the dynamic range of image grayscale.The experimental results show that the method proposed in this paper can better retain image information than traditional image enhancement,and has strong robustness to noise.Secondly,the accuracy and real-time performance of different models based on convolutional neural network are compared,and the YOLO v3 model is selected to build the network.In order to make the model meet the security requirements,this paper has carried out a series of optimization processing on the YOLO v3 model.First,the concept of dense links is introduced to avoid the problem of small target features being lost during training in the YOLO v3 model.Second,the K-means cluster is optimized.Io U is used instead of Euclidean distance to avoid the problem of local optimization.Finally,Using Soft-NMS improves the ability of the model to detect occluded objects when multiple objects occlude each other.