Research On X-ray Security Image Dangerous Goods Identification Method Based On Deep Learning

X-ray security technology is widely used in public transport places and logistics express and other security inspection work,playing an important role in social public safety.However,the current security work mainly relies on the active identification of X-ray security images through the security personnel,vulnerable to the influence of various factors,leading to false checks and omissions and occurrence,which has greater security risks.This paper explores the application of two object detection methods in the field of X-ray security screening,the two-stage method Faster R-CNN,which has higher accuracy but slower speed,and the one-stage method RetinaNet,which has faster speed but lower accuracy.It is found that Faster R-CNN has the problem of inconsistent data distribution in the training and testing domains resulting in its high false detection rate when detecting normal images.In response to the above problems,combined with the majority of X-ray security images belong to the characteristics of normal images,we propose of the X-ray security image recognition method with a pre-classified head to reduce the rate of false detection while improving the efficiency of detection.Meanwhile,taking into account the real-time requirements of security screening,this paper explores the reasons for the low accuracy of RetinaNet and proposes a single-stage dense X-ray security image detection network,which combines the characteristics of X-ray security images and effectively improves the detection accuracy of this network while ensuring its detection speed.The details are as follows:(1)To address the problem that Faster R-CNN cannot learn the normal image features,a pre-classification module sharing the backbone network is designed so that the model can learn the feature distribution of normal images effectively.Meanwhile,in prediction,the normal and abnormal images are firstly classified by the pre-classification module,and when an abnormal image is identified,it is then subjected to hazardous material detection.Experiments show that the pre-classification module effectively reduces the false detection rate of the model for normal images from 27.83% to 3.80%,a reduction of 24.03%.Meanwhile,for normal images,the detection efficiency of the model is greatly improved as further subsequent target detection is avoided,and the detection speed of a single normal image is shortened from the original 0.1843 s to 0.0113 s.(2)To address the problem of deviation in region suggestion due to twice quantization error of RoIPooling,the region suggestion normalization method of Faster R-CNN is improved by introducing RoIAlign,and the bilinear interpolation method is used to directly calculate the eigenvalues at floating point coordinates to improve the detection performance of Faster RCNN.The experimental results show that the detection accuracy of the new model is greatly improved,and the m AP is increased from 48.30% to 57.33%.(3)Combined with the dense distribution of X-ray security image items and more overlapping occlusions,the RetinaNet edge selection method is improved with soft-NMS to reduce the occurrence of missed detections and improve the detection of overlapping occlusion targets,resulting in a 3.27% increase in the average detection accuracy of the model.(4)RetinaNet has a low regression accuracy due to the lack of RPN network to pre-sample the images.In this paper,we improve the border regression performance of RetinaNet by GIo U loss,and reduce the sensitivity of the model to scale changes,resulting in a 1.2% improvement in the performance of the model.(5)By improving the edge selection method and edge regression of RetinaNet,the detection accuracy of RetinaNet for hazardous materials is improved from 52.29% to 56.34%,which is a 4.05% improvement.Meanwhile,it is experimentally verified that the detection speed of the model before and after the improvement remains basically the same for images.Therefore,this paper makes the detection accuracy improved while ensuring the detection speed,which is more suitable for real X-ray security screening application scenarios.