Pattern Recognition Of Typical Prohibited Items In X-Ray Images Based On Deep Learning

Nowadays,X-ray security inspection machines mainly rely on the experience of security inspection staff to identify prohibited articles by observing X-ray images with naked eyes.Only in some areas,semi-intelligent recognition assisted by artificial intelligence algorithms has been realized.However,because the work itself is boring,and people will be fatigued or affected by some external factors,it is inevitable that there will be wrong or missed inspection,which may lead to unimaginable consequences.Therefore,the complete intelligent recognition of X-ray security inspection machines has always been a key proposition in the field of public security.In this paper,X-ray images as the research object,automatic recognition of the typical prohibited items in X-ray images as the purpose,to launch the research of this topic.The main research contents and achievements are as follows:(1)Understanding the current research status of the identification of prohibited items and determineing the research direction By consulting relevant literature,I have learned the current research status at home and abroad.The identification method of prohibited items is summarized into two major categories:traditional X-ray image recognition methods and X-ray image recognition methods based on deep learning.The traditional method has the disadvantages of poor generalization ability and low accuracy,so it is difficult to meet the requirements of detection.The recognition method based on deep learning does not need too much human intervention,and the computer learns autonomously based on the convolutional neural network and extracts enough rich features.Good results can be achieved.Therefore,a method based on deep learning for the identification of typical prohibited items in X-ray images is proposed in this paper.(2)Learning the imaging principle of X-ray security machine and completing the production of X-ray image datasets.Going to Shenyang Ditai Inspection Equipment Limited Company to learn the imaging principle of X-ray security inspection machine,and to complete the collection and preprocessing of the dataset.According to the principle of X-ray imaging,a data enhancement method based on X-ray energy value is proposed.And combined Mosaic data enhancement and traditional data enhancement methods to complete the production of the X-ray image VOC2007 format datasets.(3)Learn the relevant theoretical knowledge of deep learning,determine the basic model of deep learning,and optimize it to achieve the purpose of automatic identification of typical prohibited items.According to the characteristics of existing deep learning models and the requirement of high accuracy in this study,the Faster R-CNN model is determined.The experimental data were analyzed,and the feature extraction network was replaced in combination with the characteristics of dense and disintegrated X-ray images and many small targets.The multi-layer feature fusion was carried out using FPN,Roialign optimized Roipooling,soft-NMS optimized NMS,Diou loss function optimized boundary box regression loss function and other improvements were made.The model used in this paper is proposed.After comparing experiments with existing models,it is verified that the model in this paper has the highest accuracy,and the mAP value is 90.45%.a The higher robustness of this model is also verified both by changing the training set ratio and the IoU threshold.In addition,the detection speed of the model is 380ms per image,which means an FPS of 2.6.Therefore,the model can meet the needs of X-ray safety inspection machine to automatically identify typical prohibited items.