Research On X-ray Restricted Items Monitoring Algorithm Based On Improved Cascade Mask R-CNN

Security inspection is an important means to protect the personal safety of passengers and prevent dangerous goods from entering the freight channels.The quality of security inspections directly affects the development of many industries such as transportation and logistics management.With the popularization of digitization and information technology,it is of great significance to reasonably use computer technology to assist security inspectors to complete security inspections such as the identification of contraband and dangerous goods.At present,the main method for checking whether there are restricted items in the package is to use an X-ray security inspection machine to cooperate with the staff to visually inspect the X-ray security images.The interference of many subjective factors not only increases the difficulty of discrimination,but also adds uncertainty to the inspection results.In the security scene,if image processing technology combined with a convolutional neural network model can be used to implement image segmentation tasks with high accuracy and fast response speed,it will help security inspectors to quickly and accurately judge each package and find the restricted products in the package.Restricted items are placed in different positions,so their imaging angles are not only different,but also easy to overlap with other items,which leads to the existing restricted products identification methods based on X-ray security images have many problems such as low processingefficiency,serious missed detection and false alarm.In addition,few people currently consider instance segmentation algorithms in this issue.If the computer provides visualization results of image classification,object detection,or semantic segmentation tasks,in many cases,the security inspectors can not directly see the shape and specific location of the restricted product object,which affects the efficiency of taking out the restricted product from the package in actual work.In this paper,two improved Cascade Mask R-CNN models are combined,and a restricted items instance segmentation algorithm based on X-ray security images is proposed.Firstly,by taking the intersection of the detection frame of the target object and the smallest circumscribed rectangle,the rough segmentation labels in the data set are optimized,so that the background area in the initial weak label is effectively reduced,and the features of the region of interest are highlighted.Then the unlabeled and weak labeled images in the dataset will be weighted and fused with the precise labeled images to increase the number of samples of overlapping items,which could largely increase the recognition ability of network to overlapping areas.The author design and implement two network models.One is a hybrid task cascade network formed by introducing mask information flow for Cascade Mask R-CNN,and the other is the second generation of deformable convolution networks used on the last layer of the backbone network without changing the structure of Cascade Mask R-CNN.The network uses the results of the model fusion as the final results of the method in this paper.Finally,during the training phase of the network,a random multi-scale transformation is performed on the input image within a certain range to simulate the different sizes of the same object in real life.Multi-scale testing strategy is adopted to optimize the accuracy of the model.Experiments were performed on an X-ray security image data set containing 16,138 images.The experimentalresults confirmed that the strategies selected in this paper improve the accuracy of the model,and finally achieved an accuracy of 76.12%.