Research On Recognition Algorithms Of X-ray Images Of Container Cargo

X-ray imaging is a necessary means for custom container inspection.Intelligent image inspection is an inexorable trend.X-ray cargo image classification is a fundamental work to achieve intelligent image inspection.In order to classify cargo images regionally,we study image segmentation algorithms using graph-based methods and invent a segmentation method based on DGM(Dynamic Graph Merging).Experiments on visible images and X-ray cargo images obtain comparable or better results.As a future work,we will look into ways of applying the proposed segmentation method based on DGM to cope with the challenging task of multi-class cargo container inspection.Using traditional classification methods,we develop a classification algorithm using ensemble of exemplar-SVMs which incorporates WTA-hash and HS Code semantic tree.We select representive typical image patches from each class and train an exemplar-SVM for each of them.A fast detection algorithm based on WTA-hash is applied to determine which typical patch an image block under test belongs to.Furthermore,a balance between specificity and accuracy of classification can be achieved by the HS Code semantic tree.A supervised deep learning method based on 152-Res Net is developed.The supervising weights are adaptive to the relativity between the classification logits extracted from the supervised layers and output layer.A new clustering loss is defined to strengthen the cross entropy loss.We also propose two ways to calculate 2-digit,4-digit and 6-digit HS Code classification logits from 8-digit HS Code classification logits and add the corresponding loss to the total loss function,and classification results at diffent levels are obtained.Experiments display higher accuracy for our method.A manifest verification algorithm using triplet network is developed.A new distribution loss and an improved triplet loss are presented to improve the performance of triplet network.In order to meet the need of real-world applications,online training and negative sample training are designed.Online training can compensate for the weakness of offline feature dataset construction so that our algorithm is applicable to variable application system and can get more and more adaptive in use.Negative sample training can increase the accuracy of detection smuggling items in cargo.Finally the concept of foreign object detection and its corresponding algorithm based on manifest verification are figured,due to the fact that customs is very concerned about those small entrainments.Because of the complexity of the cargo image and the particularity of the custom inspection,we feel that grammar and low-level semantic verification method is not enough for entrainments prediction,and high-level semantic and pragmatics verification method is proposed in the outlook.