Research On X-ray Image Contraband Detection And Recognition Method Based On Supervised Contrastive Learning And Cross-distillation

X-ray security inspection is an important method to protect human’s live.With the rapid development of the transportation,there are more and more scenarios of security inspection applications.But there is a lack of a method that can effectively improve the efficiency of security inspection.The detection of contraband in X-ray images has great significance for protecting public safety and improving traffic efficiency.With the development of science and technology,more and more studies have been carried out in the field of X-ray images contraband detection,but these studies cannot effectively solve the problems existing in X-ray image contraband detection.Combined with the characteristics of X-ray images,a X-ray contraband detection and continuous learning algorithm based on supervised contrastive learning is proposed.Aiming at the problems of small differences and overlapping occlusions among the categories of contraband in X-ray images,a detection method of contraband in X-ray images based on supervised contrastive learning and weighted boxes fusion is proposed.First,a supervised contrastive branch is designed at the head of the classification network to optimize the feature vectors of region proposals,which solves the similarity problem between contraband,and the Io U aware function is used to optimize the supervised contrastive loss,which further improves the classification performance;secondly the weighted boxes fusion method optimized by the Top K algorithm is used to perform multi-boxes fusion,which effectively solves the problem of repeated detection caused by overlapping occlusions and improves the detection performance.Aiming at the problem of increasing categories of contraband in X-ray images,the thesis proposes a continuous learning method based on cross-distillation and multi-level memory replay.First,the cross-distillation module generates a classification layer for each new task,uses the distillation loss function for the old task,improves the memory ability of the old tasks,and uses the cross-entropy loss function for the new tasks to learn the knowledge in the new tasks,which improves the continuous learning.Then,multi-level memory replay is used,including three parts: coarse-grained batch-level replay strategy,balanced memory update strategy,and final replay strategy.The coarse-grained batch-level replay strategy avoids frequent memory retrieval and update.Secondly,the balanced memory update strategy balances the number of categories in the memory buffer during memory update to avoid the problem of data imbalance;finally,the final playback strategy is used.After all task training is completed,the data in the memory area is perform full replay to improve test performance.The results show that the methods has great detection ability and continual learning ability.Compared with the mainstream target detection algorithm and continual learning algorithm,the methods in the thesis all have certain advantages.