| With the continuous increase of the population density and the travel demand,X-ray dangerous objective detection(X-DOD)has been one of the necessary equipment in major institutions for safety reasons.However,the current safety inspection is mainly based on the manual inspection,whose efficiency strongly depends on the experience of the staff,and it is prone to misdetection and has a great potential safety hazard.As a result,this work aims to apply the state-of-the-arts deep-learning objective detection method(YOLOV4)to XDOD to promote the automation process of the safety inspection.The main research contents are presented as follows.(1)According to the related datasets of X-DOD,the dangerous objective is often covered and its number is small,which hinders the deep learning methods to extract the effective features.Hence,we have designed a new model DU by combing the “U”-shape structure of U-Net and YOLOV4,which improves the representation ability of the network and increases the accuracy of objective discrimination.In addition,we have proposed a data repartitioning loss function S-T CHM-C,which is based on the weight gradients and further solves the problem of the object coverage.The experiments show that compared with YOLOV4,The m AP of our algorithm has been better improved on the OPIXray dataset and the SIXray dataset.(2)Aiming at the problem that small objects are difficult to distinguish,we have designed a X-DOD algorithm based on the multi-scale feature fusion to improve the backbone of YOLOV4,where the spatial feature fusion module SFF uses the fusion features of the space and the channel to reduce the misdetection,and the channel feature fusion module CFF is used to solve the problem of the object coverage.The experiments show that compared with YOLOV4,The m AP of our algorithm has been better improved on the OPIXray dataset and the SIXray dataset.(3)Based on the second method,we have further proposed a new feature extraction module MS-SFF to extract edge information and contour information,effectively improving the utilization efficiency of the shallow features.Moreover,we have introduced the BIFPN module to perform the fast multi-scale fusion and added a new penalty item on the original loss function to focus on the center point difference between the predicted objective boxs and the ground-truth ones.The experiments show that compared with YOLOV4,The m AP of our algorithm has been better improved on the OPIXray dataset and the SIXray dataset.It is worth noting that the above three methods all improve the detection accuracy of X-DOD to varying degrees.It further promotes the automation of the security inspection process,saves human resources and improves the security inspection performance. |
