Detection Of Contraband In X-ray Image Based On Anchor-free Network

At present,the main security means is to check whether there are contraband in packages and luggage through X-ray images,and a lot of inspection work still depends on manual work.In order to reduce the work intensity,the development of auxiliary detection system has become an inevitable trend.However,at present,the distance of automatic identification of contraband is mature,and there is still a lot of work to be completed.This thesis attempts to apply Center Net,a target detection algorithm without anchor frame,to the auxiliary detection of contraband,and puts forward an improved model combined with the actual engineering scene and the characteristics of X-ray image.(1)Aiming at the problem of large amount of calculation and redundant parameters of Center Net model,the model structure used by the algorithm is adjusted,and lightweight Shuffle Net-Center Net and Mobile Net-Center Net are designed.Firstly,the backbone network is replaced by Shuffle Net V2 and Mobile Net V2 with compact parameters,and the feature pyramid network(FPN)is used to fuse multi-layer features to reduce the overall parameters and calculation of the model.After lightweight,the amount of calculation and parameters of Shuffle Net-Center Net and Mobile Net-Center Net are only one third of Res Net18-Center Net.Secondly,although the amount of calculation of the original model is huge,it has strong learning ability and can contain the knowledge that the lightweight model cannot learn.Therefore,channel distillation is introduced,and Hourglass-Center Net model is used to guide the training of Shuffle Net-Center Net and Mobile Net-Center Net,so as to increase their accuracy by 2%(2)Aiming at the problems of various horizontal and vertical ratios and different sizes of contraband,an improved Hourglass-Center Net model based on pyramid convolution and strip pooling is proposed.Firstly,based on Hourglass Network,pyramid convolution is introduced to enrich the receptive field of the backbone network,extract the mesoscale invariant features of the contraband image,and enhance its multi-scale feature extraction ability.Secondly,due to the characteristics of X-ray imaging,the real frame is often large horizontally and vertically in the image.By introducing the banded pooling module,taking into account the local detail information,it provides contextual global information for the result prediction,and its unique banded pooling core can prevent the interference of irrelevant areas.The prediction accuracy of the improved model is improved from 86.6% to 88.0%.