Prohibited Items Detection In X-ray Images Based On Multi-scale Feature Fusion SSD

At present,most of the security check work in the actual scene still needs to be carried out manually.On the one hand,manual detection is not efficient,on the other hand,mistakes will inevitably occur in the state of human exhaustion.Therefore,it is of great social significance to realize the automatic detection of prohibited items in X-ray security check images for improving the quality of security check.Although some domestic and foreign scholars have studied the relevant contents in the field of prohibited items detection of X-ray images,how to further improve the detection effect of prohibited items is still the key content that we need to study in the future.Therefore,in order to improve the detection accuracy of prohibited items,in this thesis,proposes two improved algorithms based on convolutional neural network.(1)To solve the problem of poor detection effect of small-scale prohibited items,a multiscale Feature Fusion Network(MFFNet)for prohibited items detection of X-ray images is proposed.This network is based on SSD model.And a deeper feature extraction network,Res Net-101,is adopted.Two lightweight feature fusion modules are added to integrate the highlevel features and the low-level features of the network through jump connection,which adds context information for small-scale prohibited items and effectively improved the detection accuracy of all kinds of prohibited items,especially small-scale prohibited items,without greatly sacrificing the detection speed.(2)Based on the characteristics of X-ray images,a prohibited items detection algorithm based on attention mechanism is proposed to solve the problems of complex background and disorderly placement of prohibited items in X-ray images.Based on the MFFNet network model,this algorithm has made two improvements.Firstly,an image attention module is added before the input image enters into the feature extraction network.The network can better learn the key information in the image through the attention mechanism,enhance the expression ability of image features,and effectively solve the problem that prohibited items is difficult to detect in the chaotic image background.In addition,the last two convolution layers in the network are replaced by deformable convolution,and the receptive field of target sampling is adjusted adaptively according to the shape and scale of prohibited items,so as to realize the automatic detection of prohibited items with different shapes and scales.Through the verification on Sixray＿OD dataset,compares with the existed methods,the two improved algorithms can improve the detection accuracy to different degrees while maintaining a faster detection speed,and solve some difficulties in the detection of prohibited items specifically,which proves the effectiveness of the work in this thesis.