Semantic Segmentation Of Security Inspection Image Based On Full Convolutional Network

The security check is crucial to protect the safety of public spaces.However,most detection tasks rely mainly on humans,which could lead to the missed detection and false detection.In the recent years,computer vision technology has developed rapidly,driving changes in all walks of life.Therefore,we try to use computer vision technology to assist security staffs and achieve automated detection.In this thesis,the detection of prohibited items is regarded as a semantic segmentation task.In order to improve the semantic segmentation network,the joint attention module is embedded into the network,and the upsampling module is further improved.The main contributions in this thesis are as follows:1)The semantic segmentation model of prohibited items is built.The Deeplab network is used as the basic network framework,and the feature extractor is replaced by a lightweight network to ensure the model parameters are not too large.2)The joint attention is proposed to improve the segmentation network.The channel attention is used to weight the channels,while the spatial attention is used to weight the space.The network would pay attention to important information while ignoring other irrelevant information,which can improve the segmentation accuracy.3)The joint pyramid upsampling module is introduced.In the process of upsampling,the low-level semantic information is used.The joint pyramid upsampling module can learn different scale information to make the feature fusion more sufficient and further improve the segmentation accuracy.The semantic segmentation method in this thesis not only achieves the classification of prohibited items,but also locates the prohibited items,which greatly facilitates the open-pack inspection.From the perspective of experimental accuracy,our method could achieve super accuracy,which is a relatively successful exploration of automated security.