Research On Fast Semantic Segmentation Method Of Prohibited Items Image In X-ray Security Inspection

The identification of prohibited items plays a very important role in ensuring public safety.In order to reduce the missed detection and false detection caused by human detection,many researchers have done a lot of work in the field of X-ray security inspection of prohibited items automatic detection.Semantic segmentation technology is the fastest-growing branch in the field of computer vision,and its theoretical research has become increasingly mature.The semantic segmentation method is used to realize the automatic detection of prohibited items,and the shape,category and location of the target object can be obtained.However,in the task of identifying prohibited items in security check,it is faced with the difficulty of complex and changeable images and the huge difficulty of algorithm calculation.In response to the above problems,this thesis focuses on the study of real-time semantic segmentation methods for X-ray security screening of prohibited items images.The specific research content is as follows:(1)In order to reduce the loss of real semantic information in the process of downsampling,an N-type encoder is built for single target recognition task with simple background.According to the characteristics of X-ray security images,the utilization rate of shallow semantic information by network is improved,and the supervision of downsampling by network is increased to improve the recognition accuracy.(2)In this thesis,an improved encoder based on 16 layers Visual Geometry Group Network(VGG16)is built to quickly identify single objects in a simple background.The Dilated Convolution Module(DCM)is used to replace the traditional Convolution layer to improve the network perceptive field.At the same time,Asymmetric Convolution Module(ACM)is used to replace the convolution of series to reduce the computation.(3)Aiming at the multi-object recognition task under simple background,the attention mechanism is introduced,and the coding module of Atrous Spatial Pyramid Convolution(ASPC)is designsed to enlarge the network perceptive field and enhance the information mining of multi-object images.At the same time,1×1 convolution is used in the upsampling stage to reduce the channel dimension and reduce the computation.(4)For the multi-target fast recognition task under complex background,the model retains the ASPC module and ACM module,and at the same time,the 1×1 convolution dimension reduction operation is used in the upsampling stage.It balances the recognition accuracy and running speed of the algorithm.The research results of each stage show that the recognition algorithm designed based on ASPC module and ACM module has a good performance for the rapid detection of various targets in a complex background,combining with the characteristics of the security inspection contraband images.