| With the improvement of social security standards,access control of some environments such as public transportation,museums,and event venues has become increasingly strict.X-ray imaging technology-based security inspection equipment plays a key role in protecting personnel and venues from the risks of crime and attack:because of the penetration of X-ray,objects inside the package can be observed in the X-ray image of the package,and then these objects can be checked for prohibited items.However,for objects that overlap within the package,their X-ray images will also overlap,making it difficult for security personnel to identify them,resulting in false or missed detection of some prohibited objects,posing a hidden danger to social security.In response to this overlapping problem,the thesis proposes a ViT based method for separating overlapping objects in X-ray images,which can process the X-ray images of overlapping objects into X-ray images of each individual object.The main work and innovations of the thesis are as follows:(1)An X-ray image dataset named Mask Xray was created to fill the gap in the solution of X-ray image separation problem in training data.After selecting and reprocessing existing security inspection images,a dataset with 50000 training data was created,each of which contains X-ray images of overlapping objects,X-ray images of individual object and corresponding trimap images;(2)A GAN model based on ViT was proposed for separating overlapping objects in X-ray images.The ViT structure was used to implement the encoder network and decoder network of the generator.Based on the prior information provided by the trimap image,the generator can differently learn the overlapping and non-overlapping regions in the input image,so that the generated X-ray images of individual object has both texture features of overlapping regions and color features of non-overlapping regions;Convolutional neural networks and fully connected networks were used to implement the discriminator to help the generator to train more effectively.The experimental results and comparison results show that the model can effectively separate overlapping objects in X-ray images;(3)A network model for end-to-end separation of overlapping objects in X-ray images was proposed.A Mask RCNN model that has completed the instance segmentation task on the coco dataset was trained using the reprocessed Mask Xray dataset,enabling it to distinguish between overlapping and non-overlapping regions in X-ray images,thus achieving a mask images generation network that outputs trimap images corresponding to the input image.The network was combined with the separation network that requires the trimap image to obtain a network model with endto-end separation effect.The experimental results show that the model can achieve the effect of end-to-end separation of overlapping objects in X-ray images,and the separation effect is satisfying.(4)A graphical user interface for the separation network model and end-to-end separation network model was designed and implemented,meeting the convenience requirements of users for separation operations and improving the feasibility of the proposed method in practical applications. |
