Research On Multiple Prohibited Items Synthesis Method Of X-ray Security Image

Automatic detection of prohibited item based on Convolutional Neural Network(CNN)is of great significance to improve security check efficiency.However,the appearance of X-ray security images with prohibited items is a small-probability event,which leads to the diversity and quantity of the existing security image dataset unable to meet the requirements of reliable CNN model training.In order to build the large security image dataset,we propose a method of synthesizing X-ray security image with multiple prohibited items from semantic label images basing on Generative Adversarial Network(GAN).The main contributions are as follows:1)A method of expanding the semantic label image is proposed to synthesize as many X-ray security images as needed.First,Self-Attention Generative Adversarial Network(SAGAN)is used to generate a large number of single-object semantic label images with different posturing from noise.Then,a method of Semantic Image Projection(SIP)is proposed to synthesize a large number of multi-object Semantic label images.2)An image synthesis network of Multiple Prohibited Items without occlusion based on GAN is constructed.A generator based on Res2 Net is designed to solve the problem of poor synthesis effect caused by different sizes of prohibited item images.A multi-scale discriminator is constructed to achieve high resolution image synthesis.Feature matching loss is introduced into the objective function to improve the stability of model training.3)An image synthesis network of occluded Multiple Prohibited Items based on mixed attention is constructed.Mixed attention is composed of channel attention and spatial attention in parallel structure.Considering that the mapping of occluded images is more complicated,the channel attention module is used to improve the efficiency of feature learning.Spatial attention enhances the global dependence of the generator to solve the problem of blurred boundaries of overlapping regions in the synthesized image.4)A synthetic image evaluation method based on deep CNN recognition algorithm of is proposed.By comparing the recognition effects of a Fully Convolutional Network(FCN)trained by real images and synthesized images,the effect of the synthesized image training model is evaluated.The dataset of the Single Shot Multi Box Detector(SSD)is expanded with the synthesized image to verify the data augmentation effect of the synthesized image.The experimental results show that the synthetic image can effectively train the recognition algorithm and solve the problem of insufficient positive samples and small number of X-ray security datasets.