Research On Network Models Of Security Prohibited Items Images Generation And Evaluation

In recent years,detecting prohibited items of X-ray security images automatically has attracted attention increasingly.It is of great significance to maintain public transport safety and improve security check efficiency.With the rapid development of deep learning,detecting prohibited items of X-ray security images automatically based on Convolutional Neural Networks(CNNs)has made some progress.However,security X-ray images with prohibited items are not easy to obtain,which makes it difficult to train a reliable CNN model from the existing X-ray security image database.In order to solve the problem that the X-ray security image database is small,we propose a method of generating X-ray security prohibited items image based on the Generative Adversarial Networks(GANs).The main contributions are as follows:1)Two X-ray security databases are constructed: X-ray security prohibited items image database and X-ray security image database.In addition,those two X-ray security databases are preprocessed and the characteristics of two databases are analyzed.2)An X-ray prohibited items images generation network based on the improved GAN model is constructed.By improving the network structure and loss function,the improved GAN model can be better applied to generate X-ray security images.Combining with the qualitative and quantitative analysis of the real X-ray security images database,the performance of different GAN models is judged.The improved GAN model is used to generate a large number of high-quality and diverse X-ray security prohibited items images and X-ray security images.3)The mapping transformation model of natural image and X-ray image is constructed.Natural images are converted from natural imaging space to X-ray imaging space based on the improved Cycle GAN framework.This method is used to generate many X-ray images with a wide variety of shapes and posture.4)Two method for evaluating generated images are proposed.The first method uses the GAN-train and GAN-test indicators to evaluate the generated images of authenticity,diversity and feature information so that the generated image whether can be used to expand the database.The second one is based on the object detection network model.This method can determine whether the generated images have data augmentation effect.