Research On Security X-Ray Item Image Generation Based On GAN

Recognizing prohibited items automatically is of great significance for intelligent X-ray baggage security screening.Convolutional Neural Networks(CNNs)have been verified as the powerful models in image content analysis.However,the current X-ray baggage image databases suitable for CNN training are not enough in sample quantity and sample variety.To build a large database of security X-ray item images,an image generation method based on Generative Adversarial Networks(GANs)is studied in this thesis.The main contributions are as follows:1)An X-ray item image preprocessing method is proposed.The KNN matting and Cycle GAN model are applied to extract the item foreground interested from collected security X-ray images.Then,the poses of the extracted foregrounds are estimated using a space rectangular coordinate system,and categorized into 4 or 8 classes for building a training database.2)A new GAN model suitable for X-ray item images is constructed by some improvements for Deep Convolutional Generative Adversarial Network(DCGAN).Different GAN models are evaluated using Frechet Inception Distance(FID)scores.Then,15 classes of new X-ray item images with good visual quality are generated by the proposed GAN model.3)Two methods are proposed to enrich the diversity of generated images.First,the improved pix2 pix model is used to change the item poses in generated images.Then,the Cycle GAN model is used to transform the natural images of items into X-ray images.4)The CNN-based evaluation method for generated images is proposed.A cross-validation experiment is implemented based on CNN model,while the real images and generated images are respectively used as training set and testing set.Experimental results show that most generated images can be recognized correctly by CNN model.Furthermore,the generated images are verified that they can improve the CNN training.