Augmented CycleGAN With an Improved Loss Function For Generative method of Security X-Ray Images in Security For knives

Abstract

In recent years, as the research field of deep neural networks has become popular, more and more researchers are focusing on the automatic detection of dangerous goods in security inspection X-ray images. However, sample imbalance in the datasets is one of the difficulties in this research. In particular, the number of knives is small, resulting in data imbalance. To solve the above problem, we put forward the double-cycle consistency loss function and an enhanced CycleGAN image generation method, which can transform natural images of dangerous cutters into security inspection X-ray images as an enhancement of the dataset data. Not only can that enrich the diversity of samples in the public dataset and the diversity of poses, but also solve the problems of unbalanced data samples. Finally, we compare and evaluate the images generated by our model with the CycleGAN model, and confirm that our method can generate high-quality security inspection X-ray images with better effect than the CycleGAN model. We also use the FID index to evaluate the quality of images generated by the model, which proves that our model has good performance.