A VQ-VAE Based Recognition Algorithm for Terahertz Security Inspection Images

Terahertz imaging technology has been used in the field of security inspection due to its excellent perspective ability. However, the terahertz security inspection images have the problem of poor image quality and cannot cover all types of dangerous objects. In this paper, a terahertz image recognition algorithm based on Vector Quantized-Variational Autoencoder (VQ-VAE) is proposed. The algorithm realizes image data enhancement through dynamically learned prior, and discrete potential space is used to efficiently identify abnormal objects outside the training set. In order to improve the network recognition performance, the attention CBAM module is introduced to increase the attention ratio of the key target area. The algorithm embeds Center Loss metric learning module into the network frame to increase the inter-class distance and reduce the intra-class distance, thereby improving the clustering and generalization performance. The experimental verification is carried out on the actual collected terahertz security image data set. The experimental results show that the improved and optimized VQ-VAE image recognition network is significantly better than the convolutional neural network in recognition accuracy and has better generalization ability.