Super-Resolving X-Ray Security Images with Variational Autoencoder for Prohibited Item Detection

Most existing super-resolution methods are designed for natural visible light images, relying on subjective visual perception and objective image quality metrics to evaluate their performance. This paper introduces a super-resolution network model, XSR-VAE, specifically designed for pseudo-color X-ray security inspection images. XSR-VAE consists of three main components: Feature Extraction, channel-attention adaptive instance normalization (CAdaIN), and Upsample OneStep. Specifically, the variational autoencoder architecture is employed to construct the prior and posterior distributions of the input. The novel channel-attention adaptive instance normalization is designed to enhance details related to the color and contrast of X-ray security images. Additionally, the denoising function is integrated into the super-resolution process. We conduct experiments on three tasks: image contrast enhancement, image denoising, and image super-resolution. Furthermore, we evaluate the performance of the super-resolution X-ray security images generated by XSR-VAE in detection tasks. The experimental results demonstrate that XSR-VAE effectively enhances image contrast, removes noise, and improves the detection accuracy of prohibited items in object detection tasks.