Detail Restoration and Tone Mapping Networks for X-Ray Security Inspection

Abstract

X-ray imaging is one of the most widely used security measures for maintaining airport and transportation security. Conventional X-ray imaging systems typically apply tone-mapping (TM) algorithms to visualize high-dynamic-range (HDR) X-ray images on a standard 8-bit display device. However, X-ray images obtained through traditional TM algorithms often suffer from halo artifacts or detail loss in interobject overlapping regions, which makes it difficult for an inspector to detect unsafe or hazardous objects.To alleviate these problems, this article proposes a deep learning-based TM method for X-ray inspection. The proposed method consists of two networks called detail-recovery network (DR-Net) and TM network (TM-Net). The goal of DR-Net is to restore the details in the input HDR image, whereas TM-Net aims to compress the dynamic range while preserving the restored details and preventing halo artifacts. Since there are no standard ground-truth images available for the TM of X-ray images, we propose a novel loss function for unsupervised learning of TM-Net. We also introduce a dataset synthesis technique using the Beer-Lambert law for supervised learning of DR-Net. Extensive experiments comparing the performance of our proposed method with state-of-the-art TM methods demonstrate that the proposed method not only achieves visually compelling results but also improves the quantitative performance measures such as FSITM and HDR-VDP-2.2.INDEX TERMS Convolutional neural network, high dynamic range, tone mapping, unsupervised learning, X-ray imaging. HYO-YOUNG KIM (Member, IEEE) received the B.S. degree in electrical engineering from Korea University, in 2013, where he is currently pursuing the Ph.D. degree. His current research interests include image processing, computer vision, and artificial intelligence. SEUNG PARK received the B.S. and Ph.D. degrees in electrical engineering from Korea University, Seoul, South Korea, in 2013 and 2020, respectively. He was a Research Professor with Korea University, in 2020. He currently joined the Samsung Medical Center as an AI Researcher. His research interests include X-ray imaging and medical image analysis.

Human Visual System Model-Based Optimized Tone Mapping of High Dynamic Range Images

“…for H and H (k+1) T , respectively, using (26) and (27). Finally, we obtain the TMC T (k+1) (m) for the input pixel value m by…”

“…Recently, inspired by the success of deep learning in various image processing tasks, several deep learning-based tone mapping techniques that use convolutional neural networks (CNNs) or generative adversarial networks (GANs) have been developed [23]- [27]. For example, Zhang et al [23] developed a tone-mapping network by employing the improved Wasserstein GAN and constructed a dataset of HDR images and the corresponding tone-mapped images, manually adjusted by photographers.…”

“…Panetta et al [26] further improved the performance of the GAN-based tone-mapping algorithm by developing an attention-guided generator architecture. Kim et al [27] proposed a detail-preserving tone-mapping algorithm that consists of two CNNs; one restores the details in the input HDR image, and the other compresses the dynamic range. However, the main disadvantage of the learning-based approaches is that their performance is highly dependent on the training datasets.…”

“…Although conventional tone-mapping algorithms have attempted to obtain high-quality results [10]- [27], little effort has been made to consider the characteristics of the display devices on which the tone-mapped images are presented. The display-adaptive tone-mapping algorithm [12] is computationally inefficient, because the computation of the HVS response to the input HDR image requires high computational complexity.…”

“…The HVS response model-based tone-mapping algorithm [16] uses a guided tone-mapping curve (TMC) to preserve the absolute luminance that may cause detail loss in the bright regions. We also note that, although recent deep learning-based algorithms have shown significant tone-mapping performance improvements [23]- [27], they too have been developed without considering the characteristics of the display devices.…”