Due to the recent development of deep-learning-based computer vision technology, AI-based X-ray inspection system development is attracting attention. Object detection and semantic segmentation are applicable computer vision techniques for X-ray inspections. Specifically, semantic segmentation has the advantage of obtaining detailed experimental results compared to object detection, given its ability to undertake pixel-level classification of images.

Representative network structures for semantic segmentation include UNet [9], FPN [10], DeepLabV3+ [11], and HRNet-OCR [12]. UNet is a network structure composed of an encoder that generates representation features and a decoder that generates segmentation results using the representation features. It is characterized by how it combines the feature maps of the decoder and the encoder, which have identical resolutions, to produce good segmentation results. FPN performs well for multi-scale images by performing multi-scale feature fusion using a feature pyramid network structure. DeepLabV3+ uses atrous separable convolution, which combines depth-wise separable convolution and atrous convolution, showing a good segmentation performance while dramatically reducing the number of parameters and the amount of computation. HRNet-OCR is characterized by its ability to generate segmentation results considering object regions by utilizing object-contextual representation, showing state-of-the-art performance on many segmentation benchmarks. However, for this AI-based system, there is a limitation in that high-quality X-ray image datasets of large size must be available to achieve good results [5,13]. In addition, because most of the publicly available X-ray inspection benchmarks are X-ray images for baggage inspection purposes, another problem is that they cannot be used for cargo inspections with different X-ray imaging energy levels [5,14,15,16,17,18,19].

The first study to introduce a publicly available benchmark for cargo inspections was MFA-net [13], which provides the CargoX dataset, compiled by synthesizing prohibited items (e.g., knives) after randomly cropping cargo X-ray images. More specifically, the CargoX dataset is a benchmark for cargo inspections created by inserting prohibited items into random locations. However, because the CargoX dataset does not include X-ray images of nuclear items and only one banned item is synthesized per image, it cannot readily detect overlapping items. Therefore, the CargoX dataset is not suitable for nuclear item detection. In this study, multiple X-ray images of nuclear items are inserted into a cargo X-ray image but randomly rotated, randomly scaled, and randomly located to respond to more diverse scanning environments and item changes.

X-ray datasets of high quality and large size are needed to train a semantic segmentation model for the purpose of X-ray cargo inspections. However, X-ray images from cargo inspections are rare. In particular, X-ray images of nuclear items taken with a cargo inspection scanner and accurate label information are not publicly available. Therefore, in order to prevent the illegal transfer of nuclear items in the future, it is crucial to create a synthetic dataset using data augmentation and train a semantic segmentation model.

Various augmentations for image classification have been introduced [20,21,22,23]. Cutout [22] is a simple regularization method that randomly masks patches from the input image. Cutout allows the model to see the entire area of the input image rather than a specific area. AugMix [23] is a method that improves the robustness and uncertainty and mixes the results of data augmentation in a convex combination to prevent image degradation while maintaining diversity. These augmentations are mainly used for encoding invariances to data transformations and are well-suited for image classification.

Various augmentation methods have been introduced for semantic segmentation, most notably CutMix [24] and Copy-Paste [25]. CutMix [24] replaces the area to be mixed with a patch of another image to solve the problem of Cutout augmentation, significantly reducing the informative pixels of an image. With CutMix, every pixel within the image becomes informative, and the benefits from local dropout are obtained. Copy-Paste [25] is a way to paste objects from one image to another to create many new combinations of training data. These augmentations are used for semantic segmentation of general images, meaning that there is a limitation in that they cannot respond to the X-ray transmittance according to the physical properties of overlapping objects.

To train a segmentation model that performs automatic cargo inspections without cargo X-ray images, a means of synthesizing the textures of backscatter X-ray (BSX) images using a GAN was also introduced [14]. However, it is assumed that general cargo X-ray images and X-ray images of nuclear items taken by the same high-energy electromagnetic radiation used for cargo inspections are held. Hence, artificial texture generation is not required.

It is crucial at present to develop a semantic segmentation technique for X-ray cargo inspections in order to prevent the illegal transfer of nuclear items. While X-ray datasets of good quality and a large size are required to train a deep-learning-based segmentation model, X-ray images from cargo inspections are rare. Specifically, X-ray images of nuclear items loaded into a cargo are almost impossible to obtain.