HiXray: Security Benchmark & Diffuse Scanning

• HiXray is a polysemous term that refers both to a high-quality airport security X-ray benchmark and to a high-resolution scanning method employing Huffman-patterned diffuse probes.
• The security benchmark version of HiXray includes over 45,000 RGB scans with detailed bounding box annotations, enabling robust evaluation of detection protocols using metrics like mAP and advanced modules such as LIM.
• The diffuse-probe scanning approach of HiXray minimizes radiation dose by spreading exposure with structured masks and uses deconvolution techniques to reconstruct sharp images, balancing resolution with computational accuracy.

HiXray is a polysemous term in recent X-ray imaging literature. In security inspection, it denotes a high-quality benchmark of real-world airport baggage X-ray images for prohibited-item detection, introduced together with the Lateral Inhibition Module (LIM) for boundary-aware detection under overlap and clutter. In computational X-ray imaging, it also denotes a high-resolution scanning method that uses a diffuse, Huffman-patterned probe to spread dose spatially and recover sharp images by deconvolution. These two usages are technically unrelated, but both have become reference points in their respective subfields (Tao et al., 2021, Aminzadeh et al., 2024).

1. Terminological scope and research contexts

In the security-imaging literature, HiXray was introduced as a High-quality X-ray security inspection image dataset containing real-world airport scans and professional bounding-box annotations for prohibited items (Tao et al., 2021). Subsequent comparative studies use HiXray as one of the standard public benchmarks for X-ray illicit-object detection, typically alongside datasets such as EDS and PIDray (Cani et al., 1 May 2025).

In a distinct line of work, HiXray is the abbreviated name of “High-resolution x-ray scanning with a diffuse, Huffman-patterned probe to minimise radiation damage”, a coded-scanning framework in which a broad structured beam is rastered across an object and the resulting bucket signals are decoded to reconstruct a sharp image (Aminzadeh et al., 2024).

A common source of confusion is to treat all occurrences of “HiXray” as referring to the same system. The supplied literature does not support that usage: one body of work concerns airport security inspection and object detection, while the other concerns coded-aperture-like X-ray scanning and dose management.

2. HiXray as a real-world security-inspection benchmark

The HiXray dataset was collected from the carry-on luggage of an international airport’s security X-ray machines, and every image is a genuine scan rather than a staged or synthetic overlay (Tao et al., 2021). It contains 45,364 RGB X-ray scans and 102,928 bounding boxes across 8 categories, with 2.27 objects per image on average. The classes are PO1 (portable charger 1, a prismatic lithium-ion cell), PO2 (portable charger 2, a cylindrical lithium-ion cell), WA (water), LA (laptop), MP (mobile phone), TA (tablet), CO (cosmetic), and NL (nonmetallic lighter) (Tao et al., 2021, Cani et al., 23 Jul 2025).

The annotation format is axis-aligned bounding boxes with class labels. The annotations were drawn by professional airport security inspectors following a Pascal-VOC-style protocol with exhaustive labeling, consistency checks, and cross-inspection (Tao et al., 2021). The image data are described as pseudo-color RGB images from a dual-energy X-ray scanner, single-view, high-quality captures collected at an international airport (Cani et al., 23 Jul 2025). Reported image sizes are JPEG with average 1,200 × 900 px, up to 2,000 × 1,040 px (Tao et al., 2021).

The benchmark’s visual domain is materially encoded. The dataset description explicitly notes color-by-material mapping with orange for organics, blue for metals, and green for mixtures (Tao et al., 2021). In later evaluations, the dataset is characterized as single-scanner data with no explicit domain shift within HiXray, and with moderate clutter in real baggage scans but no deliberately hidden items, unlike PIDray (Cani et al., 1 May 2025). This matters because HiXray is used less as a benchmark for adversarial concealment than as a benchmark for fine-grained detection under realistic overlap, everyday baggage clutter, and material ambiguity.

3. LIM: boundary-aware feature refinement on HiXray

The paper that introduced HiXray also introduced the Lateral Inhibition Module (LIM), a plug-in feature module intended to improve prohibited-item detection in X-ray images by suppressing irrelevant information and emphasizing identifiable boundaries (Tao et al., 2021). LIM is described as comprising a Bidirectional Propagation (BP) submodule and a Boundary Activation (BA) submodule. It can be inserted between a convolutional backbone and the detection head of one-stage or anchor-free detectors such as SSD, YOLOv5, and FCOS (Tao et al., 2021).

For the top-down BP pathway, if $L$ is the total number of feature levels and $\mathcal{F}^l(x)$ is the feature map at level $l$, the reported formulation is

$$A^{l} \;=\;\mathcal{V}\bigl(\mathcal{F}^{l}(x)\bigr) \;+\;\sum_{m=1}^{L-l}\mathcal{U}^{m}\bigl(A^{l+m}\bigr)\,.$$

After BA, the bottom-up BP pathway is written as

$$C_t^{\,l} \;=\; \mathcal{V}\bigl(B^{l}\bigr) \;+\;\sum_{m=1}^{l-1}\mathcal{D}^m\bigl(C_t^{\,l-m}\bigr), \qquad C^{l} \;=\; C_t^{\,l} \;+\;\mathcal{F}^{l}(x)\,.$$

The BA submodule applies directional running-max operations to emphasize boundary structure. For the left-to-right pass on $A^l\in\mathbb{R}^{H\times W\times C}$, the paper gives

$B^{l}_{ijc} = \begin{cases} A^{l}_{iWc}, & j = W, \[6pt] \max\{A^{l}_{ijc},\, A^{l}_{i,j+1,c},\,\dots,\,A^{l}_{iWc}\}, & \text{otherwise.} \end{cases}$

The interaction of BP and BA is described in explicitly neuro-inspired terms: the top-down path filters spurious activations, BA re-activates salient edge cues, and the bottom-up path propagates the refined features while suppressing context noise (Tao et al., 2021).

Quantitatively, the reported gains on HiXray are measured by [email protected] IoU. On HiXray, SSD improves from 71.4% to 73.1%, FCOS from 75.3% to 76.9%, and YOLOv5s from 78.6% to 80.1% when LIM is added. The paper further states that, against the prior specialized module DOAM on the same backbones, LIM consistently outperforms by approximately 0.6–2.6 percentage points (Tao et al., 2021).

4. Comparative detector behavior on HiXray

Later studies situate HiXray within a broader benchmarking framework for X-ray illicit-object detection. The common detection metrics are based on Intersection over Union,

$$\mathrm{IoU}(A,B)\;=\;\frac{\lvert A\cap B\rvert}{\lvert A\cup B\rvert}\,,$$

and on mean average precision variants such as mAP@50 and mAP@[50:95] (Cani et al., 23 Jul 2025). In a focused hybrid-architecture study, the HiXray protocol uses an 80% training / 20% testing split, batch size 18, and SGD with learning rate $0.01$ and momentum $0.937$; early stopping is based on validation loss, and data augmentation follows the default Ultralytics routines of random flips, mosaic, and mixup (Cani et al., 1 May 2025).

A later large-scale comparison reports the following HiXray results for ten detectors (Cani et al., 23 Jul 2025):

Detector	mAP@50	mAP@[50:95]
YOLOv8 + CSPDarkNet53	0.845	0.564
YOLOv8 + HGNetV2	0.833	0.557
YOLOv8 + CHR	0.811	0.523
YOLOv8 + DOAM	0.830	0.545
YOLOv8 + LIM	0.828	0.525
DINO + Swin-B	0.849	0.535
Co-DETR + Swin-B	0.857	0.531
RT-DETR + HGNetV2	0.839	0.510
YOLOv8 + Next-ViT-S	0.841	0.551
RT-DETR + Next-ViT-S	0.818	0.483

These numbers establish a nuanced ranking. The highest mAP@50 is achieved by Co-DETR at 0.857, whereas the highest mAP@[50:95] is achieved by YOLOv8 + CSPDarkNet53 at 0.564 (Cani et al., 23 Jul 2025). In the more targeted hybrid-architecture evaluation, the same YOLOv8 baseline with its default backbone is reported to outperform all hybrid variants on HiXray, including object-level mAP$\mathcal{F}^l(x)$0 for every class (Cani et al., 1 May 2025).

The stricter localization metric therefore favors the CNN-only baseline on HiXray, even though a transformer detector attains the best single-threshold mAP@50. This suggests that, on this dataset, local appearance modeling and localization fidelity remain particularly consequential.

5. Dataset-specific challenges and empirical interpretation

The comparative papers describe HiXray as a homogeneous single-scanner dataset with no large cross-domain shifts, in contrast to EDS (Cani et al., 1 May 2025, Cani et al., 23 Jul 2025). This has direct architectural consequences. One study concludes that on HiXray-like scenarios the recommended choice is YOLOv8 with CSP-DarkNet53 backbone for the highest mAP and inference speed, and that hybrid CNN-transformer backbones should be reserved for cases with known domain shifts (Cani et al., 1 May 2025).

Class-level behavior is also uneven. The later comparative evaluation reports that the best detected classes are high-density, geometrically simple items, such as laptops and portable chargers, while the worst detected classes are low-density or highly variable items, such as water, cosmetic, and non-metallic lighter (Cani et al., 23 Jul 2025). It further states that portable charger 1 and portable charger 2 are often confused because of similar shape and material, and that fine-grained electronics such as phones and tablets benefit from local feature modeling in CNNs more than from global-context transformers (Cani et al., 23 Jul 2025).

Object size is not the dominant factor in relative model ranking, but it is still a difficulty axis. The object-size analysis on HiXray uses COCO definitions of small, medium, and large objects and reports that small objects show a slight dip of approximately 5–7 points relative to medium and large objects, whereas medium and large objects show nearly identical performance across detectors; the variance across scales is reported as less than 3% between the best and worst models (Cani et al., 1 May 2025). The same study concludes that small objects remain the most challenging.

Another important domain property is that HiXray does not define explicit occlusion subsets, yet ordinary baggage overlap remains common: chargers may be on phones, cables may cross devices, and low-$\mathcal{F}^l(x)$1 materials such as water and plastics produce faint signals and higher false-negative rates (Cani et al., 23 Jul 2025). The difficulty is therefore not simply heavy clutter, but the conjunction of overlap, material attenuation variability, and fine-grained inter-class similarity in a relatively clean single-scanner domain.

6. HiXray as diffuse Huffman-patterned X-ray scanning

In a separate research program, HiXray denotes a high-resolution X-ray scanning strategy that replaces a tightly focused probe with a broad, structured probe derived from Huffman sequences (Aminzadeh et al., 2024). The starting point is a 1D Huffman sequence $\mathcal{F}^l(x)$2 of length $\mathcal{F}^l(x)$3, whose aperiodic autocorrelation

$\mathcal{F}^l(x)$4

is as close as possible to zero for all nonzero shifts except for two inevitable end terms. The guiding idea is that

$\mathcal{F}^l(x)$5

with $\mathcal{F}^l(x)$6 denoting aperiodic correlation (Aminzadeh et al., 2024).

A 2D mask is formed by an outer product, $\mathcal{F}^l(x)$7, yielding a 2D pattern with strongly $\mathcal{F}^l(x)$8-like aperiodic autocorrelation (Aminzadeh et al., 2024). Because raw integer Huffman sequences can have a very large dynamic range, the authors report the use of controlled scaling, integer rounding, and small random dither perturbations to compress the entries to $\mathcal{F}^l(x)$9. Quality is monitored using RMS, mean absolute value, zero-fraction $l$0, merit factor $l$1, peak-to-sidelobe ratio $l$2, spectral flatness $l$3, and the condition number $l$4 of the convolution operator (Aminzadeh et al., 2024).

The masks are fabricated on SiO$l$5 wafers with sputtered tantalum (Ta) up to approximately 5 $l$6m thick. Two implementations are reported. Binary masks use a single 5 $l$7m Ta layer and encode gray values by subdividing each pixel into $l$8 subpixels, with pixel pitches of 8 $l$9m, 10 $$A^{l} \;=\;\mathcal{V}\bigl(\mathcal{F}^{l}(x)\bigr) \;+\;\sum_{m=1}^{L-l}\mathcal{U}^{m}\bigl(A^{l+m}\bigr)\,.$$0m, or 15 $$A^{l} \;=\;\mathcal{V}\bigl(\mathcal{F}^{l}(x)\bigr) \;+\;\sum_{m=1}^{L-l}\mathcal{U}^{m}\bigl(A^{l+m}\bigr)\,.$$1m. Quaternary masks realize four transmission levels $$A^{l} \;=\;\mathcal{V}\bigl(\mathcal{F}^{l}(x)\bigr) \;+\;\sum_{m=1}^{L-l}\mathcal{U}^{m}\bigl(A^{l+m}\bigr)\,.$$2 through three successive lithography-plus-etch steps, with pixel pitches of 10 $$A^{l} \;=\;\mathcal{V}\bigl(\mathcal{F}^{l}(x)\bigr) \;+\;\sum_{m=1}^{L-l}\mathcal{U}^{m}\bigl(A^{l+m}\bigr)\,.$$3m, 15 $$A^{l} \;=\;\mathcal{V}\bigl(\mathcal{F}^{l}(x)\bigr) \;+\;\sum_{m=1}^{L-l}\mathcal{U}^{m}\bigl(A^{l+m}\bigr)\,.$$4m, or 20 $$A^{l} \;=\;\mathcal{V}\bigl(\mathcal{F}^{l}(x)\bigr) \;+\;\sum_{m=1}^{L-l}\mathcal{U}^{m}\bigl(A^{l+m}\bigr)\,.$$5m (Aminzadeh et al., 2024).

Validation proceeds in two ways. First, the transmitted beam is imaged directly using a uniform monochromatic synchrotron beam: 12.4 keV for quaternary masks and 20 keV for binary masks, with a 6.5 $$A^{l} \;=\;\mathcal{V}\bigl(\mathcal{F}^{l}(x)\bigr) \;+\;\sum_{m=1}^{L-l}\mathcal{U}^{m}\bigl(A^{l+m}\bigr)\,.$$6m pixel detector (Aminzadeh et al., 2024). Second, a small pinhole of 5–20 $$A^{l} \;=\;\mathcal{V}\bigl(\mathcal{F}^{l}(x)\bigr) \;+\;\sum_{m=1}^{L-l}\mathcal{U}^{m}\bigl(A^{l+m}\bigr)\,.$$7m is rastered across the mask, and bucket signals are formed from the full-field detector image. For a 32×32 quaternary mask and a 20 $$A^{l} \;=\;\mathcal{V}\bigl(\mathcal{F}^{l}(x)\bigr) \;+\;\sum_{m=1}^{L-l}\mathcal{U}^{m}\bigl(A^{l+m}\bigr)\,.$$8m pinhole, the resulting 32×32 bucket map recovers all seven gray levels, and its cross-correlation with the ideal mask has a sharp central peak with 100% at the center and near zero elsewhere (Aminzadeh et al., 2024).

The forward model for scanning an object $$A^{l} \;=\;\mathcal{V}\bigl(\mathcal{F}^{l}(x)\bigr) \;+\;\sum_{m=1}^{L-l}\mathcal{U}^{m}\bigl(A^{l+m}\bigr)\,.$$9 with a known pattern $$C_t^{\,l} \;=\; \mathcal{V}\bigl(B^{l}\bigr) \;+\;\sum_{m=1}^{l-1}\mathcal{D}^m\bigl(C_t^{\,l-m}\bigr), \qquad C^{l} \;=\; C_t^{\,l} \;+\;\mathcal{F}^{l}(x)\,.$$0 is

$$C_t^{\,l} \;=\; \mathcal{V}\bigl(B^{l}\bigr) \;+\;\sum_{m=1}^{l-1}\mathcal{D}^m\bigl(C_t^{\,l-m}\bigr), \qquad C^{l} \;=\; C_t^{\,l} \;+\;\mathcal{F}^{l}(x)\,.$$1

or, in matrix form, $$C_t^{\,l} \;=\; \mathcal{V}\bigl(B^{l}\bigr) \;+\;\sum_{m=1}^{l-1}\mathcal{D}^m\bigl(C_t^{\,l-m}\bigr), \qquad C^{l} \;=\; C_t^{\,l} \;+\;\mathcal{F}^{l}(x)\,.$$2 (Aminzadeh et al., 2024). Deblurring can then be performed by cross-correlation with the mask:

$$C_t^{\,l} \;=\; \mathcal{V}\bigl(B^{l}\bigr) \;+\;\sum_{m=1}^{l-1}\mathcal{D}^m\bigl(C_t^{\,l-m}\bigr), \qquad C^{l} \;=\; C_t^{\,l} \;+\;\mathcal{F}^{l}(x)\,.$$3

The same paper notes that one may also write $$C_t^{\,l} \;=\; \mathcal{V}\bigl(B^{l}\bigr) \;+\;\sum_{m=1}^{l-1}\mathcal{D}^m\bigl(C_t^{\,l-m}\bigr), \qquad C^{l} \;=\; C_t^{\,l} \;+\;\mathcal{F}^{l}(x)\,.$$4 or use Tikhonov-type regularization $$C_t^{\,l} \;=\; \mathcal{V}\bigl(B^{l}\bigr) \;+\;\sum_{m=1}^{l-1}\mathcal{D}^m\bigl(C_t^{\,l-m}\bigr), \qquad C^{l} \;=\; C_t^{\,l} \;+\;\mathcal{F}^{l}(x)\,.$$5 (Aminzadeh et al., 2024).

7. Performance, dose redistribution, and limitations of the diffuse-probe method

The spatial resolution of the diffuse-probe HiXray method is set by the mask pixel pitch and the scan step size, both reported in the range 8–20 $$C_t^{\,l} \;=\; \mathcal{V}\bigl(B^{l}\bigr) \;+\;\sum_{m=1}^{l-1}\mathcal{D}^m\bigl(C_t^{\,l-m}\bigr), \qquad C^{l} \;=\; C_t^{\,l} \;+\;\mathcal{F}^{l}(x)\,.$$6m in the experiments (Aminzadeh et al., 2024). A 32×32 mask with 20 $$C_t^{\,l} \;=\; \mathcal{V}\bigl(B^{l}\bigr) \;+\;\sum_{m=1}^{l-1}\mathcal{D}^m\bigl(C_t^{\,l-m}\bigr), \qquad C^{l} \;=\; C_t^{\,l} \;+\;\mathcal{F}^{l}(x)\,.$$7m pitch therefore yields a 32-pixel field of view with 20 $$C_t^{\,l} \;=\; \mathcal{V}\bigl(B^{l}\bigr) \;+\;\sum_{m=1}^{l-1}\mathcal{D}^m\bigl(C_t^{\,l-m}\bigr), \qquad C^{l} \;=\; C_t^{\,l} \;+\;\mathcal{F}^{l}(x)\,.$$8m resolution over a 640 $$C_t^{\,l} \;=\; \mathcal{V}\bigl(B^{l}\bigr) \;+\;\sum_{m=1}^{l-1}\mathcal{D}^m\bigl(C_t^{\,l-m}\bigr), \qquad C^{l} \;=\; C_t^{\,l} \;+\;\mathcal{F}^{l}(x)\,.$$9m span (Aminzadeh et al., 2024). The numerical conditioning is reported as $A^l\in\mathbb{R}^{H\times W\times C}$0 for 2D masks up to 32×32, which the authors present as ensuring stability against noise and misalignment (Aminzadeh et al., 2024).

The method’s central motivation is dose redistribution. The reported scaling is by mask area: a 10×10 mask reduces local dose rate by approximately 100× compared with a 1-pixel focused beam of the same total flux, and a 32×32 mask yields approximately 1024× reduction (Aminzadeh et al., 2024). Demonstrations include imaging of sub-pixel pinholes of 5–20 $A^l\in\mathbb{R}^{H\times W\times C}$1m under very low SNR, as well as binary and gray test objects including patterned Ta and aluminium foils; reconstructed images are said to match ground truth with mean absolute errors of approximately 0.1 gray level per pixel (Aminzadeh et al., 2024).

The limitations are equally explicit. The method requires over-scan beyond the object edges by at least the mask footprint, which increases acquisition time. Mask fabrication becomes more complex as the number of gray levels increases and the pixel size decreases. In the reported beamline experiments, stage and data-transfer overheads limited scan speed to approximately 3 s per point, preventing large-area scans (Aminzadeh et al., 2024).

The paper also suggests several extensions: application to X-ray fluorescence and other scanning modalities, static coded-aperture imaging with large Huffman masks, 3D Huffman arrays such as 11×11×11 arrays with $A^l\in\mathbb{R}^{H\times W\times C}$2, and adaptive mask design via hybrid reverse-Monte-Carlo optimization (Aminzadeh et al., 2024). A plausible implication is that HiXray, in this second sense, should be understood less as a conventional focusing architecture than as a structured-illumination and inverse-problem framework for trading local energy concentration against computational reconstruction fidelity.