PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Real-world X-ray Security Inspection: A High-Quality Benchmark and Lateral Inhibition Module for Prohibited Items Detection

Renshuai Tao, Yanlu Wei, Xiangjian Jiang, Hainan Li, Haotong Qin, Jiakai Wang, Yuqing Ma, Libo Zhang, Xianglong Liu

arXiv:2108.09917v1cs.CV

TL;DR

Prohibited-item detection in X-ray images is under-studied in part because specialized high-quality datasets are lacking. The paper presents HiXray and the Lateral Inhibition Module (LIM), and reports improved performance with state-of-the-art detection methods on HiXray and OPIXray.

  • Problem

    X-ray prohibited-item detection has few specialized high-quality datasets, and overlapping items create occlusion and heavy image noise.

    Source: §1 · pp. 1–2

  • Method

    The paper builds HiXray from real-world airport inspections and proposes LIM, which filters neighboring-region noise and activates item boundaries.

    Source: §1 · p. 2

  • Results

    LIM works with state-of-the-art detection methods and improves their performance on HiXray and OPIXray.

    Source: §1 · p. 2

  • Takeaways & Limitations

    HiXray provides a high-quality evaluation benchmark for prohibited-item detection, with 102,928 labeled instances across eight categories.

    Source: §1 · p. 2

Abstract

from arXiv · show

Prohibited items detection in X-ray images often plays an important role in protecting public safety, which often deals with color-monotonous and luster-insufficient objects, resulting in unsatisfactory performance. Till now, there have been rare studies touching this topic due to the lack of specialized high-quality datasets. In this work, we first present a High-quality X-ray (HiXray) security inspection image dataset, which contains 102,928 common prohibited items of 8 categories. It is the largest dataset of high quality for prohibited items detection, gathered from the real-world airport security inspection and annotated by professional security inspectors. Besides, for accurate prohibited item detection, we further propose the Lateral Inhibition Module (LIM) inspired by the fact that humans recognize these items by ignoring irrelevant information and focusing on identifiable characteristics, especially when objects are overlapped with each other. Specifically, LIM, the elaborately designed flexible additional module, suppresses the noisy information flowing maximumly by the Bidirectional Propagation (BP) module and activates the most identifiable charismatic, boundary, from four directions by Boundary Activation (BA) module. We evaluate our method extensively on HiXray and OPIXray and the results demonstrate that it outperforms SOTA detection methods.

1. Introduction

X-ray prohibited-item detection is difficult because items are often overlapped and noisy, while existing datasets and models do not adequately meet real-world needs. The paper introduces HiXray and LIM to address these data and detection challenges.

1. Introduction

  • Overlapping luggage contents occlude object regions and create heavy noise, while X-ray items often lack strong identifying visual properties.The paper identifies these characteristics as reasons the task requires high-quality datasets and models.Source: §1 · p. 2
  • HiXray contains 102,928 labeled instances across eight categories, gathered from real-world airport inspections and annotated by professional inspectors.The dataset is presented as a high-quality benchmark for prohibited-item detection.Source: §1 · p. 2
  • LIM combines Bidirectional Propagation to filter neighboring-region noise with Boundary Activation to capture and aggregate boundaries from four directions.The module is inspired by human recognition of identifiable characteristics amid irrelevant information, especially for overlapping objects.Source: §1 · p. 2
  • On HiXray and OPIXray, LIM is reported to work with state-of-the-art detection methods and improve their performance.Source: §1 · p. 2

2. Related Work

Prior work includes X-ray datasets with different tasks, scales, and image sources, alongside general object-detection approaches. The paper situates its work amid limitations in available security-inspection benchmarks.

2. Related Work

  • GDXray includes 19,407 grayscale images, SIXray has fewer than 1% positive samples and classification labels, and OPIXray contains 8,885 synthetic images across five cutter categories.Source: §2 · p. 2
  • The reviewed object detectors include one-stage and two-stage methods; SSD uses default boxes, YOLO emphasizes real-time performance and accuracy, and FCOS predicts per pixel.Source: §2 · p. 2

3. HiXray Dataset

HiXray is presented as a real-world, professionally annotated benchmark for prohibited-item detection, with data that also supports evaluation of small and occluded objects.

3.1. Construction Principles

  • The dataset is gathered from daily inspections at an international airport and uses privacy removal, eight frequent-item categories, expert box labels, and annotation checks.Inspectors follow shared guidelines, and another inspector checks each annotation, including for omitted objects.Source: §3.1 · p. 3

3. HiXray Dataset

3.2. Data Details

  • HiXray averages 2.27 instances per image, versus 1.37 in SIXray positive samples and 1 in OPIXray and GDXray.The authors describe the higher average as providing more contextual information.Source: §3.2 · p. 3, Table 3 · p. 3
  • The train/test ratio is about 4:1, images average 1200*900 pixels, and Table 4 reports object colors under X-ray.The paper notes that X-ray colors are mainly determined by chemical composition and may vary across machine models.Source: §3.2 · p. 3, Table 4 · p. 4

3.3. Potential Tasks

  • HiXray supports evaluation of small-object and occluded-object detection, among other detection tasks.For occlusion studies, researchers can divide the dataset into subsets by occlusion level.Source: §3.3 · p. 4
  • For small-object evaluation, boxes below 0.1% of image area are small, those above 0.2% are large, and the remainder are medium.Portable Charger 2 and Mobile Phone are divided into these size subsets; Table 5 reports their category distribution.Source: §3.3 · p. 4, Table 5 · p. 4

4. The Lateral Inhibition Module

LIM translates lateral inhibition into a module that filters neighboring-region noise and strengthens object-boundary information. Its bidirectional pathways combine feature maps across scales, while Boundary Activation detects boundary changes from four directions.

4. The Lateral Inhibition Module

  • LIM mimics lateral inhibition by filtering noisy information from neighboring regions and activating object boundaries as identifying features.It comprises Bidirectional Propagation (BP) and Boundary Activation (BA).Source: §4 · p. 4

4.1. Network Architecture

  • LIM takes a single-scale image of arbitrary size and produces proportionally sized feature maps at multiple levels, independently of the backbone architecture.The module’s processing combines its feature maps with corresponding backbone features for prediction.Source: §4.1 · p. 4

4.2. Bidirectional Propagation

  • BP uses dense connections in top-down and bottom-up pathways to filter neighboring-region noise across feature-map levels.The top-down pathway aggregates features from higher levels; the bottom-up pathway takes boundary-activated features as input.Source: §4.2 · p. 5

4.3. Boundary Activation

  • BA detects sudden changes at object boundaries and strengthens boundary information from four directions before aggregating it into a whole shape.The module evaluates whether positions are boundary points in the feature map.Source: §4.3 · p. 6

5. Experiments

Experiments test LIM across HiXray and OPIXray against detection baselines and feature-pyramid alternatives, then examine its components and visual performance.

5.1. Experiment Setting Details

  • Experiments use SGD and evaluate mAP at an IoU threshold of 0.5; LIM is integrated into SSD, FCOS, and YOLOv5 with their respective backbones.Source: §5.1 · p. 6

5.2. Comparing with SOTA Detection Methods

  • On HiXray, LIM improves average performance by 1.7%, 1.6%, and 1.5% over SSD, FCOS, and YOLOv5; on OPIXray, mean performance gains are 3.7%, 1.1%, and 2.8%.The comparisons cover both datasets and all three base detectors.Source: §5.2 · pp. 6–7, Table 6 · p. 7
  • LIM also outperforms DOAM by 1%, 1.1%, and 1% with SSD, FCOS, and YOLOv5 on HiXray, and by 0.6%, 0.7%, and 2.6% on OPIXray.Source: §5.2 · p. 7, Table 6 · p. 7
  • Performance remains particularly low for CO and NL across models; NL is small and blends into the background, while cosmetics vary in shape and are easily confused.Source: §5.2 · p. 7

5.3. Comparing with Feature Pyramid Mechanisms

  • LIM improves over FPN and PANet by 1.1% in SSD, over FPN by 1.1% in FCOS, and over PANet by 1.5% in YOLOv5.Portable Charger 1, Portable Charger 2, and Water show larger feature-specific gains over FPN, which the authors connect to their richer boundary information.Source: §5.3 · p. 7, Table 7 · p. 7

5.4. Ablation Study

  • Single-directional propagation improves performance by 0.7% over baseline; bidirectional propagation reaches 1.2% over baseline and 0.5% over single-directional propagation.Source: §5.4 · p. 7, Table 8 · p. 8
  • Adding Boundary Activation improves performance by 1.7% over baseline and 0.5% over Boundary Propagation alone.The ablations support the effectiveness of each LIM component.Source: §5.4 · pp. 7–8, Table 8 · p. 8

5.5. Visualization

  • Visualizations show improved localization over SSD, including a cosmetic detected at 91% confidence and two prohibited items detected where SSD finds one.Source: Figure 5 · p. 8, §5.5 · p. 8

6. Conclusion

The conclusion presents HiXray and LIM as contributions to X-ray prohibited-item detection. Evaluation on HiXray and OPIXray shows improved performance of SOTA detectors with LIM.

  • HiXray contains 102,928 common prohibited items across eight categories, gathered from real-world settings and manually annotated by professional inspectors.Source: §6 · p. 8
  • LIM targets overlap in X-ray images by suppressing noisy neighboring regions and activating item boundaries.Source: §6 · p. 8
  • Evaluation on HiXray and OPIXray demonstrates that LIM improves the performance of SOTA detection methods.Source: §6 · p. 8
Loading 2108.09917v1…