Under Review
DERA Detached Edge-Residual Adaptation for Prohibited Item Detection
DERA learns edge features and injects a small, controlled correction into a grounded detector.
Built for weak texture, severe overlap, and material clutter.
1 Khalifa University, Abu Dhabi, United Arab Emirates
2 The University of Western Australia, Perth, Australia
Improvement over the baseline under matched evaluation settings.
00 In brief
Find the boundary that matters.
Transmission X-rays obscure threats through overlap, weak texture, and material clutter, while generic edges often highlight benign structures. DERA combines a Swin contextual hierarchy with a parallel PiDiNet edge pyramid, then learns an object-specific boundary prior from instance contours available during training.
Detached gating isolates boundary supervision from the foundation detector. Zero-initialized adapters inject detection-guided corrections only at the first two feature levels. At inference, DERA needs only an X-ray image and category prompts.
01 The problem
Not every edge is evidence.
Threat contours can be faint or fragmented, while luggage frames, electronics, and benign objects often create stronger boundaries under X-ray imagery.
The challenge is not simply to add edges, but to let selected boundaries improve localization without rewriting the representation that already works.
02 Core idea
Understand. Select. Correct.
Three deliberate operations turn noisy X-ray structure into a controlled, detection-optimized update.
-
01
Understand the scene
A hierarchical Swin branch captures context while a parallel PiDiNet pyramid preserves local intensity transitions across four scales.
Context + local structure -
02
Select the boundary
A lightweight head learns an object-specific prior from the two highest-resolution edge features and training-time inner contours.
Task-relevant edges only -
03
Correct, don’t overwrite
The detached prior gates early edge features. Zero-initialized residual heads inject only the correction into the first two fused levels.
Preserve the foundation
Architecture
Detached Edge-Residual Adaptation
03 Training choreography
Three stages. Clear responsibilities.
Each objective gets its own stage so semantic adaptation, contour learning, and residual correction do not compete.
-
01
Learn the X-ray domain
Adapt contextual and edge branches, cross-stream fusion, and detection modules with the detection objective.
- Trained
- Visual + edge foundation
- Frozen
- Language encoder
-
02
Learn what to retain
Freeze the foundation and train only the side heads and fusion layer using inner-contour supervision.
- Trained
- 26.1K parameters
- Frozen
- Foundation + residual heads
-
03
Learn only the correction
Freeze the foundation and boundary branch; optimize zero-initialized residual heads with detection loss.
- Trained
- 14.7K parameters
- Frozen
- Everything else
Inference stays simple
X-ray image + category prompts only. No instance masks, contour targets, or external edge maps.
04 Evidence
Sharper localization across three benchmarks.
DERA consistently improves its Grounding DINO baseline under matched partitions, preprocessing, and category prompts.
PIDray12 categories
Baseline72.1→75.2
+3.1 APAP50 +3.4 · AP75 +3.0
CLCXray12 categories
Baseline64.0→65.6
+1.6 APAP50 +2.1 · AP75 +1.9
STCray21 categories
Baseline9.3→11.7
+2.4 APAP50 +7.4 · AP75 +4.2
Where DERA helps most+7.5 APS
Small-object improvement on CLCXray
PIDray+4.6
CLCXray+7.5
STCray+2.2
Positive gains at every reported scale show that stronger small-object sensitivity does not come at the expense of large objects.
Selected benchmark values Expand table
| Method | PIDray | CLCXray | STCray | ||||||
|---|---|---|---|---|---|---|---|---|---|
| AP | AP50 | AP75 | AP | AP50 | AP75 | AP | AP50 | AP75 | |
| Grounding DINO | 72.1 | 82.8 | 77.6 | 64.0 | 75.2 | 72.5 | 9.3 | 12.1 | 10.3 |
| BEF-Swin | 74.5 | 85.7 | 79.9 | 65.3 | 76.9 | 73.2 | 11.2 | 18.8 | 12.8 |
| DERA | 75.2 | 86.2 | 80.6 | 65.6 | 77.3 | 74.4 | 11.7 | 19.5 | 14.5 |
05 Efficient correction
Change less.
Keep more.
The final stage tunes a tiny residual pathway. The complete inference model adds modest capacity while preserving the foundation representation.
06 What the ablations teach
The restraint is part of the design.
Learn the edge
Generic Sobel gradients amplify irrelevant transitions. The learned PiDiNet branch improves strict localization by +4.6 AP75 over Sobel.
Sobel 75.3PiDiNet 79.9
Keep it shallow
The two-level pathway is both strongest and smallest. Deeper injection adds capacity while slightly reducing localization accuracy.
Residual-tuning parameters
P0–P114.7KP0–P261.2KP0–P3154.1K
Begin at zero
Zero initialization preserves the detector before tuning and improves strict localization by +0.5 AP75 over standard initialization.
R0useful correction
07 Qualitative microscope
See where the correction lands.
08 Honest boundaries
What boundary guidance does not solve.
DERA improves structural evidence, but it does not explicitly enforce instance separation or query consistency.
- 01Instance merging
Closely packed objects can merge into a single prediction.
- 02Partial localization
Recovered items may remain insufficiently localized at strict IoU thresholds.
- 03Duplicate predictions
Multiple queries can still respond to the same structure.
Citation
If you use DERA in your research, please cite our paper:
@misc{michael2026deradetachededgeresidualadaptation,
title={DERA: Detached Edge-Residual Adaptation for Prohibited item Detection},
author={Yonathan Michael and Mohamad Alansari and Mohammed Bennamoun and
Dwarikanath Mahapatra and Andreas Henschel and Naoufel Werghi},
year={2026},
eprint={2609.12411},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2609.12411},
}