.png){kind=icon}
.png){kind=icon}
Return
Enhanced X-ray image object separation via residual diffusion modeling
.png){kind=icon}
PRE
AI
T
Z
F
H
W
Z
DOI：10.1117/1.JEI.35.2.023029.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
X-ray imaging plays a critical role in modern security systems. However, its high penetration capability frequently leads to mutual occlusion of objects in baggage scans. This occlusion degrades the accuracy of manual inspection and complicates automated detection. Existing approaches primarily rely on feature enrichment strategies to enhance object detection. These methods improve model performance by augmenting image features, particularly in overlapping regions. Nevertheless, they do not resolve the fundamental issue of object overlap. We introduce a framework that integrates image segmentation with generative modeling to separate overlapping objects in X-ray images. The proposed approach employs a diffusion-based model with residual prediction. An attention mechanism is incorporated into the U-Net architecture to enhance feature focus. The model takes carefully annotated segmentation masks as input. Furthermore, an end-to-end automated pipeline is developed, combining convolutional neural networks with separation networks. This design enables fully automatic separation of overlapping objects based on segmented regions. Due to the absence of standardized benchmark datasets for separation network training, the SplitXray dataset is introduced for overlapping object separation. Experimental results on the SplitXray dataset demonstrate superior performance (structural similarity index measure: 0.9831, learned perceptual image patch similarity: 0.0318, and Fr & eacute;chet inception distance: 143.4200) compared with baseline methods, marking a significant advancement in X-ray image analysis.
Keywords:
object separation
X-ray image
diffusion model
residual learning
end-to-end network
Journal
J
IF:
1
Papers:
109
Citations:
2.7K
Organization
Cited Papers