A deep learning-based recognition for dangerous objects imaged in X-ray security inspection device.

Area of Science:

• Computer Vision
• Image Analysis
• Security Technology

Background:

• X-ray security inspection accuracy is limited by current algorithms and datasets.
• There is a lack of dedicated datasets and studies for dangerous object segmentation in X-ray images.

Purpose of the Study:

• To improve the accuracy of dangerous object segmentation in X-ray security inspection.
• To address the scarcity of annotated datasets for dangerous object detection.

Main Methods:

• Manual pixel-level semantic segmentation of the SIXRay dataset.
• A novel composition method using affine transformation and HSV features for data augmentation.
• Integration of Softer-Non-Maximum Suppression (Softer-NMS) with Mask RCNN, creating the Softer-Mask RCNN model.

Main Results:

• The Softer-Mask RCNN model achieved a 3.4% increase in mean Average Precision (mAP) compared to the original Mask RCNN.
• Performance improved by 6.2% when incorporating synthetically generated data.
• The proposed methods effectively enhance the recognition of adjacent and overlapping dangerous objects.

Conclusions:

• The developed manual segmentation and data augmentation techniques provide valuable resources for X-ray security inspection research.
• The Softer-Mask RCNN model offers a significant improvement in detecting dangerous objects within X-ray imagery.
• This work contributes to more accurate and reliable security screening through advanced image analysis.