Semi-supervised contour-driven broad learning system for autonomous segmentation of concealed prohibited baggage...

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Manual X-ray baggage screening is error-prone and raises privacy concerns.
• Current deep learning models require extensive data and pixel-wise annotations, with separate fine-tuning for each dataset.
• There is a need for efficient threat segmentation in aviation security.

Purpose of the Study:

• To propose a semi-supervised contour-driven broad learning system (C-BLX) for X-ray baggage security threat instance segmentation.
• To enhance representation learning and achieve faster training with limited data and minimal supervision.
• To address challenges like severe occlusion and class imbalance in threat detection.

Main Methods:

• Developed a contour-driven broad learning system (BLS) named C-BLX.
• Employed resource-efficient image-level labels for training.
• Generated candidate region segments based on local intensity transitions to identify concealed items.
• Utilized multi-convolutional BLS to extract complementary features for object categorization.
• Applied contours of predicted threat segments for final segmentation results.

Main Results:

• The C-BLX system achieved high performance on three imbalanced public datasets.
• Achieved mean Intersection over Union (mIoU) scores of 90.04% on GDXray, 78.92% on SIXray, and 59.44% on Compass-XP.
• Outperformed competitive approaches in baggage-threat segmentation.

Conclusions:

• The proposed C-BLX system offers an efficient and accurate solution for X-ray baggage threat segmentation.
• The framework successfully localizes illegal items using minimal supervision and resource-efficient labels.
• Future work will explore post-processing techniques to overcome limitations in noisy settings.