Semi-supervised YOLO-DEP for high-resolution X-ray component localization and counting

Area of Science:

• * Nuclear science and radiation imaging.
• * Industrial quality control and automated inspection.

Background:

• * Accurate localization and counting of small electronic components in high-resolution X-ray images is crucial but challenging.
• * Traditional methods struggle with cluttered scenes, and deep learning methods require extensive annotated data.
• * Existing approaches are limited in handling dense, small targets in complex industrial settings.

Purpose of the Study:

• * To develop a semi-supervised object detection framework for high-precision localization and counting of tiny electronic components in large X-ray images.
• * To reduce manual annotation costs by utilizing a novel semi-supervised label propagation strategy.
• * To introduce a new large-scale dataset, LEEC, for X-ray electronic component counting.

Main Methods:

• * Proposed YOLO-DEP, a novel object detector combining YOLOv11 with a Deep Encoding Processor (DEP) and Graph Attention Network (GAT).
• * DEP module employs half-channel and spatial attention for enhanced feature discrimination of small, dense targets.
• * Developed a semi-supervised label propagation strategy using feature similarity graphs and GAT for pseudo-label generation from minimal annotations.

Main Results:

• * YOLO-DEP demonstrated superior performance compared to state-of-the-art detectors on the LEEC and DOTAv1 datasets.
• * YOLO-DEP-x achieved 79.2% mAP50 and 70.9% mAP50-95 on the LEEC dataset.
• * Achieved a low counting error rate of 0.8%, indicating high accuracy in component counting.

Conclusions:

• * The proposed YOLO-DEP framework offers an effective solution for accurate electronic component localization and counting in high-resolution X-ray images.
• * The semi-supervised approach significantly reduces annotation effort, making it practical for real-world applications.
• * YOLO-DEP provides a deployable solution for industrial automation, nuclear inspection, and quality control.