Work

Pseudo-label optimization and DIoU-NMS for semi-automatic annotation in scrap steel detection

Year: 2026

Type: article

Abstract: Industrial scrap steel detection is challenged by dense stacking, fine-grained morphological differences, and severe category imbalance, which make fully manual annotation prohibitively labor-intensive and leave even carefully labeled datasets with missing or imprecise objects. To address this bottleneck, we propose a Semi-Supervised Pseudo-Label YOLO11 detector (SSPL-YOLO11). Built on an offline teacher–student pseudo-labeling framework, SSPL-YOLO11 applies a three-level pseudo-label quality control mechanism—confidence thresholding, DIoU-NMS deduplication, and lightweight manual verification—to recover unlabeled objects at a small manual cost. Controlled experiments quantify the impacts of the confidence threshold, the deduplication strategy, and the pre-trained weight initialization. On the public Scrap Steel Image Dataset (SSID), SSPL-YOLO11 achieves 88.25% $$\text {mAP}_{50}$$ and 65.58% $$\text {mAP}_{50--95}$$ on the independent test set, surpassing the fully supervised YOLO11 baseline by 3.92 and 7.74 percentage points, respectively. The proposed workflow substantially reduces manual annotation cost and provides a practical route toward intelligent scrap steel sorting and recycling.

Source: Scientific Reports

DOI: 10.1038/s41598-026-76096-6