Semi-Supervised Learning Method for Drill Pipe Detection

Abstract

In underground coal mine drilling operations, complex environmental factors such as low illumination, dust interference, and severe occlusion significantly degrade the performance of visual perception systems. Meanwhile, large-scale surveillance video data generated in mining environments are difficult to fully annotate, leading to insufficient scene coverage in labeled datasets and consequently limiting the generalization capability of detection models. To address these challenges, this paper proposes a semi-supervised object detection method with a multi-strategy pseudo-label optimization mechanism based on a teacher–student framework, aiming to improve drill-related object detection performance under low-labeling conditions. Specifically, in the pseudo-label generation and filtering stage, a dual-teacher consistency verification mechanism is introduced, along with a scene-aware noise evaluation factor to adaptively adjust pseudo-label selection thresholds, thereby reducing the impact of noisy labels in complex environments. Furthermore, a pseudo-label localization optimization module is designed from the perspective of regression stability, where the spatial consistency of multiple regression predictions is exploited to evaluate localization quality and eliminate unreliable pseudo-labels with large deviations. In addition, an uncertainty-aware weighting strategy is proposed to dynamically adjust training loss by jointly modeling pseudo-label confidence, environmental noise, and localization quality, enabling more effective utilization of high-quality pseudo-labels while suppressing the influence of low-quality samples. Experiments are conducted on a self-constructed underground drilling dataset under different labeling ratios. The results demonstrate that the proposed method significantly improves detection accuracy in low-label scenarios and exhibits strong robustness and generalization capability in complex underground environments. This study provides an effective solution for leveraging large-scale unlabeled data in intelligent visual monitoring for coal mine applications and offers valuable insights for semi-supervised object detection in complex industrial scenarios.