Answers are grounded in this paper and cite their sources.
Labeled rail surface defect samples are scarce, defect types are diverse, and pixel-level annotations are costly to obtain, which limits the practical applicability of supervised learning methods. To address identity mapping, unstable reconstruction under limited training data, and abrupt gradient changes caused by hard-threshold mining in reconstruction-based unsupervised anomaly detection, we propose Adaptive Momentum Teacher Reconstruction (AMTR). AMTR comprises three complementary modules: the Adaptive Perturbation Bottleneck (APB) dynamically scales perturbations according to feature magnitude and removes the attention-input residual connection in the first decoder block to suppress identity mapping; Feature Discrepancy-Guided Gradient Mining (FDGM) uses continuous position-wise weights to modulate backward gradients; and Momentum Teacher Consistency Learning (MTCL) employs an exponential moving average teacher to provide stable reconstruction targets. On the evaluated RSDDs-AUG split, AMTR achieves the highest mean result for each of the seven metrics among the compared methods. In three independent runs, it achieves mean image-level AUROC and pixel-level AUPRO values of 99.30±0.04% and 77.98±0.51%, respectively. Ablation studies provide evidence for the contributions of the proposed modules. The experimental evaluation is conducted on the RSDDs-AUG dataset, while generalization to other rail datasets and inspection conditions remains to be further investigated.
No takes yet. Share an insight, caveat, or question.
Song et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: