Safeguarded online self-training for continual tunnel-face blast-hole detection

Continual full-image blast-hole detection is required as tunnel-face imagery accumulates under weak or unreliable hole-position priors, while hidden batch annotations prevent direct supervision during online updating. This study presents safeguarded online self-training, an artificial intelligence method that provides visual inputs for hole matching and automatic charging. Its core contribution is an update-level fixed-development-set gate that accepts or rejects a coupled detector–pseudo-label-memory state while batch references and test evidence remain inaccessible. The study uses 956 images and 9728 blast-hole targets from four tunnel–lithology source groups, three partitions of the same corpus and four label-hidden batches in a fixed order. For each batch, the active detector generates full-image pseudo-labels, raw confidence weights modulate pseudo-label losses, and one candidate is trained using initial labeled replay, committed memory, and the staged batch. The gate accepted 8 of 12 candidate updates. Relative to unconditional updating, mean final cumulative average precision increased by 0.0197 at an intersection-over-union threshold of 0.50 (AP50) and by 0.0192 when averaged over thresholds 0.50–0.95 (AP50–95), while forgetting decreased from 0.0135 to 0.0007. After model selection, a final-test campaign evaluated control and gate checkpoints; the fixed-development-set gate yielded modest partition-paired mean differences of +0.0067 AP50 and + 0.0110 AP50–95 relative to unconditional updating. Mean test precision, recall, their harmonic mean, AP50, and AP50–95 were 0.8810, 0.8521, 0.8663, 0.8834, and 0.4314. These results support joint detector–memory admission control for improved cumulative detection and reduced forgetting under the evaluated protocol. Future work includes unseen-tunnel evaluation and downstream charging integration.