Evaluation of a Learning Method Combining a Penalty for False Negatives and Online Hard Negative Mining for Automatic Concrete Crack Detection

Abstract

As a growing proportion of road bridges continue to age, the demand for automated crack inspection is becoming increasingly urgent. However, in automated crack detection for concrete bridge decks, reducing false negatives often leads to an increase in false positives, whereas suppressing false positives tends to increase the risk of missed detections. In addition, surface-specific noise in real inspection environments, such as color irregularities, structural boundaries, and efflorescence, frequently triggers spurious detections, making it difficult to achieve a level of performance suitable for field deployment. To address this challenge, this study proposes a two-stage penalty learning framework for semantic segmentation, in which false negatives (FNs) are heavily penalized during the early stage of training, followed by the introduction of Online Hard Negative Mining (OHNM) to suppress false positives. Comparative experiments using images acquired from actual bridge decks showed that the proposed method improved both IoU and F1 score. Furthermore, paired t-tests with Holm correction confirmed that these improvements were statistically significant. The proposed approach can therefore be regarded as a promising learning strategy for practical crack detection systems that must simultaneously cope with missed detections and false alarms.