Deep learning–based automated fine-grained and quantitative detection of defects in submerged concrete

Concrete immersed in water degrades under coupled mechanical, hydraulic, and chemical actions, with visible defects often preceding hidden material loss and durability decline. Traditional manual inspection methods have limitations such as low efficiency and insufficient accuracy, while underwater optical inspection technology inevitably generates a large amount of redundant video inspection data. To address this, this study develops a fine-grained detection and size quantification of defects on submerged concrete using deep learning and computer vision. Firstly, a general-purpose above-water dataset covering diverse defect types was built, and a dual-stage transfer-learning pipeline was proposed to adapt representations to underwater imagery, thereby substantially reducing annotation needs. Then, an improved defect pixel-wise segmentation network for submerged concrete combining YOLACT and the Swin Transformer is developed for fine-grained processing of video inspection data, while a tailored morphological analysis extracts geometric features of multi-category defects across different scales. A large water-related concrete structure in long-term service was used as a case study. Validation on the long-serving hydraulic concrete structure under turbidity, occlusion, and uneven illumination across four defect categories (crack, holes, exposed aggregate, depressions) demonstrates consistent superiority over six benchmarks, with gains up to 6.8% in mAP. The framework enables near-real-time generation of quantitative defect indicators from ROV video, supporting durability assessment, trend monitoring, and condition-based maintenance while reducing reliance on hazardous, labor-intensive manual surveys.