Underwater Crack Detection and Segmentation via Transformer-Based Style Transfer With Global–Local Feature Fusion

Abstract

Deep-learning-enabled autonomous underwater vehicles (AUVs) have shown strong potential for autonomous underwater structural inspection, yet reliable crack detection and segmentation remain constrained by scarce underwater crack data and degraded imaging conditions. This article presents a crack-aware framework that combines structure-preserving style transfer for underwater crack image synthesis with DBE-YOLO for underwater crack detection and segmentation. In the style transfer stage, terrestrial crack images are translated into underwater-style samples using a Transformer-based model, where a quality-aware consistency loss is designed to suppress patchwise style artifacts and color distortion while preserving crack structures. In the perception stage, we propose DBE-YOLO, a YOLOv7-based detector-segmenter tailored to underwater crack geometry. DBE-YOLO incorporates deformable convolution v2 (DCNv2), bilevel routing attention (BRA), and Explicit Visual Center (EVC) to align adaptive sampling, contextual reasoning, and global–local feature fusion with the elongated morphology, weak visual contrast, and background interference of underwater cracks. Focal-EIoU and binary cross-entropy (BCE)-Dice losses are further adopted to improve bounding-box localization and mask delineation, respectively. A dedicated underwater crack dataset containing both synthesized and real underwater crack images is constructed for training and evaluation. Experiments on unseen real underwater crack images show that DBE-YOLO improves detection mean average precision (mAP) from 82.0% to 87.1% over the YOLOv7 baseline, while the BCE-Dice loss improves average precision (AP)-mask from 86.8% to 88.1% for segmentation. These results demonstrate the effectiveness of the proposed framework for underwater crack detection and segmentation under challenging underwater conditions.