Morphology-Invariance-Driven Transfer between sonar and optical images for underwater defect detection

Abstract

Underwater defect detection (UDD) from images is essential for the safe operation and maintenance of hydraulic structures. However, UDD remains challenging because of substantial variations in appearance, limited annotated data, and degraded underwater image quality. To mitigate these issues, this paper presents a deep learning-based Morphology-Invariance-Driven Transfer Network (MID-TransNet) for sonar-to-optical underwater defect detection. MID-TransNet uses defect morphology as the transferable cue and integrates a Sonar–Optical Morphological Perception Module (SOMPM), a Polarity-aware Convolution-Attention Mixer (PCAM), and a Semantic Enhancement Unit (SEU) to enhance morphology-aware representation, response balancing, and boundary-sensitive reconstruction. We construct the 2000-pair Underwater Sonar–Optical Defect Image Dataset (USODID) and evaluate the proposed method on the USODID benchmark and an independent external field test set. Experimental results show that MID-TransNet achieves macro-averaged mean Average Precision (mAP) values of 0.955 at an Intersection over Union threshold of 0.5 (mAP@0.5) and 0.661 averaged over thresholds from 0.5 to 0.95 (mAP@0.5:0.95), outperforming the evaluated baselines across crack, spalling, and hole detection. The external field evaluation further provides preliminary evidence of field-transfer capability under the current single-site setting.