Fine Segmentation and Morphological Feature of Underwater Concrete Structure Defects

To address the challenges in accurately identifying and quantifying defects in underwater concrete structures of operational dams,this study proposes a method for pixel-wise segmentation and quantitative analysis of defect geometric morphology under complex noise conditions.First,given the high computational cost,dependency on high-quality data,and labor-intensive nature of precise defect annotation,we introduce a cross-domain transfer learning approach for defect pixel-level segmentation.This approach aims to reduce computational requirements and data dependency.The method leverages the deep residual network ResNet-50 and the lightweight convolutional neural network MobileNet-V2 as backbone feature extraction networks within the YolactEdge model to develop a multi-category defect pixel-level segmentation model specifically for underwater concrete dam structures.Furthermore,digital image processing techniques,such as connected component analysis,image refinement,and morphological operations,are employed to enable pixel-level quantitative analysis and extraction of geometric morphological features for various defect types.Experimental results demonstrate that the ResNet-50-based YolactEdge model achieves a bounding box detection accuracy of 97.34% (box_mAP@0.5) and a mask segmentation accuracy of 96.81% (mask_mAP@0.5) on the test dataset,while the MobileNet-V2-based YolactEdge model achieves a video inference speed of 47.87FPS.Additionally,the proposed method effectively identifies and segments multiple categories of defects on underwater concrete dam structures in complex inspection scenarios,including those involving obstacle interference,reagent color occlusion,and low visibility.Following segmentation,a quantitative analysis of defect pixel-level dimensions is conducted,extracting geometric features such as defect area,width,and length.