Lightweight Instance Segmentation for Underwater Grout Overflow Monitoring in Offshore Wind Turbine Jacket Installation

The installation of grouted connections in deep-water offshore wind turbine jackets faces severe visibility constraints and limited diver operation time, necessitating an unmanned, real-time monitoring solution using remotely operated vehicles (ROVs). However, the lack of dedicated datasets and the challenge of balancing segmentation accuracy with edge-deployment efficiency remain major obstacles. This paper addresses these gaps by constructing, to the best of our knowledge, the first underwater grout overflow instance segmentation dataset reported in the published literature, comprising 2088 images and 5733 polygon instances annotated under a two-level semantic system (“overflow” and “overflow warning”). On this basis, we propose a collaborative optimization method based on YOLOv11n-seg and validate it with a strict 12-fold leave-one-group-out cross-validation (LOGOCV; 2088 images, video-level disjoint groups, unweighted macro averages). We first design a lightweight Multi-Path Feature Extractor (MPFE), MPFE-lite, retaining only the B3 asymmetric strip-convolution branch (depthwise 1 × 5 and 5 × 1 kernels). MPFE-lite improves mask mAP@0.5 over the YOLOv11n-seg baseline by 3.85 percentage points (35.9% → 39.8%; paired t-test p = 0.017, Wilcoxon p = 0.027, 10 wins out of 12 folds) while using 11.6% fewer parameters (2.54 M vs. 2.88 M) and slightly fewer GFLOPs (9.80 vs. 9.90). The full three-branch MPFE is also significantly better than baseline (mask mAP@0.5 +3.53 points, p = 0.007/0.002), but adding its dilated multi-scale and fixed Sobel branches to B3 gives no significant cross-group gain; the Dual-Branch Channel-Spatial Attention (DCSA) attention module and DySample upsampling, which appeared beneficial on a single fixed split, likewise produce no significant LOGOCV improvement and are therefore removed. For deployment-oriented compression, a Bidirectional Feature Pyramid Network (BiFPN)-based variant (1.21 M parameters, 7.38 GFLOPs) offers a 58.0% parameter reduction at a cost of 3.10 mask-mAP points, providing an optional lightweight configuration. Against peer lightweight models over an eight-fold partial comparison, MPFE-lite is significantly more accurate than YOLOv5n-seg and YOLOv6s-seg and comparable to YOLOv8n-seg while being lighter. Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations provide qualitative supporting evidence consistent with the strip-convolution design rather than definitive confirmation of its mechanism. This work provides a complete engineering closed-loop from model inference to on-site visualization, filling critical data and algorithmic gaps in underwater grout monitoring.