MSDC‐Net: Multiscale Optimization and Dynamic Convolution for Instance Segmentation in Underwater Dam Cracks

The detection and segmentation of dam cracks are crucial for water conservancy project maintenance and safety assessment. However, underwater image often suffers from poor lighting and noise, causing blurred edges and low contrast. This increases the difficulty of segmentation. Additionally, cracks have complex shapes and varying scales, making it hard for existing methods to capture fine details. To address these issues, we propose a new crack instance segmentation framework. First, in the feature extraction stage, we propose a multiscale optimization module (MSOM). This module adjusts the receptive field’s shape and size based on the input image, improving the extraction of crack features at different scales. Then, we introduce a spatial feature pyramid aggregation FPN (SFPAFPN) to enhance multiscale feature fusion with a redesigned fusion strategy. We also design a dynamic perception convolution module (DPC), which optimizes sampling point distribution through a flexible sampling mechanism, improving the segmentation of crack edges and fine details. Experimental results show that our method performs well in underwater dam crack segmentation and underwater object detection, with AP 50 in mask reaching 52.9% and 41.5%.