LSFNet: Lightweight Slicing-Fusion Network for Pixel-Wise Damage Segmentation in Dam from UAV Imagery

Automated damage detection in a dam is important, while existing models and datasets like CrackForest or DeepCrack predominantly focus on crack detection, so which leads to frequent misclassification of spalling as cracks due to their shared visual characteristics. To address this, we constructed the Dam Damage Benchmark (DDB) and proposed a novel network, LSFNet. LSFNet is a lightweight slicing-fusion network for pixel-wise segmentation of damage in a dam from UAV Imagery. The encoder integrates the lightweight efficient aggregation feature network LEAFNet, spatial-slicing feature pyramids, and a lightweight efficient channel attention module LECA, dynamically enhancing cross-scale channel interdependence while maintaining computational efficiency based on spatial-slicing and channel-slicing. The decoder progressively merges hierarchical features through cascaded upsampling modules interleaved with channel recalibration blocks, effectively bridging the semantic gap of multi-granularity features. We introduce the Dam Damage Benchmark (DDB), comprising 1,396 high-resolution UAV images with pixel-wise damage annotations covering extreme lighting and two types of damage (cracks and spalling). Extensive experiments validate that LSFNet achieves 86.05% mIoU and 92.56% mPA‌ at 48 FPS with 8.81MB parameters, outperforming the other models by +7.05% mIoU on two damage segmentations.