Few-shot learning for image-based bridge damage detection
Yan Gao, Haijiang Li, Weiqi Fu
Abstract
Autonomous bridge visual inspection is a real-world challenge due to various materials, surface coatings, and changing light and weather conditions. Traditional supervised learning relies on massive annotated data to establish a robust model, which requires a time-consuming data acquisition process. This work proposes a few-shot learning (FSL) approach based on improved ProtoNet for damage detection with just a few labeled examples. Feature embedding is achieved through cross-domain transfer learning from ImageNet instead of episodic training. The ProtoNet is improved with embedding normalization to enhance transduction performance based on Euclidean distance and a linear classifier for classification. The approach is explored on a public dataset through different ablation experiments and achieves over 94% mean accuracy for 2-way 5-shot classification via the pre-trained GoogleNet after fine-tuning. Moreover, the proposed fine-tuning methods based on a fully connected layer (FCN) and Hadamard product are demonstrated with better performance than the previous method. Finally, the approach is validated using real bridge inspection images, demonstrating its capability of fast implementation for practical damage inspection with weakly supervised information.