Few-shot road surface defect detection using enhanced multi-level multi-scale graph neural network

Abstract

Detecting road surface defects in complex environments remains challenging due to and data scarcity and insufficient detail perception. In this paper, a stable diffusion and fractal texture hybrid data enhancement algorithm (DFH-Enhancement) is proposed to generate images of harsh environments that retain key semantic information for model training. An enhanced multi-scale graph neural network (MSKA-GNN) is also proposed for few-shot scenarios. This model converts images into graph data containing multi-scale information through graph neural network (CNN) and superpixel region aggregation, employs kolmogorov–arnold network (KAN) in place of multi-layer perceptron (MLP) to enhance the non-linear fitting capability of graph convolution, and combines multi-scale feature fusion (MSFF) and dual attention mechanism (DAM) to optimise feature extraction. Our experiments on two public datasets show that MSKA-GNN performs better than seven existing state-of-the-art models, achieving an accuracy of 81.1%.