GACNet: Defect Detection with Graph Attention Mechanism Convolutional Network Model

Abstract

Convolutional Neural Networks (CNNs) are classic models for image classification. In modern industrial and manufacturing fields, automated defect detection and classification often rely on CNNs and their variant networks. However, due to the diversity and complexity of defects, traditional image processing methods often struggle to perform this task effectively. Graph Neural Networks (GNN), as effective tools for handling graph data, can capture relationships and local structures among nodes in a graph. This is valuable for describing the distribution and interconnections of defects in images. This paper introduces a deep learning framework that combines the advantages of both CNNs and GNNs, along with a method for transforming 2D images into graph data. In defect classification tasks, this framework outperforms ResNet-50 with pre-trained weights, achieving 4.07% higher precision.