MGABLS: a broad learning system combining a graph attention autoencoder and manifold learning for few-shot defect detection

As a novel neural network learning framework, a broad learning system (BLS) offers advantages such as simple network architecture, few model parameters and rapid training speed. However, this inherent simplicity hinders its ability to effectively capture structural information among samples, particularly when the sample size is limited. Such limitations impede the model capacity to accurately learn the data distribution of training samples, ultimately reducing classification accuracy in defect detection tasks. To address this issue, this paper proposes a novel BLS that combines a graph attention autoencoder and manifold learning (MGABLS). This approach enhances the feature space representation of BLS by leveraging a graph attention autoencoder (GATE) to better capture the structural relationships among samples. Furthermore, a dual-graph regularisation term based on manifold learning is introduced to strengthen the model constraints on output weights, thereby improving its ability to learn the intrinsic structure of the data structures. MGABLS is evaluated on four surface defect datasets. The experimental results for visual inspection of surface defects demonstrate that MGABLS achieves higher classification accuracy than competing methods. The effectiveness of this approach is further validated through ablation studies.