Convolutional neural networks (CNNs) have achieved remarkable performance in industrial image-based surface defect inspection in recent years. However, many state-of-the-art (SOTA) networks have become increasingly complex and computationally expensive, which limits their deployment in resource-constrained environments or high-throughput, real-time, practical industrial applications. To address this challenge, this paper proposes a novel approach, Light SDI-NAS, to automatically design lightweight CNN architectures for real-time industrial surface defect inspection through neural architecture search (NAS). First, a task-oriented search space for industrial image inspection is constructed by integrating prior knowledge of neural network architecture design with empirical observations. Second, a novel loss function is introduced to balance model accuracy and computational efficiency during the architecture search process. Finally, the lightweight networks generated by Light SDI-NAS demonstrate strong performance on three industrial image datasets. Experimental results show that the proposed models achieve comparable or superior accuracy to manually designed SOTA networks while significantly reducing the number of parameters and improving inference speed by 1.8 times, making them highly suitable for real-time industrial inspection applications.
Chen et al. (Thu,) studied this question.