CSDD-Net: A cross semi-supervised dual-feature distillation network for industrial defect detection

Abstract

Detecting defects in industrial products is crucial to the strict quality control of products. Most current methods focus on supervised learning, relying on large-scale labeled samples. However, the forms of defects in industrial scenarios vary, and the data collection cost is high, which makes it difficult to meet the high requirements of massive labeled data. Therefore, we propose a Cross Semi-Supervised Dual-Feature Distillation Network (CSDD-Net), which aims to cross-use supervised and semi-supervised networks to learn rich feature representations and the distribution of large-scale features, respectively. CSDD-Net can transfer the defect feature distribution learned on partially labeled data in supervised branch to unsupervised branch, achieving simultaneous modeling and distillation based on partially labeled data. Firstly, this paper proposes a cross-local-global feature extraction network. By designing double interaction and ghost linear attention structure, it aims to force the network to be able to focus on local detail texture in global features and local features to perceive global semantics. Secondly, this paper proposes a Closed-Loop Cross-Aggregation Network (CLCA-Net), which considers deep and shallow semantics and fine-grained information. Thirdly, this paper designs a dynamic adaptive distillation loss, which could automatically adjust a more suitable regression loss function according to the defect characteristics, ensuring that the model could accurately locate and regress defects of various scales. Finally, this paper proposes a Glass Bottleneck defect dataset and verifies the feasibility of CSDD-Net in practical industrial applications. CSDD-Net achieved [email protected] of 80.41%, 76.42%, and 97.12% on the Glass Bottleneck, Wood, and Aluminum datasets with only 13.5 GFLOPs.