Particleboard surface defect recognition using shared-weight feature learning

Rare defect categories in particleboard manufacturing often have limited labeled samples, reducing the effectiveness of conventional image recognition methods. To address this issue, this study proposed shared-weight supervised contrastive learning for particleboard surface defect recognition (SSCL-PSDR), a particleboard surface defect recognition framework combining a shared-weight representation learning with supervised contrastive optimization. The framework focuses on five defect categories – Glue Spot, Scratch, Sand Leakage, Pollution, and Damage – and is evaluated on a single-source industrial dataset containing 393 defect patches. DINOv2 Transformer is adopted as the feature extraction backbone to capture global representations, while a lightweight multi-scale fusion and downsampling module enhances defect recognition at different spatial scales. An improved coordinate attention strategy further strengthens directional feature perception and highlights discriminative defect information. Supervised contrastive loss reduces intra-class variation and improves inter-class separability. Experimental results show that SSCL-PSDR achieves 88.8% accuracy, 88.9% precision, 88.7% recall, and 88.8% F1-score under fixed single-split evaluation, while maintaining robust average accuracy of 88.4 ± 0.8% across three independent data partitions and moderate computational complexity with 23.0M parameters and 25.2G FLOPs. Compared with a conventional pairwise Siamese baseline, SSCL-PSDR achieves improved recognition performance and remains competitive with other advanced deep learning models, demonstrating effectiveness for small-sample particleboard defect recognition under investigated industrial conditions.