This study introduces a novel hybrid deep learning framework, CNN–LPMPSO–GATVAE, designed to address the inherent challenges of wood surface defect classification, where high visual variability and complex texture patterns often hinder traditional inspection systems. The proposed model integrates a Convolutional Neural Network for hierarchical feature extraction, a Label Propagation–based Multi-objective Particle Swarm Optimization (LPMPSO) algorithm for feature selection based on k-kernel mean and ratio-cut objectives, and a Graph Attention Variational Autoencoder (GATVAE) for structure-aware representation learning. A key contribution of this work is the introduction of a closed-loop feedback mechanism between LPMPSO and GATVAE, allowing iterative refinement of both feature embeddings and optimization objectives to enhance classification robustness. Experiments conducted on a dataset of 20,275 labeled wood surface images demonstrate that the framework achieves high performance, with 95.89% accuracy, 95.79% precision, and a 96.84% F 1-score, although the CNN–LPMPSO–GATVAE model achieved higher mean accuracy, predictive accuracy, and F 1-score values compared to the baseline models, one-way ANOVA results showed that these differences were not statistically significant (such as accuracy p = 0.727 > 0.05). Therefore, performance improvement is understood in terms of quantitative enhancement and model stability, rather than confirming statistically significant superiority. The findings highlight the effectiveness of combining evolutionary optimization with graph-based deep learning, confirming the framework’s potential for deployment in real-time smart manufacturing environments. Overall, this research advances the methodological foundation of industrial visual inspection by delivering a more reliable, interpretable, and generalizable approach to automated defect detection.
Duc et al. (Thu,) studied this question.