Multi-Class Wood Surface Defect Classification Using Transfer Learning and CNN-SVM Model

Abstract

The automated identification of wood surface defects is crucial for maintaining product quality and maximizing resource efficiency in the timber industry. Yet, conventional deep learning approaches often face limitations due to the scarcity of labeled data and the considerable visual diversity across wood species. To tackle these issues, this study aims to explore a transfer learning-based CNN-SVM approach for multi-class classification of nine wood surface defect categories across four Malaysian hardwood species with distinct visual characteristics, using a dedicated wood defect dataset obtained from the Universiti Teknikal Malaysia Melaka (UTeM) wood defect database. Additionally, the wood samples were collected from several secondary wood product manufacturing facilities in the Bukit Rambai Industrial Area, Melaka, Malaysia, while image acquisition was conducted under controlled conditions to ensure consistent defect visibility and image quality. In this study, five pre-trained Convolutional Neural Networks (CNNs), such as AlexNet, VGG16, ResNet50, GoogLeNet, and MobileNetV2, are employed as fixed feature extractors, while a linear Support Vector Machine (SVM) serves as the classifier. This study evaluates the effectiveness of different CNN feature representations when integrated with SVM for multi-class classification of nine wood surface defect categories. Thus, the five pre-trained CNN-SVM model stability and generalization capability are evaluated through repeated experiments using both fixed and randomly generated train-validation splits. The experimental results show that the ResNet50-SVM approach achieves the highest classification accuracy while maintaining stable performance across different data partitions. Representative confusion matrix analysis further revealed that most defect categories were accurately classified, although minor confusion persisted between visually similar defects, such as Knot and Bark Pocket. In addition, a paired t-test further confirms that these improvements are statistically significant. The findings demonstrate that integrating deep convolutional feature extraction with SVM classification provides a reliable and effective framework for automated wood surface defect classification under limited and imbalanced dataset conditions, while also establishing a strong benchmark for automated wood surface inspection systems.