Timber Defect Identification: Enhanced Classification with Residual Networks

Abstract

This study investigates the potential enhancement of classification accuracy in timber defect identification through the utilization of deep learning, specifically residual networks. By exploring the refinement of these networks via increased depth and multi-level feature incorporation, the goal is to develop a framework capable of distinguishing various defect classes. A sequence of ablation experiments was conducted, comparing our proposed framework's performance (R1, R2 and R3) with the original ResNet50 architecture. Furthermore, the framework's classification accuracy was evaluated across different timber species and statistical analyses such as independent t-tests and one-way ANOVA tests were conducted to identify the significant differences. Results showed that while the R1 architecture demonstrated slight improvement over ResNet50, particularly with the addition of an extra layer ("ConvG"), the latter still maintained superior overall performance in defect identification. Similarly, the R2 architecture, despite achieving notable accuracy improvements, slightly lagged behind R1. Integration of fully pre-activation activation functions in the R3 architecture yielded significant enhancements, with a 14.18% increase in classification accuracy compared to ResNet50. The R3 architecture showcased commendable defect identification performance across various timber species, though with slightly lower accuracy in Rubberwood. Nonetheless, its performance surpassed both ResNet50 and other proposed architectures, suggesting its suitability for timber defect identification. Statistical analysis confirmed the superiority of the R3 architecture across multiple timber species and this underscores the significance of integrating network depth and fully preactivation activation functions in improving classification performance. In conclusion, while the wood industry has made strides towards automation in timber grading, significant challenges remain. Overcoming these challenges will require innovative approaches and advancements in image processing and artificial intelligence to realize the full potential of automated grading systems.