<![CDATA[Wood type classification in the timber industry is frequently hindered by the high visual similarity of grain patterns, a structural challenge that is particularly pronounced with small-scale datasets. This study systematically resolves this fine-grained visual ambiguity by developing a rigorously optimized InceptionV3 framework. Initial baseline evaluations conducted on a dataset of 800 images encompassing four wood species, namely Pterospermum javanicum (Bayur), Magnolia champaca (Cempaka), Tectona grandis (Jati), and Melia azedarach (Mindi), revealed that a standard ResNet50 architecture experienced a severe performance degradation, generating a random guess accuracy of merely 25 percent. Addressing this severe architectural inadequacy, the proposed InceptionV3 model integrates highly targeted spatial augmentations, a Dropout rate of 0.6, and L2 regularization. Furthermore, the optimization strategy deploys Stochastic Gradient Descent with Nesterov momentum and Label Smoothing to explicitly mitigate intra-class visual similarities. Consequently, the proposed framework achieved a robust overall accuracy of 89 percent, surpassing the baseline by a substantial 64 percent, alongside a macro-averaged F1 Score of 0.89. These empirical findings substantiate that specific architectural fine-tuning and advanced stochastic regularization are highly essential for capturing subtle wood-grain patterns, thereby offering a highly reliable, automated quality-control solution for the timber industry.]]>
Copyrights © 2026
