Abstract Accurate material recognition with low computational overhead is critical for edge applications such as autonomous drones, mobile robots, and smart manufacturing systems. Direct fine-tuning of deep backbones often leads to early saturation in validation accuracy due to overfitting on small, domain-specific datasets. To address this, we propose a structured multi-phase fine-tuning strategy for EfficientNetV2-S, progressively unfreezing layers over four stages with adaptive learning rate scheduling. The approach also incorporates label smoothing, dropout, and data augmentation to enhance generalization. We evaluated the method on a curated dataset of 1,730 images across four material classes: glass, metal, paper, and plastic. The resulting model achieves a validation accuracy of 95.66%, demonstrating that the proposed pipeline effectively balances accuracy and computational efficiency, making it suitable for real-time deployment on resource-constrained edge devices.
Tomić et al. (Wed,) studied this question.