FARR: An Efficient Frozen-Feature Learning Framework for Wood Species Identification with Applications to Texture Recognition

The scarcity of high-value timber resources and increasing market fraud have intensified the demand for efficient wood species identification technologies. Traditional methods and deep learning approaches using X-ray imaging are constrained by efficiency limitations and insufficient training samples. This paper proposes FARR, a wood species identification framework that completely freezes pre-trained networks as feature extractors. The framework achieves efficient feature learning through cross-block feature aggregation and learnable residual connections , attention mechanisms, randomized autoencoders, with low computational complexity requiring optimization of only a few parameters. Experimental results demonstrate that the proposed method achieves 99.86% accuracy on wood identification tasks, with training efficiency improved by 20–68× over partial fine-tuning by freezing the backbone and reducing parameter updates. The method exhibits excellent robustness and generalization capability across multiple datasets and complex environmental conditions. This study provides an efficient solution for intelligent wood identification in resource-constrained scenarios, reducing computational costs while maintaining identification accuracy and offering new insights for deep learning applications under small-sample conditions.