@article{LI2026, 
author = {Fang LI and Zhangheng WANG and Delin SUN and Xiangdong DAI and Zhou LYU and Tao TAO},
title = {Research on material selection optimization method for solid wood furniture based on visual language large model},
year = {2026},
journal = {Journal of Central South University of Forestry & Technology},
volume = {46},
number = {2},
pages = {229-236},
keywords = {solid wood furniture material selection, visual language large model, linear probing, intelligent manufacturing},
url = {https://www.sciopen.com/article/10.14067/j.cnki.1673-923x.2026.02.021},
doi = {10.14067/j.cnki.1673-923x.2026.02.021},
abstract = {【Objective】Wood selection represents a crucial stage preceding the manufacturing process of solid wood furniture. Currently, the solid wood selection in production lines is predominantly manual. To address this issue, an intelligent classification model has been devised. This model is capable of significantly enhancing the accuracy of solid wood identification, thereby establishing a solid foundation for ensuring the quality and quality related aspects of wooden furniture.【Method】Four types of wood frequently employed in solid wood furniture, namely Finnish pine, larch, beech, and ash, were selected as the research subjects. Based on the surface characteristics such as color and texture of the wood, nine distinct solid wood datasets were constructed. The task of solid wood furniture material selection was accomplished by leveraging the transfer learning approach and the image-text pair prediction and classification capabilities of the visual language large model. To further augment the recognition accuracy of wood for solid wood furniture, the contrastive language image pre-training (CLIP) visual language large model was introduced and optimized using Linear Probing. This optimization enabled the efficient identification and classification of multiple surface features of the wood.【Result】A total of 1 740 solid wood images were randomly partitioned into training and test sets at a ratio of 3∶1 and then compared and analyzed with the ResNet101 network and Full Fine-Tuning. By adopting the linear probing method within the CLIP model, the overall recognition accuracy of the model reach 98.29%, the mAP is 96.72%, the iterative parameter size of the model is 0.26 MB, and the output time per image is 37.48 ms. Both the classification performance and speed are optimized.【Conclusion】The results indicated that in the task of solid wood furniture material selection with limited sample training, the learning performance and stability of the proposed method have been effectively validated. It holds the potential for application in intelligent wood selection equipment, which can effectively boost the wood utilization rate and the production efficiency of solid wood furniture. Moreover, it provides significant guidance for promoting the intelligent manufacturing of solid wood furniture.}
}