Wood ViT: the vision transformer for segmenting and identifying wood micro-structure

• Proposes Wood-ViT with CNN & Transformer to segment & recognize wood micro-structure. • Designs and implements Convolution and Vision Transformer block, i.e. PCB & WTB. • Balances wood micro-structure detail recognition and long-range dependency modeling. • Experiments show Wood-ViT outperforms existing methods in wood microscopic image. Accurate identification of wood species is the foundation for ensuring the quality of wood products. Technological progress integrates deep learning with attention mechanisms to enable precise microstructure image segmentation, vital for wood species identification and product quality assurance. In this paper in order to address the issues of ViT's (Vision Transformers) poor recognition of wood grain details and CNN's (Convolution Neural Networks) weak long-range dependency the ViTs and the CNNs are combined to implement complementary advantages, namely both a strong ability to distinguish wood details and long-term dependency. Wood micro-structure image processing requires the ability to recognize or segment micro-structures such as wood vessels and wood rays. For this task, we design and implements the Wood Vision Transformer (Wood-ViT) specifically proposed for wood micro-structure images. Extensive experiments have demonstrated that the performance of Wood-ViT surpasses the existing CNNs, ViTs, and CNN-Transformer hybrid architectures in terms of the performance metrics of various visual tasks. In the training and testing of the self-labeled wood microscopic images, the average intersection over union (mIoU) of our method's segmentation was 3.1% higher than that of ResNet and 3.7% higher than that of ViT, achieving a satisfactory result.