Wood species identification based on mask R-CNN with multi-feature extraction networks and hyperspectral imaging

Abstract

Different tree species exhibit significant variations in physical properties, uses, and economic value, making accurate species identification crucial. Traditional methods relying on human visual inspection are time-consuming and susceptible to subjective experience and fatigue. This paper proposes an RGB image expansion method based on hyperspectral data and an optimized Mask R-CNN model for wood species identification. First, 600 hyperspectral images of wood blocks of four tree species (Larch, Spruce, Birch, and Poplar) were collected. Principal Component Analysis was used to reduce the dimensionality of the hyperspectral images, followed by spectral band recombination to enhance texture features, resulting in a dataset of 1873 RGB images. Secondly, Leaky ReLU was used in place of ReLU as the activation function for the residual blocks. The ResNet50 and ResNet101 networks, combined with Feature Pyramid Networks were served as the two foundational feature extraction networks, and Convolutional Block Attention Module (CBAM) and Squeeze-and-Excitation Network (SENet) were inserted at different layers of the feature extraction network. Experimental results show that appropriate integration of attention mechanisms at different layers of the backbone can improve model accuracy and reduce loss rates. The ResNet101-CBAM3-SENet4 model exhibited the best overall performance, with precision of 0.9574, 0.9778, 0.9592, and 0.9783 for the four wood species in the test set, and an average precision of 0.9680. The mean Average Precision was calculated as 0.9657, and the mean Average Recall was 0.9806. This research provides new directions for dataset expansion in image identification and accurate identification of wood species with similar textures.