Targeting Complex Textures in Guangxi Native Tree Species Wood Defects Intelligent Diagnosis: A Multi-Feature Fusion and Multi-Task Collaborative Learning Method

The wood of Guangxi native tree species exhibits unique material properties and textures. It poses challenges for intelligent defect recognition and surface quality evaluation during mechanical processing. Traditional methods perform poorly in handling complex texture features. They ignore the intrinsic relationship between defect information and surface roughness. This study proposes an intelligent diagnosis method for wood defects based on multi-feature fusion and multi-task collaborative learning. The proposed method innovatively combines multi-level texture features extracted by local binary patterns (LBP), Gabor filters, and a deep convolutional neural network (ResNet-50). It builds a robust feature representation adapted to complex texture backgrounds. It also constructs a multi-task collaborative optimization framework. It leverages shared low-level features and designs dedicated network structures for defect classification and roughness prediction. This enables synchronized and precise prediction for both tasks. Experimental results on the public WSDD dataset and a self-built dataset covering four dominant Guangxi native tree species (Dingguomu, Goushu, Kumian, and Yinhua) indicate that the proposed method achieves an accuracy of 94.5 % for defect classification. It shows improvements of approximately 16 % and 6 % over traditional methods and a single CNN model, respectively. For surface roughness prediction, it obtains an RMSE of 0.52 and an MAE of 0.41. These results are significantly better than existing methods. The study effectively addresses the challenges of intelligent diagnosis of wood defects in Guangxi native tree species under complex texture backgrounds. It also provides novel technical means and theoretical support for quality control and process optimization in wood processing.