Wood Defect Classification Based on Lightweight Convolutional Neural Networks

Abstract

Different types of wood defects correspond to different processing methods. Good classification means can transform defective boards into practical boards after appropriate processing. The detection accuracy of the wood surface defects is particularly important for improving the utilization rate and speed of processing the boards. The RegNet stands out in the field of computer vision. It automatically designs the network model based on the design space and applies it to wood defect detection, which can improve the classification accuracy. When the convolutional structure of the RegNet network is applied to industrial detection and classification, the problems of long real-time detection time and large algorithm parameters persist. This study focuses on collecting wood material images of common coniferous and broad-leaved trees in Northeast China with three types of defects: wormholes, slip knots, and dead knots. To improve the allocation of computing resources, based on the RegNet network model, an attention mechanism module was added, and the Ghostconv structure was introduced. The structure quickly and accurately highlighted the types of wood defects, improved the classification accuracy, reduced the parameters of the network, and exhibited generalization ability. To verify the performance of the improved network, MobileNet-v2, EfficientNet, and Vision-Transformer networks were introduced for comparative analysis. The improved RegNet network had smaller weight and higher accuracy, with a classification accuracy of 96.58%.