基于深度学习的轻量化木材种类分选模型  

Lightweight Wood Species Sorting Model Based on Deep Learning

作　　者：王正[1] 杨帆 江莺[1] WANG Zheng;YANG Fan;JIANG Ying(College of Mechanical and Electronic Engineering,Nanjing Forestry University,Nanjing Jiangsu 210037,China)

机构地区：[1]南京林业大学机械电子工程学院,江苏南京210037

出　　处：《林业机械与木工设备》2025年第1期84-89,共6页Forestry Machinery & Woodworking Equipment

摘　　要：为了提高木材种类行业生产的智能化程度,基于GhostNetv2设计了轻量化的改进注意力机制模型对木材种类进行识别分类。针对深度可分离卷积给模型带来更高的内存访问和更低的整体计算速度的缺点,设计了基于部分卷积的Bottleneck模块;针对GhostNetv2模型缺乏高性能特征融合的缺点,引入了基于金字塔分割注意力(Pyramid Split Attention,PAS)的特征融合模块。设计的改进注意力机制的木材种类分选的Ghost-FasterNet轻量化模型,综合考虑了模型的识别效果、参数大小、推理时间以及训练时间,使用Top-1准确率和Top-5准确率作为评价指标。实验结果表明:提出的Ghost-FasterNet轻量化模型在推理时间和训练时间与其他轻量型网络基本保持一致的同时,减少了大量参数,在强注意力机制和部分卷积的精度补偿下,模型准确率大幅度增加,最高准确率达到87%,相较于其它传统的深度学习模型,提高了近10%。In order to improve the intelligence level of wood industry production,this paper designs a lightweight improved attention mechanism model based on GhostNetv2 to identify and classify wood types.A Bottleneck module based on partial convolution is designed to solve the shortcomings that deep separable convolution brings higher memory access and lower overall computing speed to the model.Aiming at the lack of high-performance feature fusion in GhostNetv2 model,a feature fusion module based on Pyramid Split Attention(PAS)is introduced.The Ghost-FasterNet lightweight model for wood species sorting with improved attention mechanism was designed in this paper.The recognition effect,parameter size,inference time and training time of the model were comprehensively considered,and Top-1 accuracy and Top-5 accuracy were used as evaluation indexes.The experimental results show that:While the inference time and training time of the Ghost FasterNet lightweight model proposed in this paper are basically consistent with other lightweight networks,a large number of parameters are reduced.Under the precision compensation of the strong attention mechanism and partial convolution,the accuracy of the model is greatly increased,with the highest accuracy reaching 87%.That's an increase of nearly 10 percent.

关 键 词：深度学习 木材种类分选 轻量化模型 

分 类 号：S781.61[农业科学—木材科学与技术] TP183[农业科学—林学] TP391.4[自动化与计算机技术—控制理论与控制工程]