基于快速傅里叶增强深度网络的树皮种类识别

Bark species recognition using fast Fourier transform-enhanced deep network

快速且准确的树种识别为生态系统分析提供重要指导,但现有基于深度学习的方法无法满足林场实际作业需求,需要更简化、可部署的模型,为此笔者提出一种基于快速傅里叶增强深度神经网络的双分支网络。在第1个主干分支ResNet34中引入了门控通道注意力模块(GCT),以保证使用轻量模型且识别精度稳定;第2个分支利用快速傅里叶变换进行辅助监督,将树皮图像的灰度分布转换为频谱图分析图像纹理,通过辅助监督主干网络学习纹理差异性特征。本方法在3个树皮图像数据集上进行实验,其中,在23类树皮图像数据集中,准确率、预测精度和召回率3项评估指标分别达到了91.01%,89.65%和88.60%。此外,本方法显著减少了模型参数量,与原模型ResNet34相比,减少了33%的权重参数量。结果表明,该方法不仅提高了识别效率,降低了人力资源浪费,而且通过减少模型参数量,便于在各种终端设备上进行部署。

Rapid and accurate tree species identification is crucial for analyzing forest ecosystems,providing guidance for biodiversity understanding,habitat health monitoring,and conservation efforts.However,manual identification is time-consuming,inefficient,and requires specialized expertise,resulting in significant human resource wastage.To address these challenges,researchers have explored deep learning techniques for automated tree species identification using bark images.Yet,existing methods struggle with similar tree bark images in practical field operations.Moreover,models for identifying a larger number of tree species are complex and not easily deployable on resource-constrained end devices commonly used in the field.This study proposed a novel approach to tree species identification using a dual-branch network based on fast Fourier transform-enhanced deep neural networks.The first main branch incorporated a gate-controlled channel attention module(GCT)within the ResNet34 backbone.This module ensured the utilization of lightweight models while maintaining stable recognition accuracy.By focusing on informative regions of the bark images,GCT enhanced the model's ability to discriminate between similar tree species effectively.The second branch of the proposed network employed the fast Fourier transform as an auxiliary supervision mechanism.It transformed the grayscale distribution of bark images into spectrogram representations,enabling the analysis of image textures.This auxiliary supervision helped the main network learn features that were specifically related to texture differences,enhancing the discrimination capability of the overall model.Through the field collection of bark image datasets and the network collection of public data sets,the experiments were carried out on three different bark image datasets,and good results were achieved.In a 23-class bark image dataset,the proposed method achieved an impressive accuracy of 91.01%,precision of 89.65%,and recall rate of 88.60%.Moreover,this approach significantly reduced the parameter size of the model by approximately 33%compared to the original ResNet34 architecture.This reduction in parameter size made the proposed method more suitable for deployment on end devices with limited computational resources.Overall,the results demonstrated that the proposed method improved the efficiency of tree species identification while reducing human resource wastage.Additionally,the reduction in parameter size facilitated the deployment of the model on various end devices,making it more accessible for real-world applications in forest management and ecological research.